A digital twin framework for predicting and simulating type 2 diabetes onset using retrospective lifestyle data

Abstract

Introduction:

Type 2 Diabetes Mellitus (T2DM) is a rising global health concern, heavily influenced by modifiable lifestyle and psychosocial factors. However, most predictive tools focus on biomedical markers and rely on real-time data from wearables or electronic health records, limiting their scalability in resource-constrained settings. This study presents a novel digital twin (DT) framework that uses retrospective lifestyle, behavioral, and psychosocial data to forecast T2DM onset and simulate the estimated effects of preventive interventions.

Methods:

Data were drawn from 19,774 participants in the UK Biobank cohort, followed for up to 17 years. A penalized Cox proportional hazards model was employed to estimate individual time-to-event risk trajectories based on 90 candidate predictors. Predictors were selected through univariate screening, multicollinearity assessment, and variance filtering, yielding a final model with 14 significant variables. Causal inference techniques, including directed acyclic graphs (DAGs) and counterfactual simulations, were used to explore intervention effects on disease progression.

Results:

The model demonstrated strong predictive performance (C-index 0.90, SD 0.004). Psychosocial stressors such as loneliness, insomnia, and poor mental health emerged as strong independent predictors and were associated with estimated increases in absolute T2DM risk of approximately 35 percentage points individually and nearly 78 percentage points when combined, under the modeled assumptions. These effects were partly reinforced through diet, with high intake of processed meat, salt, and sugary cereals acting as risk amplifiers within the modeled causal pathways. Cheese intake was protective overall, but its estimated benefit was attenuated under psychosocial stress, where reduced consumption produced a small, directionally harmful mediation effect. Counterfactual simulations suggested that improvements in psychosocial conditions could reduce estimated T2DM risk by approximately 11.6 percentage points within the modeled cohort, with protective dietary patterns such as cheese consumption re-emerging as psychosocial stress was alleviated. The model also revealed pronounced ethnic disparities, with South Asian, African, and Caribbean participants exhibiting significantly higher estimated risk than White counterparts within this cohort. These findings highlight the potential of integrated, stress-informed prevention strategies that address both psychosocial and dietary pathways.

Conclusion:

This study introduces a transparent, simulation-enabled DT framework for estimating T2DM risk and exploring behavioral intervention scenarios without reliance on real-time data streams. It enables interpretable, personalized prevention planning and supports exploration of scalable deployment in public health, particularly in underserved or low-infrastructure environments. The integration of psychosocial and lifestyle data represents an important step toward more equitable and behaviorally informed digital health solutions.

1 Introduction

Type 2 Diabetes Mellitus (T2DM) is a chronic and progressive metabolic disorder marked by persistently elevated blood glucose levels, or hyperglycemia. According to the International Diabetes Federation (IDF), T2DM remains one of the most pressing global public health concerns, currently affecting over 537 million people worldwide. The IDF projects that this number will rise sharply to 643 million by 2030 and further to 783 million by 2045 [1]. If undiagnosed or poorly managed, T2DM can lead to a range of severe health outcomes, including kidney failure, vision loss, limb amputations, and increased mortality risk [2].

While genetic factors such as variants in the TCF7L2, FTO, and PPARG genes contribute to the development of T2DM [3, 4], there is strong evidence that modifiable lifestyle factors play a more dominant role in its onset and progression [5, 6]. These are behaviors or conditions that individuals can change through personal action or public health interventions, including physical inactivity, unhealthy dietary habits, excessive body weight (obesity), and chronic psychosocial stress. These factors contribute to metabolic dysfunction and are strongly associated with long-term complications such as cardiovascular disease, kidney damage, and stroke [7, 8].

Emerging evidence also underscores alcohol consumption as a vascular risk factor, with recent UK Biobank analyses showing positive linear associations between alcohol intake and arterial stiffness, challenging earlier assumptions of protective effects at moderate levels [9]. Additionally, chronic psychosocial stress, which may arise from financial hardship, social isolation, job related pressure, or caregiving responsibilities, is increasingly linked to elevated T2DM risk. Such stressors, together with mental health issues such as depression, anxiety, and sleep disturbances, can negatively affect daily routines, impair self regulation, and disrupt biological processes. These disruptions contribute to insulin resistance and poor glycemic control through both behavioral and physiological pathways [10–13].

1.1 Problem statement

Despite this growing understanding, current risk prediction models and clinical decision tools for T2DM remain predominantly centered on conventional biomedical indicators such as body mass index (BMI), age, fasting glucose, and blood pressure [14, 15]. This narrow clinical focus often overlooks the interconnected behavioral and emotional factors that precede and shape disease onset, creating a significant gap in the design of prevention strategies [16].

To improve prediction and personalization in chronic disease management, technologies such as artificial intelligence (AI), machine learning (ML), and digital twins (DTs) have gained significant attention in recent years. Research is currently being conducted on DTs to explore the possibility of generating dynamic models and simulations of human physiology, with the goal of enhancing patient care and treatment options [17–19]. These approaches show promise, but they also face important limitations [20–23]. Many AI and ML models operate as “black boxes.” They can make accurate predictions but often do not explain how or why certain outcomes occur, which limits their interpretability in clinical settings [24]. While some studies [10–13] have begun to incorporate psychosocial factors such as stress, loneliness, and mental health, these variables remain underrepresented in many mainstream models. As a result, the broader influence of behavioral and emotional factors on outcomes like T2DM is often overlooked or insufficiently modeled, despite growing evidence of their clinical relevance.

DT systems, which are virtual models that replicate individual health profiles to simulate disease progression and intervention effects, offer a promising approach for advancing personalized and preventive medicine. By integrating diverse health data, DTs enable clinicians to test ”what-if” scenarios and tailor care to individual needs [25, 26]. However, most existing DT implementations rely heavily on real-time data streams from wearables, biosensors, or electronic health records (EHRs), which presents significant limitations. These include the need for continuous data acquisition, costly technological infrastructure, and substantial concerns around data privacy and interoperability [27]. Such constraints restrict the scalability and adoption of DT systems, especially in low-resource or decentralized healthcare settings.

1.2 Proposed solution

To address the limitations of existing prediction models and DT systems for T2DM, this study introduces a novel DT prototype that operates entirely on historical behavioral and psychosocial data and does not require real-time data streams or specialized sensors, making it suitable for deployment in settings with limited technical infrastructure. By utilizing existing datasets, such as the UK Biobank [28], the prototype supports population-level scalability while retaining the capacity for individualized simulation. This positions the framework as a cost-effective and widely accessible solution for early-stage prevention, particularly in public health settings and underserved communities.

Unlike traditional models that focus on physiological markers (BMI, blood glucose, and age etc.,) the framework presented in this study includes psychosocial risk factors such as loneliness, sleep disturbances (e.g., insomnia), and depressive symptoms. These variables are not only included in survival analysis for outcome prediction but are also positioned as causal mediators and moderators within the disease pathway, offering a richer and more behaviorally contextualized understanding of T2DM risk.

To maintain transparency and clinical relevance, the system uses penalized Cox regression [29] for survival modeling and incorporates causal inference methods such as backdoor adjustment [30] and mediation analysis. This allows clinicians and researchers to trace how modifiable risk factors influence outcomes through biologically and behaviorally plausible pathways, enhancing model credibility, clinical trust, and interpretability.

Beyond predictive accuracy, this study lays the groundwork for a new class of simulation-enabled, behavior-aware DT systems for chronic disease prevention. It offers actionable insights for a broad spectrum of stakeholders: clinicians gain tools for tailored risk management, researchers benefit from a validated framework for behavioral modeling, and policymakers are empowered with scalable strategies for public health intervention, particularly in settings where real-time data infrastructure is limited. Finally, the proposed DT system was validated using a series of techniques recommended by Sharma et al. [31]. These included placebo tests, subset sampling, bootstrap refutation tests, and sensitivity analysis to detect hidden confounding, thereby ensuring the robustness of the findings.

Together, these elements contribute to a novel, simulation enabled DT prototype capable of generating personalized, interpretable, and clinically actionable prevention strategies for T2DM, particularly in environments where real-time data capture is not feasible.

2 Related work, research gap and key contributions

DTs have demonstrated considerable potential for enhancing patient monitoring, risk stratification, and individualized treatment planning [32, 33]. While effective in high resource or clinical settings, these implementations face significant challenges related to cost, scalability, and privacy, which hinder their broader deployment, particularly in community level or resource limited environments.

2.1 DT-driven precision nutrition and remission strategies

One of the most advanced applications of DTs in T2DM lies in precision nutrition. Shamanna et al. [34, 35] demonstrated that integrating continuous glucose monitoring data with dietary inputs improved glycemic outcomes, reduced BMI, and decreased reliance on diabetes medication. Their Twin Precision Nutrition (TPN) and Twin Precision Treatment (TPT) platforms supported rapid and sustained metabolic benefits [36], and introduced a structured, DT-guided seven-stage remission model [37].

However, these systems rely on continuous biosignal input and clinical oversight, which limits their use in preventive or low infrastructure contexts. In contrast, our study develops a retrospective, data driven DT prototype that simulates disease risk and intervention outcomes using behavioral, dietary, and psychosocial variables, without requiring real-time inputs. This enhances scalability and enables deployment in community level prevention programs.

2.2 Digital twins for comorbidity management

DT applications are also expanding to address multimorbidity, particularly in patients with T2DM and related conditions like hypertension and cardiovascular disease. Shamanna et al. [38] showed that DT-enabled systems reduced antihypertensive medication use while improving metabolic and cardiovascular markers. However, like their earlier platforms, these systems remain tied to real-time biosensor data and frequent clinical engagement, limiting broader applicability.

2.3 Offline and simulation-driven DT models

Recent work has explored DT frameworks that do not require real-time monitoring. For example, Silfvergren et al. [39] modeled glycemic responses to macronutrient intake using trial data, and Vaskovsky et al. [40] developed an adaptive food recommendation system based on genetic predisposition. While promising, these studies focus primarily on nutrition and do not fully incorporate behavioral simulation, causal reasoning, or long-term disease forecasting.

2.4 Ethical, regulatory, and interpretability considerations

As DTs become more prevalent in health systems, concerns around transparency, explainability, and data governance grow. Prior research [41] stresses the importance of interpretability in digital health. Our study addresses this by using transparent statistical and causal inference methods that yield clinically meaningful, traceable outputs. This supports ethical alignment and enhances trust in clinical and policy decision-making.

2.4.1 Research gap

2.4.2 Key contributions

To address the identified gaps, this study makes the following contributions:

3 Proposed digital twin framework

Figure 1

4 Research methodology

Figure 2

5 Dataset preparation

The foundational dataset was sourced from the UK Biobank [28], a large-scale, prospective cohort of over 500,000 participants aged 40–69 at baseline. It includes extensive behavioral, psychosocial, demographic, and clinical variables such as validated measures of depression, insomnia, loneliness, self-reported mental health, physical activity, sleep, and diet, critical for modeling behavioral pathways and simulating intervention outcomes.

To construct a cohort suitable for time-to-event analysis, participants were conceptually separated into two groups based on disease status during follow-up: individuals who developed T2DM after baseline assessment (T2DM cohort) and individuals who remained non-diabetic throughout the observation period (Healthy cohort). The full workflow is illustrated in Figure 3.

Figure 3

5.1 Healthy stream

For the healthy stream, individuals with any recorded chronic disease at baseline were excluded using linked clinical records and self reported diagnoses from the UK Biobank. Disease status was identified using field 41,270, which captures ICD-10 coded diagnoses reported across hospital records and participant medical history. This process excluded all participants with T2DM and other major long term conditions. Applying this criterion removed 463,550 participants, leaving 36,450 individuals free of diabetes and other chronic diseases at baseline, as shown in the left-hand branch of Figure 3. These participants formed the healthy cohort prior to follow up verification and further data quality filtering.

This exclusion strategy was designed to ensure that observed behavioral, dietary, and psychosocial differences could be attributed specifically to diabetes rather than to the presence of multiple chronic conditions. Non-diabetic illnesses, including cancer, cardiovascular disease, kidney disease, and chronic respiratory disorders, are known to independently alter lifestyle behaviors and psychological health through disease related symptoms, functional limitations, and treatment burden [42–44]. Retaining participants with such conditions would therefore introduce confounding effects unrelated to diabetes itself, complicating the interpretation of diabetes specific behavioral patterns and reducing internal validity [45, 46]. Restricting the healthy cohort to participants without chronic disease thus provided a robust reference group for examining behavioral changes associated with the onset of T2DM.

5.2 T2DM stream

For the T2DM stream, participants were identified based on the presence of at least one ICD-10 diagnosis code E11.0 to E11.9 in linked hospital inpatient records. This yielded 56,115 individuals with a recorded diagnosis of T2DM. Participants without any T2DM diagnosis, approximately 443,885 individuals, were excluded from this stream. To reduce confounding from non diabetes related illnesses, individuals with major chronic comorbid conditions were subsequently excluded, resulting in a final T2DM cohort of 33,957 participants prior to missing data and outlier exclusion, as illustrated in the right-hand branch of Figure 3.

Eligibility further required the availability of sufficient follow-up information to define either an event time or a censoring time. Participants who remained free of T2DM throughout follow-up were treated as censored observations. For these individuals, follow-up time was defined from the baseline assessment to the most recent subsequent assessment or linked record confirming continued non-diabetic status, as indicated by UK Biobank field 53. Participants without any follow-up assessment or linked hospital record confirming non-diabetic status were excluded due to the inability to define a valid censoring time. After applying this criterion, 14,927 healthy participants had valid follow-up time.

For participants who developed T2DM during follow-up, incident disease was identified using ICD-10 codes E11.0–E11.9 recorded in UK Biobank linked hospital inpatient data [47]. The date of first recorded diagnosis, captured in UK Biobank field 130,708, was used to define the event time. Follow-up time for these individuals was calculated as the interval between baseline assessment and this first diagnosis date, provided the diagnosis occurred after baseline. This approach yields valid survival times even for individuals without repeat assessment visits, as diagnosis dates are obtained through hospital linkage. After applying this rule, 14,392 T2DM participants had a valid event time. Together, these definitions ensured consistent and unbiased estimation of time-to-event outcomes across the cohort. Related methodological principles for cohort construction and survival eligibility using UK Biobank data have been described in our prior work [48] and are referenced here for methodological context, while all dataset-specific decisions are fully documented in the present study. The resulting sample size and number of observed events were sufficient to support stable time-to-event modeling relative to the final number of predictors included.

6 Data preprocessing and feature selection

To support the objectives of the Digital Twin framework, clinical biomarker variables such as blood glucose, glycated hemoglobin, and cholesterol were excluded from the predictor set. While these measures are clinically informative and commonly used in diabetes risk assessment [49], their inclusion would anchor the model to metabolic abnormalities that typically emerge later in the disease process. In contrast, the present study focuses on early risk forecasting and the simulation of preventive interventions using modifiable lifestyle and psychosocial factors that are observable prior to routine laboratory abnormalities.

After defining the predictor scope, data preprocessing steps were applied to the 14,927 healthy records and 14,392 T2DM records to ensure analytical stability and reproducibility within the Digital Twin framework. Categorical responses such as “Don’t know” and “Prefer not to answer” were recoded as missing values (NaN). Records with more than 20% missing data were excluded to balance data completeness with sample retention and to limit instability arising from extensive imputation, consistent with common practice in epidemiological and machine-learning analyses [50, 51]. Applying this threshold removed 2,350 records from the healthy group, leaving 12,557 individuals, and excluded 2,700 records from the T2DM group, resulting in 11,692 participants.

For the remaining observations, missing values were imputed using mode imputation for categorical variables, mean imputation for continuous variables, and the least frequent category for binary fields. Negatively coded responses (e.g., “” indicating “None of the above”) were recoded to preserve logical and ordinal structure across variables. Following missing data handling and imputation, additional preprocessing steps were undertaken to address structurally inconsistent records.

6.1 Outlier handling and detection

Outliers can distort multivariate relationships in behavioral health data and bias survival estimates. In this study, rather than adjusting extreme values, which can alter inter-variable dependencies and obscure higher-order behavioral structure [52–54], entire records flagged as structurally inconsistent were excluded. This full-case removal strategy preserves coherence across behavioral features and supports stable estimation in Cox proportional hazards models. Outlier detection was performed after missing-data filtering and prior to any supervised modeling to prevent data leakage and ensure valid survival inference.

6.1.1 Outlier detection pipeline

A multi-step unsupervised approach was used to detect and remove structurally inconsistent data points, ensuring a robust dataset for analysis while balancing generalizability and internal validity.

• Data standardization: All features were standardized to a mean of zero and standard deviation of one to ensure equal contribution, especially for algorithms like t-distributed Stochastic Neighbor Embedding (t-SNE) [55], which are sensitive to feature scale.

• Optimal cluster selection via Silhouette score: Silhouette analysis [56] was used to evaluate clustering quality, with the highest score at , indicating optimal cluster separation (see Figure 4a).

• t-SNE visualization:t-SNE was applied to project the dataset into two dimensions using a perplexity of 40 and 5,000 iterations. It was chosen over principal component analysis (PCA) [57] because it more effectively preserved local neighborhood structures in the data used in this study, which facilitated clearer cluster separation and outlier detection. The resulting two-dimensional projection revealed two distinct clusters corresponding to the binary target variable, along with peripheral points that may represent outliers (see Figure 4b).

• Hierarchical Clustering: Ward’s linkage method [58] was used on the t-SNE components, producing well-separated clusters with minimized intra-cluster variance (see Figure 4c).

• Outlier identification and removal:

  Outliers were identified based on their Euclidean distance from the centroid of their respective clusters:

  • Centroid calculation: Centroids of each cluster in t-SNE space were computed.

  • Distance measurement: Euclidean distances between each data point and its cluster centroid were calculated.

  • Thresholding: Points beyond the 95th percentile of these distances were flagged as outliers.

  • Visualization and removal: Outliers were marked in red in Figure 4d and subsequently removed to avoid skewing the analysis.

Figure 4

This systematic pipeline ensured the removal of structurally inconsistent data points, resulting in improved cluster cohesion and enhanced analytical integrity. While this approach enhances internal validity and model fit, it does come with a trade-off in terms of external generalizability. Removing extreme data points may limit the representativeness of certain subpopulations. However, for time-to-event modeling and simulation purposes, this trade-off was considered acceptable to improve robustness and reduce model bias. Applying the pipeline removed an additional 11,327 outliers from the healthy cohort and 1,805 from the T2DM cohort, reducing the samples from 12,557 to 1,230 healthy participants and from 11,692 to 9,887 T2DM participants. These constituted the pre balanced cohorts used for subsequent survival analysis.

6.1.2 Final study cohort

After validating follow-up times and completing missing data and outlier filtering, the Healthy and T2DM cohorts were advanced to the final stage of cohort construction. The healthy group remained moderately larger than the T2DM group. Although Cox proportional hazards models do not require balanced outcome groups, random down-sampling was applied to achieve matched sample sizes for computational balance and interpretability, without conditioning on predictor variables. This procedure altered the observed prevalence of T2DM but does not bias Cox model estimation, as inference is driven by time-to-event ordering and risk sets rather than marginal outcome proportions [59–61]. Key model estimates and performance metrics were consistent with analyses conducted on the full cohort, indicating robustness to this design choice. As shown in Figure 3, the final analytic dataset included 19,774 individuals, with 9,887 participants who were non-diabetic at baseline and subsequently developed T2DM during follow up (T2DM cohort), and 9,887 participants who remained non-diabetic across all available assessments (Healthy cohort). Time-to-event durations ranged from 1 to 17 years, providing sufficient longitudinal variability for robust and unbiased survival analysis. All preprocessing thresholds and parameter settings were specified a priori and applied uniformly, with no tuning based on outcome information.

7 Survival modeling and risk identification

This section outlines the Cox modeling pipeline, covering variable screening, multicollinearity control, proportional hazards (PH) testing, and model validation, followed by interpretation of significant predictors forming the basis for simulation and intervention in the DT system.

7.1 Cox model development and validation

A penalized Cox proportional hazards model [29] was employed to estimate T2DM onset risk. The model simultaneously identified significant predictors and generated time-sensitive risk estimates, enabling personalized simulation. The hazard function is given in Equation 1:where is the hazard at time , is the baseline hazard, and represents the hazard ratio (HR) for predictor .

7.1.1 Univariate analysis

Before building the multivariate survival model, each predictor was first evaluated individually to screen for potential associations with T2DM onset. To do this, a univariate Cox proportional hazards regression was conducted for each variable in the dataset. This process involves modeling the time to T2DM onset as a function of a single predictor at a time, allowing for an independent assessment of each variable’s association with the outcome. It should be noted that univariate screening was used solely as an initial dimensionality reduction step and not as a criterion for final variable selection, which will be determined in the multivariate modeling stage.

7.1.1.1 Variable encoding and significance testing

Variable types were handled according to their characteristics to ensure proper integration into the Cox regression model. Categorical predictors were encoded using dummy variables, with one category designated as the reference group and the remaining categories represented by binary indicators. Effects were interpreted relative to the reference category, and categorical predictors were retained if at least one associated dummy variable demonstrated statistical significance (). Continuous variables were entered in their original numerical form to capture incremental effects on risk over time, while binary variables were included directly without further transformation.

7.1.1.2 Variance filtering and multicollinearity control

7.1.2 Multivariate analysis

After identifying significant predictors through univariate analysis and addressing potential issues of multicollinearity and low variance, a multivariate Cox proportional hazards model was applied. Multivariate analysis is essential for understanding how multiple predictors collectively influence the time to the onset of T2DM. Unlike univariate analysis, which evaluates the relationship between each predictor and the outcome in isolation, the multivariate approach estimates the independent effect of each predictor while controlling for the influence of all other variables in the model.

The 28 features identified in univariate analysis were included in the initial multivariate model to assess their combined impact on T2DM onset. However, some variables lost their statistical significance in the multivariate context due to confounding or shared variance with other predictors. This adjustment reflects the model’s ability to isolate the unique contribution of each variable while accounting for correlations among predictors.

To refine the model, an iterative process was employed, sequentially removing non-significant variables and re-estimating the model until only predictors with statistically significant and independent associations remained. This procedure ensured that the final model retained only those features that robustly contributed to T2DM risk. Through this process, the feature set was reduced from 28 to 14 predictors.

The coefficients from the multivariate Cox model represent adjusted hazard ratios, quantifying the relative risk of T2DM onset associated with each predictor while holding other factors constant. For example, a dietary factor with a hazard ratio greater than 1 indicates an increased risk of T2DM with higher intake, independent of other lifestyle or demographic variables.

This stepwise refinement improves both the interpretability and reliability of the model, allowing for a focused interpretation of key predictors that significantly influence T2DM development.

7.1.3 Proportional hazards assumption

Following multivariate modeling, the proportional hazards (PH) assumption was evaluated to ensure the validity of the Cox regression framework. This assumption requires that the effect of each covariate on the hazard remains constant over time, implying that hazard ratios are time-invariant. For example, if higher physical activity reduces diabetes risk by 30% (hazard ratio 0.70), this relative effect is expected to persist throughout the follow-up period.

The PH assumption was assessed using diagnostic methods based on Schoenfeld residuals [63]. Violations can bias hazard ratio estimates; therefore, corrective strategies were applied when necessary. Two approaches were implemented: (i) stratification, allowing the baseline hazard to vary across covariate strata, and (ii) incorporation of time-dependent covariates to capture changing effects over time. In this study, non-proportionality was detected for cooked vegetable intake and long-standing illness, both of which were modeled with covariate–time interactions to preserve validity and interpretability.

These adjustments enabled the model to accommodate dynamic predictor effects, thereby enhancing explanatory accuracy while preserving the validity of the PH assumption.

7.1.4 Internal validation

Model performance was evaluated using 10-fold stratified cross-validation, training on nine folds and testing on the tenth [64]. Stratification preserved the distribution of diabetes onset and censored cases across folds. L2 regularization reduced overfitting by penalizing large coefficients [65]. Performance was measured with the concordance index (C-index) [66], yielding a mean of 0.90 (SD 0.004), indicating excellent accuracy and consistency. Low variability in performance across folds indicates that results are robust to different training and testing splits. The model was then refitted on the full dataset to maximize information, and the proportional hazards assumption was re-verified, confirming the stability and reliability of the final model.

8 Results and interpretation of significant predictors

This section presents the findings from the multivariate Cox proportional hazards model and discusses their implications for disease prediction, ethnic disparities, and behavioral intervention within the digital twin simulation framework.

The multivariate Cox model identified a set of psychosocial, behavioral, dietary, and demographic factors that significantly influence T2DM onset. Table 2 summarizes the selected predictors along with their coefficients (), hazard ratios (exp()), percentage change in hazard (HR%), confidence intervals, and -values. HR% was calculated as , and values are rounded consistently. These findings underscore the importance of modifiable non-clinical determinants alongside demographic and ethnic disparities, providing critical inputs for the digital twin simulation framework.

To facilitate clear interpretation of regression coefficients, all predictors were coded in alignment with their original UK Biobank definitions. Categorical variables were converted into binary or dummy variables as appropriate, allowing for consistent and interpretable estimation within the Cox proportional hazards model. Psychosocial, sleep, and behavioral factors, such as loneliness, difficulty waking in the morning, and insomnia, were defined as binary exposure variables.

Salt added to food, cheese intake, processed meat intake, and plays computer games were modeled as ordinal predictors to capture graded, dose–response associations with T2DM risk. For salt added to food, category-specific effects were similar and often non-significant when modeled separately, so the variable was retained as a single ordinal term. Processed meat intake, cheese intake, and plays computer games showed a monotonic increase in hazard across ordered frequency categories and was therefore also included as an ordinal variable.

“Sugary cereals” was defined as a binary variable, with a value of 1 indicating consumption of sugary or processed breakfast cereals (e.g., Cornflakes, Frosties). “Mental health” was also modeled as a binary variable, indicating whether an individual reported a history of anxiety or depression and had consulted a medical professional for these conditions.

BMI was categorized according to World Health Organization criteria [67]. Participants classified as underweight (BMI kg/m²) were excluded because of small numbers and unstable estimates. The normal-weight category (BMI 18.5–24.9 kg/m²) was used as the reference group in all BMI analyses. Age was specified using UK Biobank–defined decile groups (40–49, 50–59, and 60–70 years), with ages 40–49 years serving as the reference category. Modeling age categorically permits flexible estimation of age-related risk without assuming a linear relationship. Ethnicity was included as a categorical variable using dummy coding, with White ethnicity specified as the reference group.

8.1 Psychosocial and behavioral factors

Several psychosocial and lifestyle related variables were independently associated with the hazard of developing T2DM. Loneliness and social isolation were associated with a 24% higher hazard (HR 1.24, 95% CI: 1.17–1.32, ), while insomnia was associated with a 12% increase in hazard (HR 1.12, 95% CI: 1.07–1.19, ). Individuals who had consulted a doctor for anxiety or depression experienced a 29% higher hazard of T2DM onset (HR 1.29, 95% CI: 1.21–1.39, ). Modest but significant associations were also observed for leisure behaviors. Playing computer games (HR 1.05, 95% CI: 1.01–1.10, ) and difficulty getting up in the morning (HR 1.04, 95% CI: 1.01–1.07, ) both elevated risk, potentially reflecting sedentary patterns, circadian misalignment, or early markers of metabolic dysfunction. Collectively, these findings underscore that psychosocial distress, disrupted sleep, and subtle lifestyle rhythms are not only quality of life issues but also measurable biological risk factors. Incorporating these predictors into digital twin models allows simulation of mental health and behavior driven trajectories, supporting the design of personalized preventive interventions.

8.2 Dietary habits

Several dietary variables were independently associated with the hazard of T2DM onset. Higher processed meat intake (HR 1.05, 95% CI: 1.02–1.07, ), adding salt to food (HR 1.08, 95% CI: 1.05–1.11, ), and consumption of sugary cereals (HR 1.17, 95% CI: 1.11–1.23, ) were each associated with an increased hazard of developing T2DM, consistent with proposed mechanisms involving inflammation, insulin resistance, and gut microbiome disruption. Interestingly, cheese intake demonstrated a statistically significant inverse association with the hazard of T2DM onset (HR 0.93, 95% CI: 0.91–0.95, ). While this finding contrasts with conventional dietary guidance that often cautions against high fat dairy consumption, emerging evidence indicates that certain dairy products may exert protective metabolic effects through improvements in insulin sensitivity and reductions in systemic inflammation [68–70]. One plausible biological mechanism involves the presence of vitamin K, particularly vitamin K2, in cheese. Vitamin K2 has been associated with enhanced insulin sensitivity and a lower risk of T2DM, potentially mediated through its role in osteocalcin metabolism and downstream glucose regulation [71]. Within the digital twin framework, these dietary factors constitute modifiable causal nodes and represent high leverage targets for simulating dietary intervention scenarios.

8.3 Age and BMI

Compared with adults aged 40–49 years, individuals aged 60–70 years exhibited a substantially higher hazard of T2DM onset (HR 1.69, 95% CI: 1.40–2.04, ), whereas those aged 50–59 years did not differ significantly from the reference group (HR 0.93, 95% CI: 0.77–1.12, ). Body mass index showed a clear dose-response relationship when classified according to World Health Organization categories [67]. Overweight status was not significantly associated with the hazard of T2DM (HR 1.11, 95% CI: 0.92–1.34, ), while obesity class I (HR 1.75, 95% CI: 1.45–2.11, ) and obesity class II (HR 2.14, 95% CI: 1.77–2.60, ) were associated with progressively higher hazards of disease onset. These associations reaffirm excess weight and later life as critical accelerators of T2DM risk, consistent with evidence linking obesity to insulin resistance, systemic inflammation, and beta-cell dysfunction [72].

8.4 Ethnic disparities

Additionally, the elevated hazards observed among mixed ethnicity groups point to a complex interplay of genetic, behavioral, and environmental determinants. This highlights the need for more granular investigations into mixed ethnicity populations, which are currently underrepresented in metabolic and epidemiological research [76].

Collectively, these predictors highlight how psychosocial stress, adverse diet, excess weight, and ethnic background converge to shape T2DM trajectories. For the digital twin framework, they provide both predictive power and actionable levers for simulated interventions.

8.4.1 Risk score stratification and validation

Following development and internal validation of the multivariable Cox proportional hazards model, an individual risk score was computed for each participant as the linear predictor from the final fitted model. Regression coefficients represent covariate-specific contributions to the hazard, as shown in Equation 3:where denotes the estimated coefficient for predictor and is the corresponding value for participant . Categorical predictors were encoded using dummy variables, with reference categories defined in Table 2. The resulting score represents a log relative hazard summarizing the combined effects of behavioral, psychosocial, sleep-related, dietary, and demographic factors.

Once the risk scores were calculated, a decision tree classifier was trained to stratify individuals based on predicted risk, using the risk score as input and event occurrence as the target. This supervised binning approach identifies optimal split points to maximise outcome separation, following principles of supervised discretisations introduced by Fayyad and Irani [77]. The tree was constrained to five leaf nodes, identifying optimal thresholds that best separated individuals by outcome. These thresholds defined five discrete, non-overlapping risk groups: Very Low (1.67, 0.22), Low (0.23, 0.40), Moderate (0.41, 0.70), High (0.71, 1.11), and Very High (1.12, 2.75).

To illustrate how individual predictors contribute to risk stratification, two examples are provided. An individual reporting loneliness or social isolation and insomnia , with all other predictors at their reference levels, has a linear predictor of , which corresponds to the Low risk group. In contrast, an individual reporting loneliness, insomnia, regular consumption of sugary cereals, belonging to the 60–70 year age group, and obesity class I, with all remaining predictors at their reference levels, has a linear predictor of , placing them in the Very High risk group. These examples illustrate how specific combinations of predictors contribute additively to the overall risk score and determine risk group membership.

The discriminative performance of the resulting risk stratification was evaluated using survival analysis. Kaplan–Meier curves [78] were used to estimate diabetes-free survival across risk groups, while pairwise log-rank tests [79] assessed statistical differences between survival distributions. The Kaplan–Meier estimator accounts for censoring by incorporating information up to each participant’s last observed time point, enabling unbiased estimation of cumulative survival.

As shown in Figure 5, the resulting curves demonstrated clear and progressive separation across strata, with higher risk groups exhibiting earlier and steeper declines in diabetes free survival. Vertical dashed lines indicate the estimated 25%, 50%, and 75% survival percentiles for each group, illustrating systematic shifts toward shorter survival times with increasing risk. Pairwise log rank tests further confirmed that differences between groups were statistically significant (, Table 3a), demonstrating that the stratification captures both major and incremental variation in diabetes onset risk. To provide interpretable time-based summaries, survival percentiles were also computed (Table 3b), revealing a monotonic shift toward earlier diabetes onset with increasing risk category. For instance, individuals in the Very High group reached 25% incidence by year 4 and 75% by year 11, whereas the Very Low group crossed the 25% threshold only after 13 years and did not reach higher incidence levels during follow-up. For the very low risk group, the median (50%) and 75% survival times were not observed during follow up and are therefore reported as , indicating that fewer than 50% and 75% of participants experienced the event within the observation window.

Figure 5

Taken together, this stratification highlights the practical utility of the digital twin framework beyond statistical validation. By translating continuous outputs from the Cox model into discrete, time-sensitive categories, the system provides thresholds that can inform clinical decision-making and public health planning. These findings show that psychosocial and behavioral risk factors can be structured into groups that are both statistically distinct and clinically meaningful, reinforcing the value of this framework for individualized prevention and population-level screening.

9 Causal inference and simulation of intervention effects

While Cox models identify statistical associations, they do not establish whether modifying risk factors changes outcomes. This limitation is crucial for a DT, where simulating counterfactual scenarios (e.g., “What if loneliness were reduced?”) is essential. To address this, a causal inference framework was integrated, enabling estimation of intervention effects on T2DM progression. By quantifying the impact of modifiable variables such as sleep, diet, and psychosocial stress, the DT supports modeling of how risk may be reduced rather than only predicting who is at risk.

The framework proceeded in two stages. First, a domain informed Directed Acyclic Graph (DAG) was constructed to encode expert knowledge about hypothesized causal pathways between variables. (Figure 6). Second, causal effects were estimated via propensity score matching, backdoor adjustment, and robustness checks. This approach shifts the DT from correlation-based prediction to intervention-oriented simulation.

Figure 6

9.1 Stage 1: construction of DAG

To ensure transparency in the causal modeling, a detailed rationale is provided for the relationships embedded in the expert-driven DAG (Figure 6). Each directed edge in the graph reflects prior knowledge derived from epidemiological and clinical literature, linking psychosocial, dietary, behavioral, and demographic factors to the risk of developing T2DM. The following section presents the justification for each major domain included in the DAG.

9.1.1 Age, BMI, and metabolic pathways

Age was modeled as an upstream determinant shaping BMI, sleep, diet, long-standing illness, and T2DM, reflecting its role in multimorbidity, sleep disruption, and dietary shifts [80, 81]. BMI was positioned centrally as a mediator between psychosocial and dietary inputs and diabetes risk. It is influenced by diet and stress, interacts bidirectionally with sleep, and, when elevated, drives insulin resistance through impaired glucose regulation [82–84].

9.1.2 Dietary influences

Dietary habits were modeled as influencing both BMI and T2DM directly. High intake of sugary or processed cereals is associated with weight gain and increased T2DM risk [85]. In contrast, cheese intake may be protective due to its protein and fat content, which can enhance satiety and insulin sensitivity [69, 86], and potentially through its vitamin K2 content [71]. Processed meats are linked to both elevated BMI and direct metabolic disruption via pro-inflammatory effects [87]. Salt addition was included due to evidence that it may promote weight gain through mechanisms such as fluid retention and altered taste perception leading to higher energy intake [88].

9.1.3 Psychosocial and sleep-related factors

Loneliness and social isolation were modeled as upstream drivers influencing diet, sleep, and BMI [89]. Psychiatric history was linked to disordered eating, poor sleep, elevated BMI, and increased T2DM risk [90]. Sleep disturbances were incorporated as both causes and consequences of metabolic dysfunction: insomnia affects cortisol, appetite, and glucose metabolism, while difficulty waking was treated as a proxy for circadian disruption [91, 92].

9.1.4 Ethnicity and physical health conditions

Ethnicity was modeled as a root variable shaping dietary patterns, BMI, mental health, and T2DM risk, reflecting disparities in prevalence, cultural stressors, and lifestyle determinants [93–95]. Long-standing illness was included as a driver of mental health problems, disrupted sleep, altered diet, elevated BMI, and direct metabolic vulnerability, consistent with evidence on multimorbidity and heightened T2DM risk [96–98].

Together, these justifications ensured the DAG reflected plausible causal pathways, forming the foundation for intervention simulation.

9.2 Stage 2: causal inference and simulation of intervention effects

To move beyond association and simulate actionable interventions, causal inference methods within the potential outcomes framework were applied. The Average Treatment Effect on the Treated (ATT) [99] was estimated for key modifiable factors such as BMI and psychosocial stressors. Treatments were operationalised as binary exposures based on clinically meaningful thresholds (e.g., obesity defined as BMI  30 kg/m², and psychosocial stressors defined as presence or absence of loneliness, insomnia, or mental health consultation). This approach was chosen to approximate the effects of real-world interventions in an observational dataset, where randomized trials are not feasible. This stage involved three methodological steps: propensity score matching, regression-based backdoor adjustment, and robustness checks.

9.2.1 Propensity score matching and ATT

The ATT estimation measures the causal effect of a treatment or condition (e.g., having obesity) on those who actually experienced it. Since this is an observational dataset, confounding variables may bias naive comparisons. To address this, a 2-to-1 propensity score matching algorithm was used, pairing each treated individual with two controls with similar covariate profiles. A propensity score reflects the conditional probability of receiving the treatment given observed covariates. Matching individuals with similar scores balances the distribution of confounding variables between treated and control groups. This mitigates selection bias and allows for fairer estimation of causal effects [100].

9.2.2 Standardized mean differences

To ensure the quality of covariate balance, Standardised Mean Differences (SMDs) were computed before and after weighting for each confounder [101]. This diagnostic helps verify that the re-weighted treatment and control groups are statistically comparable on the observed covariates, which is a prerequisite for valid causal inference.

9.2.3 Backdoor adjustment and regression modeling

Once matching was performed and covariate balance confirmed, the ATT was estimated using DoWhy’s regression-based backdoor adjustment [30]. By conditioning on a suitable set of covariates, the backdoor criterion ensures that non-causal pathways between the treatment and outcome are blocked, allowing identification of the causal effect.

Regression was then used as the estimation procedure to implement this backdoor adjustment on the matched dataset, further reducing residual imbalance and improving statistical precision. This combined approach, matching followed by regression-based backdoor adjustment, reduces model dependence and enhances robustness. The resulting ATT estimates reflect the isolated causal impact of treatment on the outcome among the treated population.

9.2.4 Robustness checks

9.2.4.1 Example of the causal estimation workflow

Consider estimation of the causal effect of obesity (BMI 30 kg/m²) on the risk of T2DM. Individuals classified as obese formed the treated group, while non-obese individuals served as potential controls. Propensity score matching was first applied to pair each obese individual with two non-obese individuals (2-to-1 propensity score matching) who had similar values of age, ethnicity, long-standing illness, and relevant dietary indicators. Covariate balance between the treated and matched control groups was then assessed using SMD, and only matched samples achieving acceptable balance were retained for causal estimation.

After covariate balance was confirmed, regression-based backdoor adjustment was applied to the matched dataset, conditioning on the covariates identified by the causal DAG to block all backdoor paths between obesity and T2DM. The resulting ATT represents the estimated difference in T2DM risk that obese individuals would experience under a counterfactual scenario in which they were not obese, holding observed confounders constant. This estimate was then subjected to the refutation and sensitivity analyses described above to evaluate the robustness of the causal effect.

10 Results and discussion from causal modeling

Importantly, this causal analysis builds upon the predictive foundation established in Stage 1, which used multivariate Cox modeling to identify baseline predictors of T2DM onset. Stage 2 advances from prediction to simulation by applying formal causal inference methods (e.g., propensity score matching and backdoor regression adjustment) to estimate counterfactual effects, with effects reported on the probability scale where indicated (ATT). All findings were rigorously validated using refutation strategies, including placebo testing, subset evaluation, bootstrap resampling, and sensitivity analyses. Confounders for all paths were selected a priori from the DAG (age, ethnicity, long-term condition, and diet indicators where relevant). Diet variables were treated as pre-exposure preferences to reduce residual confounding; outcome models for diet mediators also adjusted for the exposure and other diet indicators.

10.1 Q1: effects of BMI on psychosocial stressors and T2DM mediation

This analysis examines how BMI categories interact with psychosocial stressors to shape diabetes risk within the study cohort. Individuals were classified as normal weight (18.5–24.9 kg/m²), overweight (25.0–29.9 kg/m²), or obese (30 kg/m²), with each category coded as binary (1 within range, 0 outside range). Since the dataset does not include underweight individuals (BMI < 18.5 kg/m²), these binary comparisons focus exclusively on normal weight, overweight, and obesity. Stability of the estimated effects within this cohort was supported by robustness checks, strengthening confidence in their internal consistency.

Causal mediation analysis suggests that stress-related factors, particularly insomnia, loneliness, and poor mental health, partly explain the association between BMI and T2DM risk. Effects are estimated on the probability scale using ATT; the direct effect is a controlled direct effect (BMI’s impact on T2DM when the mediator is held fixed), and indirect effects are computed as (predictor–mediator effect multiplied by mediator–outcome effect). The total effect is the sum of direct and indirect components.

10.1.1 Direct effects of BMI on T2DM risk

As shown in Table 4, in this analysis BMI exhibited strong and statistically robust controlled direct effects on T2DM risk. Normal weight was protective, ATT (95% CI: to ), corresponding to a 24.4 percentage-point lower absolute risk. Overweight increased risk, ATT 0.1229 (95% CI: 0.1114 to 0.1333), a 12.3 percentage-point higher risk, while obesity further elevated risk, ATT 0.3493 (95% CI: 0.3376 to 0.3608), a 34.9 percentage-point higher risk. These direct effects accounted for approximately 95–99% of the total estimated effect, indicating that BMI explains most of the modeled association with T2DM risk. Psychosocial mediators contributed only a small fraction on the absolute scale, though still detectable and meaningful. Robustness checks supported the stability of these estimates, strengthening confidence in their interpretation as actionable insights.

10.1.2 Psychosocial mediation pathways: insomnia, loneliness, and mental health

Table 4 shows that indirect effects were modest but consistent on the absolute scale. For insomnia, the indirect effect was 0.36 percentage points at normal BMI, 0.19 percentage points at overweight, and 1.15 percentage points at obesity. Loneliness was the largest pathway: 0.40 percentage points at normal BMI, 0.36 percentage points at overweight, and 1.18 percentage points at obesity. Mental health contributed the smallest indirect effects: 0.15 percentage points, 0.21 percentage points, and 0.92 percentage points across normal, overweight, and obesity, respectively. These magnitudes increase with BMI, consistent with a dose–response pattern in which higher BMI is associated with amplified psychosocial and metabolic risk.

10.1.3 Refutation and robustness tests

Across all BMI groups and mediators, robustness checks, including placebo, subset, bootstrap, and sensitivity analyses, consistently validated the causal estimates. Placebo effects were negligible (-values > 0.84), subset and bootstrap tests showed near-identical estimates to primary models (differences <1%), and sensitivity analyses confirmed that indirect effects remained stable even when accounting for potential unmeasured confounding. These sensitivity entries reflect robustness checks from DoWhy refuters (placebo, subset, bootstrap, unobserved common cause), not conventional confidence intervals. These robustness tests collectively support the internal reliability of the estimated direct and indirect effects within the specified causal model.

10.1.4 Discussion

10.1.5 Implications for clinical practice, and public health

10.1.6 Implications for policymakers

10.2 Q2. Individual psychosocial stressors, diet, and T2DM risk

This analysis assessed insomnia, loneliness, and poor mental health as psychosocial stressors influencing T2DM risk via direct and dietary-mediated pathways. Processed meat intake was dichotomized as high if the reported frequency was 3 on the original scale. Cheese intake was dichotomized as high if the reported frequency was 4. Salt added to food was dichotomized as high if the reported frequency was 3. Effects are estimated on the probability scale using the Average Treatment Effect on the Treated (ATT). Direct effects are controlled direct effects (conditioning on the dietary mediator), indirect effects are computed as (exposure mediator effect multiplied by mediator outcome effect), and total effects are the sum of direct and indirect components. All mediators are binary; effects are interpreted as percentage-point changes in absolute risk. In this study, positive indirect effects indicate higher T2DM risk transmitted through the mediator. Negative indirect effects indicate risk reduction through the mediator within the modeled counterfactual framework. Importantly, when the exposure reduces a protective food and that food is associated with lower T2DM risk, the indirect effect will be positive, reflecting higher diabetes risk despite lower consumption of the protective food.

As summarized in Tables 5, 6, direct effects were large and broadly similar across models, with minor variation because the controlled direct effect is estimated within each mediator model. For insomnia, the direct ATT ranged from 0.306 to 0.366 across mediators. For loneliness, the direct ATT clustered tightly around 0.363–0.365. For mental health, it was 0.349–0.350. In short, within this study population and under the modeled assumptions, each stressor was associated with an estimated increase in absolute T2DM risk of approximately 35 percentage points across models, regardless of the diet mediator chosen.

On the other hand, dietary mediation provided a smaller, reinforcing contribution within the estimated causal pathways. For instance, for insomnia the largest single mediated path was via salt (indirect ), followed by processed meat () and sugary cereals (); the cheese path was near zero and directionally harmful, reflecting reduced intake of a protective food under stress. For loneliness, processed meat and sugary cereals each mediated , salt was negligible (), and cheese again showed a near-zero, directionally harmful path for the same reason. For mental health, salt was the main mediator (), with smaller contributions from processed meat () and sugary cereals (), and cheese followed the same minimal, harmful trend.

Summing across mediators, indirect effects through diet mediated about 0.5–1.3 percentage points of absolute risk for each stressor (roughly 1%–4% of the corresponding total effect), illustrating a consistent pattern within the observed data by which psychosocial stress is associated with unhealthier eating behaviors. Greater consumption of processed, salty, and sugary foods, together with lower intake of protective foods such as cheese, appears to further reinforce the estimated diabetes risk associated with psychosocial stressors within this cohort. The robustness of both direct and mediated pathways was confirmed through placebo, subset, bootstrap, and sensitivity refutation tests, with generally high -values (mostly above 0.84, a few between 0.66 and 0.83), supporting the internal stability of these findings under the specified modeling assumptions.

10.3 Q3. Combined psychosocial stressors and T2DM risk

In real-world settings, psychosocial stressors rarely occur in isolation. Many individuals experience insomnia, loneliness, and poor mental health simultaneously, amplifying T2DM risk. To capture this clustering, a binary combined-stressor variable was defined, coded 1 when psychiatrist contact, loneliness, and insomnia co-occurred and 0 otherwise. The reported direct effect is the joint controlled direct effect (ATT) of this cluster, rather than the sum of individual stressor effects. As summarized in Table 7, the clustering of these stressors produced an ATT for the direct effect of 0.7796 (95% CI: 0.7457 to 0.8135), i.e., a 77.96 percentage-point increase in absolute risk; the corresponding total effect ATT was 0.8044. This direct effect is more than double the 0.35–0.37 risk differences (35–37 percentage points) linked to individual stressors. Balance and overlap were satisfactory: propensity score distributions showed good common support, all fitted risks lay within [0,1], and post-matching covariate balance met standard thresholds (all absolute standardized mean differences ). These results indicate that, within this study population, co-occurring stressors are associated with substantially higher estimated diabetes risk than any single stressor alone.

Dietary mediation provided a smaller, reinforcing contribution. The largest single mediated path was via salt (indirect , 3% of that model’s total effect, 0.8044). Sugary cereals and processed meat each added modest mediation (indirect and , 1%–2% of their respective totals: 0.7916 and 0.7904). The cheese pathway was negligible and directionally harmful, reflecting reduced intake of a protective food (indirect ; total effect 0.7798). In short, within the modeled framework, clustered psychosocial stressors are associated with large direct increases in estimated T2DM risk, with diet contributing additional but comparatively small mediated increments.

Although indirect effects through diet explained only about 4.8 percentage points of the total risk (approximately 6%), their consistency underscores that dietary change is a meaningful pathway linking psychosocial stress with metabolic outcomes. Under the modeled counterfactual scenarios, this 2.5–4.8 percentage point diet-mediated component corresponds to an estimated reduction of 25,000–48,000 diabetes cases per million people, assuming similar risk structures, intervention uptake, and causal transportability. The robustness of both direct and mediated effects was confirmed through placebo, subset, bootstrap, and sensitivity refutation tests, where -values were generally high, though a few fell below 0.88. Taken together, these findings emphasize that addressing stress, sleep disturbance, and poor mental health in tandem is critical, as combined stressors not only raise diabetes risk directly but also reinforce harmful dietary patterns that compound the metabolic burden.

10.3.1 Discussion

10.3.2 Implications for clinical and policy action

These findings have potential clinical and policy implications. First, they suggest the value of an expanded view of diabetes prevention that moves beyond simplistic dietary advice to consider upstream psychosocial drivers of unhealthy eating. Within the context of this study, poor food choices appear less as isolated personal decisions and more as responses shaped by chronic stress, fatigue, and emotional burdens. Interventions focused solely on one stressor (e.g., sleep loss) or one behavior (e.g., diet) may under-perform if they ignore the structural and psychological roots of these behaviors. A more comprehensive model that integrates psychological, social, and behavioral supports may be better suited for addressing multimorbidity and psychosocial clustering in vulnerable populations.

10.3.3 Pathways to integrated, stress-informed interventions

Integrated care models that combine dietary guidance with stress management, sleep hygiene, and mental health support may offer a promising approach. For example, sleep interventions could address not only sleep duration and quality but also related dietary vulnerabilities, such as cravings for salty and processed foods. Loneliness interventions might incorporate structured meal planning and social meals to strengthen community support and foster healthier eating habits. Mental health services could consider targeting emotional eating and motivational barriers, recognizing that psychological distress often shapes food choices.

10.4 Q4: Impact of simultaneously improving psychosocial factors on T2DM risk

This analysis examines the potential impact of a modeled holistic intervention targeting multiple psychosocial and behavioral risk factors simultaneously. Specifically, this study assessed the effect of improved mental health (no history of psychiatric consultation for anxiety or depression), reduced loneliness, and absence of insomnia on the overall risk of developing T2DM. These improvements were encoded as a single composite variable, ”Improved-Psychological-health”, to reflect a real-world, integrated preventive strategy.

As shown in Table 8, improved psychosocial health was associated with a strong and statistically robust estimated protective effect on T2DM. The ATT was (95% CI: to ), corresponding to an 11.6 percentage-point lower absolute risk. This finding underscores the potential value of holistic interventions that address mental health, sleep, and social connectedness together. Robustness checks supported the stability of the estimate. Placebo effects were negligible (0, ), subset analyses closely matched the main result (, ), and bootstrap replicates yielded nearly identical effects (, ). However, sensitivity analysis indicated that the protective estimate () remained within a plausible range of to . Because this range crosses zero, the protective effect could be attenuated to null or even reversed under certain levels of unmeasured confounding. Thus, while the consistency across robustness tests supports improved psychosocial health as a protective factor, caution is warranted in the causal interpretation.

10.4.1 Broader implications and mechanisms

These findings are consistent with the growing recognition that mental health, social relationships, and sleep are interconnected determinants of metabolic health. Depression and loneliness have been independently linked to systemic inflammation and hypothalamic–pituitary–adrenal (HPA) axis dysregulation [11, 107], while insomnia exacerbates sympathetic nervous system activity and impairs glucose metabolism [109, 110]. Addressing these domains together likely yields synergistic benefits that surpass the additive effects of targeting them in isolation [111].

This integrative perspective aligns with the “syndemic” framework in public health [111], which emphasizes how co-occurring psychosocial risks reinforce chronic disease pathways. For instance, improved sleep may buffer against mood disturbances and social withdrawal, while better mental health can support adherence to sleep hygiene and healthier routines. Such interdependence underscores the importance of moving beyond siloed strategies in prevention efforts.

10.4.2 Clinical and public health relevance

11 Limitations

Several limitations should be considered when interpreting the findings of this study. First, the analysis is based on a retrospective observational dataset, which limits causal identification to the assumptions underpinning the causal modeling framework. Although extensive covariate adjustment, expert-informed DAG construction, and multiple robustness checks were employed, the possibility of residual or unmeasured confounding cannot be fully excluded.

Second, several key exposures and mediators, including dietary behaviors, sleep disturbances, loneliness, and mental health indicators, were derived from self-reported data. Such measures may be subject to recall bias, reporting error, and misclassification, which could attenuate or inflate estimated effects. In addition, dietary variables were treated as baseline preferences rather than time-varying behaviors, limiting the ability to capture dynamic changes over the life course.

Third, modeling choices necessarily involved simplifications. Continuous variables such as BMI and dietary intake were dichotomized using clinically meaningful thresholds to facilitate causal interpretation and intervention simulation. While this improves interpretability, it may obscure nonlinear relationships or dose–response patterns. Similarly, psychosocial stressors were modeled as binary or composite exposures, which may not fully capture the severity or duration of these conditions.

Fourth, causal effect estimates rely on correct specification of the causal DAG and the assumption that all relevant confounders were observed and appropriately adjusted for. Although covariates were selected a priori based on domain knowledge, misspecification of causal relationships or omitted variables could bias estimates. The ATT estimates reported here reflect effects among the treated population and should not be interpreted as population-average effects or universal intervention impacts.

Finally, generalizability is limited. The findings reflect estimated counterfactual effects within the studied cohort and under the specified modeling assumptions. Population-level scaling and policy implications are illustrative and assume comparable risk structures, intervention uptake, and contextual factors. Prospective validation, incorporation of longitudinal and real-time data, and evaluation in diverse populations are needed before translating these findings into real-world intervention strategies.

12 Conclusion and future work

This study introduced a novel digital twin framework for predicting and simulating the onset of T2DM using retrospective behavioral, dietary, and psychosocial data. By moving beyond reliance on real-time monitoring and clinical biomarkers, the model demonstrates the potential for a low-burden, interpretable, and accessible approach to early disease prevention. Integrating psychosocial stressors such as insomnia, loneliness, and mental health history with dietary behaviors supported a more holistic characterization of estimated diabetes risk within this cohort. Key dietary factors, including processed meat, sugary cereals, salt intake, and cheese consumption, emerged as significant and modifiable predictors.

Causal inference analyses indicated that, under the modeled counterfactual assumptions, individual psychosocial stressors were associated with estimated increases in absolute T2DM risk of roughly 35 percentage points, while clustering of multiple stressors was associated with increases approaching 78 percentage points. Dietary mediation, though modest in absolute size, consistently reinforced these estimated effects within the modeled pathways, particularly through salt, processed foods, and reduced intake of protective foods. Importantly, the framework also revealed stark ethnic disparities, with individuals of Bangladeshi, Indian, Pakistani, African, and Caribbean descent showing substantially higher hazard ratios, highlighting the potential importance of culturally tailored prevention strategies. Validation through cross-validation (C-index 0.90), placebo checks, bootstrap resampling, and sensitivity analyses supported the internal robustness of the modeling framework.

Building directly on the limitations identified in this study, future work will focus on extending the digital twin architecture beyond static, retrospective modeling. Incorporating longitudinal and real-time data streams, where available, will enable the representation of time-varying exposures, feedback mechanisms, and behavioral adaptation, addressing current constraints related to static counterfactual assumptions and self-reported measures. Such extensions would improve the realism of simulated interventions and allow the digital twin to evolve dynamically as individual risk profiles change over time.

Personalization of the framework could be further enhanced through the integration of genomic, geographic, and socioeconomic information, allowing more precise risk stratification and reducing residual confounding related to unobserved structural determinants of health. Expanding the framework to incorporate contextual factors such as healthcare access, neighborhood environments, and occupational stressors would further strengthen its relevance for diverse populations.

Ultimately, while the present findings are derived from modeled counterfactual scenarios within a specific cohort, this work lays the groundwork for future prospective validation and real-world evaluation. With further development and empirical testing, the digital twin framework has the potential to inform accessible, interpretable, and equitable prevention strategies that address both the biological and psychosocial dimensions of chronic disease risk.

References

• 1.

  IDF. Data from: Diabetes facts and figures. IDF diabetes atlas. 10th ed. (2021). Available online at: https://idf.org/aboutdiabetes/what-is-diabetes/facts-figures.html(Accessed August 16, 2024).

• 2.

  MarxV. The big challenges of big data. Nature. (2013) 498:255–60. 10.1038/498255a

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 3.

  SangheraDKOrtegaLHanSSinghJRalhanSKWanderGS, et al. Impact of nine common type 2 diabetes risk polymorphisms in Asian Indian Sikhs: PPARG2 (Pro12Ala), IGF2BP2, TCF7L2 and FTO variants confer a significant risk. BMC Med Genet. (2008) 9:1–9. 10.1186/1471-2350-9-59

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 4.

  VermaSSrivastavaNBanerjeeM. Genetic polymorphisms in TCF7L2 and PPARG genes and susceptibility to type 2 diabetes mellitus. Meta Gene. (2021) 28:100864. 10.1016/j.mgene.2021.100864

  • CrossRef
  • Google Scholar

• 5.

  HackettRASteptoeA. Type 2 diabetes mellitus and psychological stress—a modifiable risk factor. Nat Rev Endocrinol. (2017) 13:547–60. 10.1038/nrendo.2017.64

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 6.

  WannametheeSGShaperAGPerryIJ. Smoking as a modifiable risk factor for type 2 diabetes in middle-aged men. Diabetes Care. (2001) 24:1590–5. 10.2337/diacare.24.9.1590

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 7.

  Mayer-DavisEJCostacouT. Obesity and sedentary lifestyle: modifiable risk factors for prevention of type 2 diabetes. Curr Diab Rep. (2001) 1:170–6. 10.1007/s11892-001-0030-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 8.

  XiongYZhangFZhangYWangWRanYWuC, et al. Insights into modifiable risk factors of erectile dysfunction, a wide-angled mendelian randomization study. J Adv Res. (2024) 58:149–61. 10.1016/j.jare.2023.05.008

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 9.

  SchutteRZhangJKiranMBallG. Alcohol and arterial stiffness in middle-aged and older adults: Cross-sectional evidence from the UK Biobank study. Alcohol: Clin Exp Res. (2024) 48:1915–22. 10.1111/acer.15426

  • CrossRef
  • Google Scholar

• 10.

  KnutsonKL. Impact of sleep and sleep loss on glucose homeostasis and appetite regulation. Sleep Med Clin. (2007) 2:187–97. 10.1016/j.jsmc.2007.03.004

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 11.

  Holt-LunstadJ. Social connection as a public health issue: the evidence and a systemic framework for prioritizing the “social” in social determinants of health. Annu Rev Public Health. (2022) 43:193–213. 10.1146/annurev-publhealth-052020-110732

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 12.

  MouchaccaJAbbottGRBallK. Associations between psychological stress, eating, physical activity, sedentary behaviours and body weight among women: a longitudinal study. BMC Public Health. (2013) 13:1–11. 10.1186/1471-2458-13-828

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 13.

  EatonSBEatonSB. Physical inactivity, obesity, and type 2 diabetes: an evolutionary perspective. Res Q Exerc Sport. (2017) 88:1–8. 10.1080/02701367.2016.1268519

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 14.

  TasinINabilTUIslamSKhanR. Diabetes prediction using machine learning and explainable ai techniques. Healthc Technol Lett. (2023) 10:1–10. 10.1049/htl2.12039

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 15.

  OjurongbeTAAfolabiHAOyekaleABashiruKAAyelagbeOOjurongbeO, et al. Predictive model for early detection of type 2 diabetes using patients’ clinical symptoms, demographic features, and knowledge of diabetes. Health Sci Rep. (2024) 7:e1834. 10.1002/hsr2.1834

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 16.

  KiranMXieYAnjumNBallGPierscionekBRussellD. Machine learning and artificial intelligence in type 2 diabetes prediction: a comprehensive 33-year bibliometric and literature analysis. Front Digit Health. (2025) 7:1557467. 10.3389/fdgth.2025.1557467

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 17.

  TaoFLiuWZhangMHuTQiQZhangH, et al. Five-dimension digital twin model and its ten applications. Comput Integr Manuf Syst. (2019) 25:1–18. 10.13196/j.cims.2019.01.002

  • CrossRef
  • Google Scholar

• 18.

  BagariaNLaamartiFBadawiHFAlbraikanAMartinez VelazquezRAEl SaddikA. Health 4.0: digital twins for health and well-being. In: El Saddik A, Hossain MS, Kantarci B, editors. Connected Health in Smart Cities. Cham: Springer (2020). p. 143–52.

  • Google Scholar

• 19.

  ShamannaPJoshiSThajudeenMShahLPoonTMohamedM, et al. Personalized nutrition in type 2 diabetes remission: application of digital twin technology for predictive glycemic control. Front Endocrinol (Lausanne). (2024) 15:1485464. 10.3389/fendo.2024.1485464

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 20.

  DinhAMiertschinSYoungAMohantySD. A data-driven approach to predicting diabetes and cardiovascular disease with machine learning. BMC Med Inform Decis Mak. (2019) 19:1–15. 10.1186/s12911-019-0918-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 21.

  ShresthaBAlsadoonAPrasadPAl-NaymatGAl-Dala’inTRashidTA, et al. Enhancing the prediction of type 2 diabetes mellitus using sparse balanced SVM. Multimed Tools Appl. (2022) 81:1–25. 10.1007/s11042-022-13087-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 22.

  YunJSKimJJungSHChaSAKoSHAhnYB, et al. A deep learning model for screening type 2 diabetes from retinal photographs. Nutr Metab Cardiovasc Dis. (2022) 32:1218–26. 10.1016/j.numecd.2022.01.010

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 23.

  AnjumNAlibakhshikenariMRashidJJabeenFAsifAMohamedEM, et al. IoT-based COVID-19 diagnosing and monitoring systems: a survey. IEEE Access. (2022) 10:87168–81. 10.1109/ACCESS.2022.3197164

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 24.

  HassijaVChamolaVMahapatraASingalAGoelDHuangK, et al. Interpreting black-box models: a review on explainable artificial intelligence. Cognit Comput. (2024) 16:45–74. 10.1007/s12559-023-10179-8

  • CrossRef
  • Google Scholar

• 25.

  ZhengRNgSTShaoYLiZXingJ. Leveraging digital twin for healthcare emergency management system: recent advances, critical challenges, and future directions. Reliab Eng Syst Saf. (2025) 261:111079. 10.1016/j.ress.2025.111079

  • CrossRef
  • Google Scholar

• 26.

  XamesMDTopcuTG. A systematic literature review of digital twin research for healthcare systems: research trends, gaps, and realization challenges. IEEE Access. (2024) 12:4099–126. 10.1109/ACCESS.2023.3349379

  • CrossRef
  • Google Scholar

• 27.

  HanantoALTirtaAHerawanSGIdrisMSoudagarMEMDjamariDW, et al. Digital twin and 3d digital twin: concepts, applications, and challenges in industry 4.0 for digital twin. Computers. (2024) 13:100. 10.3390/computers13040100

  • CrossRef
  • Google Scholar

• 28.

  BiobankU. Data from: About uk biobank (2014).

  • Google Scholar

• 29.

  CoxDR. Regression models and life-tables. J R Stat Soc Ser B (Methodol). (1972) 34:187–202. 10.1111/j.2517-6161.1972.tb00899.x

  • CrossRef
  • Google Scholar

• 30.

  NeubergLG. Causality: models, reasoning, and inference, by Judea Pearl, Cambridge University Press, 2000. Econ Theory. (2003) 19:675–85. 10.1017/S0266466603004109

  • CrossRef
  • Google Scholar

• 31.

  SharmaAKicimanE. DoWhy: an end-to-end library for causal inference. arXiv [Preprint] arXiv:2011.04216 (2020). 10.48550/arXiv.2011.04216

  • CrossRef
  • Google Scholar

• 32.

  QinaA. Digital twins in health and nutrition. Qina.tech Blog (2024).

  • Google Scholar

• 33.

  KatsoulakisEWangQWuHShahriyariLFletcherRLiuJ, et al. Digital twins for health: a scoping review. NPJ Digit Med. (2024) 7:77. 10.1038/s41746-024-01073-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 34.

  ShamannaPSabooBDamodharanSMohammedJMohamedMPoonT, et al. Reducing HbA1c in type 2 diabetes using digital twin technology-enabled precision nutrition: a retrospective analysis. Diabetes Ther. (2020) 11:2703–14. 10.1007/s13300-020-00931-w

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 35.

  ShamannaPDharmalingamMSahayRMohammedJMohamedMPoonT, et al. Retrospective study of glycemic variability, BMI, and blood pressure in diabetes patients in the digital twin precision treatment program. Sci Rep. (2021) 11:1–9. 10.1038/s41598-021-94339-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 36.

  ShamannaPErukulapatiRSShuklaAShahLWillisBThajudeenM, et al. One-year outcomes of a digital twin intervention for type 2 diabetes: a retrospective real-world study. Sci Rep. (2024) 14:25478. 10.1038/s41598-024-76584-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 37.

  ShamannaPJoshiSShahLDharmalingamMSabooBMohammedJ, et al. Type 2 diabetes reversal with digital twin technology-enabled precision nutrition and staging of reversal: a retrospective cohort study. Clin Diabetes Endocrinol. (2021) 7:1–8. 10.1186/s40842-021-00134-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 38.

  ShamannaPJoshiSDharmalingamMVadaviAKeshavamurthyAShahL, et al. Digital twin in managing hypertension among people with type 2 diabetes: 1-year randomized controlled trial. JACC: Adv. (2024) 3:101172. 10.1016/j.jacadv.2024.101172

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 39.

  SilfvergrenOSimonssonCEkstedtMLundbergPGennemarkPCedersundG. Digital twin predicting diet response before and after long-term fasting. bioRxiv (2021).

  • Google Scholar

• 40.

  VaskovskyAMChvanovaMS. Designing the neural network for personalization of food products for persons with genetic president of diabetic sugar. In: 2019 3rd School on Dynamics of Complex Networks and their Application in Intellectual Robotics (DCNAIR). Piscataway, NJ: IEEE (2019). p. 175–7.

  • Google Scholar

• 41.

  FagherazziG. Challenges and perspectives for the future of diabetes epidemiology in the era of digital health and artificial intelligence. Diabetes Epidemiol Manag. (2021) 1:100004. 10.1016/j.deman.2021.100004

  • CrossRef
  • Google Scholar

• 42.

  PrinceMPatelVSaxenaSMajMMaselkoJPhillipsMR, et al. No health without mental health. Lancet. (2007) 370:859–77. 10.1016/S0140-6736(07)61238-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 43.

  ZulmanDMAschSMMartinsSBKerrEAHoffmanBBGoldsteinMK. Quality of care for patients with multiple chronic conditions: the role of comorbidity interrelatedness. J Gen Intern Med. (2014) 29:529–37. 10.1007/s11606-013-2616-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 44.

  SchneidermanNAntoniMHSaabPGIronsonG. Health psychology: psychosocial and biobehavioral aspects of chronic disease management. Annu Rev Psychol. (2001) 52:555–80. 10.1146/annurev.psych.52.1.555

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 45.

  BarnettKMercerSWNorburyMWattGWykeSGuthrieB. Epidemiology of multimorbidity and implications for health care, research, and medical education: a cross-sectional study. Lancet. (2012) 380:37–43. 10.1016/S0140-6736(12)60240-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 46.

  Van den AkkerMBuntinxFMetsemakersJFRoosSKnottnerusJA. Multimorbidity in general practice: prevalence, incidence, and determinants of co-occurring chronic and recurrent diseases. J Clin Epidemiol. (1998) 51:367–75. 10.1016/S0895-4356(97)00306-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 47.

  UK Biobank. UK Biobank: Hospital inpatient data (Hospital Episode Statistics) (2023). Version 4.0; Hospital Episode Statistics Admitted Patient Care (HES APC); PDF document.

• 48.

  KiranMXieYBallGAnjumDNSchutteRPierscionekB. Type 2 diabetes prediction without labs: a systems-level neural framework for risk and behavioral network reorganization. Front Digit Health. (2026) 7:1714545. 10.3389/fdgth.2025.1714545

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 49.

  TabákAGHerderCRathmannWBrunnerEJKivimäkiM. Prediabetes: a high-risk state for diabetes development. Lancet. (2012) 379:2279–90. 10.1016/S0140-6736(12)60283-9

  • CrossRef
  • Google Scholar

• 50.

  JakobsenJCGluudCWetterslevJWinkelP. When and how should multiple imputation be used for handling missing data in randomised clinical trials—a practical guide with flowcharts. BMC Med Res Methodol. (2017) 17:162. 10.1186/s12874-017-0442-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 51.

  KuhnMJohnsonK. Applied Predictive Modeling. Vol. 26. New York: Springer (2013).

  • Google Scholar

• 52.

  HuberPJRonchettiE. Robust Statistics. 2nd ed. Hoboken, NJ: John Wiley & Sons (2009).

  • Google Scholar

• 53.

  RousseeuwPJLeroyAM. Robust Regression and Outlier Detection. New York: John Wiley & Sons (2003).

  • Google Scholar

• 54.

  BoxGE. Empirical Model Building and Response Surfaces. Vol. 2. New York: John Wiley Sons Google Scholar (1987). p. 27–37.

  • Google Scholar

• 55.

  Van der MaatenLHintonG. Visualizing data using t-SNE. J Mach Learn Res. (2008) 9:2579–605.

  • Google Scholar

• 56.

  RousseeuwPJ. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. J Comput Appl Math. (1987) 20:53–65. 10.1016/0377-0427(87)90125-7

  • CrossRef
  • Google Scholar

• 57.

  HotellingH. Analysis of a complex of statistical variables into principal components. J Educ Psychol. (1933) 24:417. 10.1037/h0071325

  • CrossRef
  • Google Scholar

• 58.

  Ward JrJH. Hierarchical grouping to optimize an objective function. J Am Stat Assoc. (1963) 58:236–44. 10.1080/01621459.1963.10500845

  • CrossRef
  • Google Scholar

• 59.

  BorganØGoldsteinLLangholzB. Methods for the analysis of sampled cohort data in the Cox proportional hazards model. Ann Stat. (1995) 23:1749–78. 10.1214/aos/1176324322

  • CrossRef
  • Google Scholar

• 60.

  DattaGAlexanderLEHinterbergMAHagarY. Balanced event prediction through sampled survival analysis. Syst Med. (2019) 2:28–38. 10.1089/sysm.2018.0015

  • CrossRef
  • Google Scholar

• 61.

  BorganØSamuelsenSO. A review of cohort sampling designs for Cox’s regression model: Potentials in epidemiology. Norsk Epidemiol. (2003) 13:239–48. 10.5324/nje.v13i2.292

  • CrossRef
  • Google Scholar

• 62.

  MontgomeryDCPeckEAViningGG. Introduction to Linear Regression Analysis. Hoboken, NJ: John Wiley & Sons (2021).

  • Google Scholar

• 63.

  SchoenfeldD. Partial residuals for the proportional hazards regression model. Biometrika. (1982) 69:239–41. 10.1093/biomet/69.1.239

  • CrossRef
  • Google Scholar

• 64.

  KohaviR. A study of cross-validation and bootstrap for accuracy estimation and model selection. In: Proceedings of the 14th International Joint Conference on Artificial Intelligence. Montreal, Canada: Morgan Kaufmann (1995). p. 1137–45.

  • Google Scholar

• 65.

  VerweijPJVan HouwelingenHC. Penalized likelihood in Cox regression. Stat Med. (1994) 13:2427–36. 10.1002/sim.4780132307

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 66.

  Harrell JrFELeeKLMarkDB. Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med. (1996) 15:361–87. 10.1002/(SICI)1097-0258(19960229)15:4%3C361::AID-SIM168%3E3.0.CO;2-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 67.

  ConsultationW. Obesity: preventing and managing the global epidemic. World Health Organ Tech Rep Ser. (2000) 894:1–253.

  • Google Scholar

• 68.

  MozaffarianDHaoTRimmEBWillettWCHuFB. Changes in diet and lifestyle and long-term weight gain in women and men. N Engl J Med. (2011) 364:2392–404. 10.1056/NEJMoa1014296

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 69.

  SlurinkIAChenLMaglianoDJKupperNSmeetsTSoedamah-MuthuSS. Dairy product consumption and incident prediabetes in the Australian diabetes, obesity, and lifestyle study with 12 years of follow-up. J Nutr. (2023) 153:1742–52. 10.1016/j.tjnut.2023.03.032

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 70.

  DrehmerMPereiraMASchmidtMIAlvimSLotufoPALuftVC, et al. Total and full-fat, but not low-fat, dairy product intakes are inversely associated with metabolic syndrome in adults. J Nutr. (2016) 146:81–9. 10.3945/jn.115.220699

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 71.

  BeulensJWSluijsISpijkermanAMvan der SchouwYT. Dietary phylloquinone and menaquinones intake and risk of type 2 diabetes. Diabetes Care. (2010) 33:1699–705. 10.2337/dc09-2302

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 72.

  EckelRHKahnSEFerranniniEGoldfineABNathanDMSchwartzMW, et al. Obesity and type 2 diabetes: what can be unified and what needs to be individualized?J Clin Endocrinol Metab. (2011) 96:1654–63. 10.1210/jc.2011-0585

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 73.

  TillinTForouhiNGMcKeiguePMChaturvediN. Southall and brent revisited: cohort profile of SABRE, a UK population-based comparison of cardiovascular disease and diabetes in people of European, Indian Asian and African Caribbean origins. Int J Epidemiol. (2012) 41:33–42. 10.1093/ije/dyq175

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 74.

  EastwoodSVTillinTDehbiHMWrightAForouhiNGGodslandI, et al. Ethnic differences in associations between fat deposition and incident diabetes and underlying mechanisms: the SABRE study. Obesity. (2015) 23:699–706. 10.1002/oby.20997

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 75.

  EnglandPH. Data from: Health matters: preventing type 2 diabetes (2016). Available online at: https://www.gov.uk/government/publications/health-matters-preventing-type-2-diabetes/health-matters-preventing-type-2-diabetes (accessed April, 2024).

  • Google Scholar

• 76.

  GuessND. Dietary interventions for the prevention of type 2 diabetes in high-risk groups: current state of evidence and future research needs. Nutrients. (2018) 10:1245. 10.3390/nu10091245

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 77.

  FayyadUMIraniKB. Multi-interval discretization of continuous-valued attributes for classification learning. In: IJCAI. Vol. 93. Citeseer (1993). p. 1022–9.

  • Google Scholar

• 78.

  StalpersLJKaplanEL. Edward L. Kaplan and the Kaplan–Meier survival curve. BSHM Bull: J Br Soc Hist Math. (2018) 33:109–35. 10.1080/17498430.2018.1450055

  • CrossRef
  • Google Scholar

• 79.

  MantelN. Evaluation of survival data and two new rank order statistics arising in its consideration. Cancer Chemother Rep. (1966) 50:163–70.

  • Pubmed Abstract
  • Google Scholar

• 80.

  CharlesS. Data from: Type 2 diabetes: statistics and facts (2025). Available online at: https://www.gov.uk/government/statistics/diabetes-profile-update-march-2025/diabetes-profile-statistical-commentary-march-2025 (accessed April 17, 2025).

  • Google Scholar

• 81.

  FernandesMAntonucciMCapecciFMercuriNBDella-MorteDLiguoriC. Prevalence of sleep disorders in geriatrics: an exploratory study using sleep questionnaires. Geriatr Nurs (Minneap). (2024) 60:107–13. 10.1016/j.gerinurse.2024.08.032

  • CrossRef
  • Google Scholar

• 82.

  DendupTFengXO’ShaughnessyPAstell-BurtT. Perceived built environment and type 2 diabetes incidence: exploring potential mediating pathways through physical and mental health, and behavioural factors in a longitudinal study. Diabetes Res Clin Pract. (2021) 176:108841. 10.1016/j.diabres.2021.108841

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 83.

  Romero-CorralACaplesSMLopez-JimenezFSomersVK. Interactions between obesity and obstructive sleep apnea: implications for treatment. Chest. (2010) 137:711–9. 10.1378/chest.09-0360

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 84.

  MartynJJKanekiMYasuharaS. Obesity-induced insulin resistance and hyperglycemia: etiological factors and molecular mechanisms. Anesthesiology. (2008) 109:137. 10.1097/ALN.0b013e3181799d45

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 85.

  ChenZKhandpurNDesjardinsCWangLMonteiroCARossatoSL, et al. Ultra-processed food consumption and risk of type 2 diabetes: three large prospective us cohort studies. Diabetes Care. (2023) 46:1335–44. 10.2337/dc22-1993

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 86.

  ZhongTHuangYWangG. The causal association of cheese intake with type 2 diabetes mellitus: results from a two-sample mendelian randomization study. Arch Med Sci. (2024) 20:1930–42. 10.5114/aoms/188068

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 87.

  SrourBFezeuLKKesse-GuyotEAllèsBMéjeanCAndrianasoloRM, et al. Ultra-processed food intake and risk of cardiovascular disease: prospective cohort study (NutriNet-Santé). bmj. (2019) 365:l145. 10.1136/bmj.l1451

  • CrossRef
  • Google Scholar

• 88.

  WuQBurleyGLiLCLinSShiYC. The role of dietary salt in metabolism and energy balance: Insights beyond cardiovascular disease. Diabetes Obes Metab. (2023) 25:1147–61. 10.1111/dom.14980

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 89.

  SteptoeAFongHLLassaleC. Social isolation, loneliness and low dietary micronutrient intake amongst older people in England. Age Ageing. (2024) 53:afae223. 10.1093/ageing/afae223

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 90.

  GalAMIatcuCOPopaADArhireLIMihalacheLGherasimA, et al. Understanding the interplay of dietary intake and eating behavior in type 2 diabetes. Nutrients. (2024) 16:771. 10.3390/nu16060771

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 91.

  DuanDKimLJJunJCPolotskyVY. Connecting insufficient sleep and insomnia with metabolic dysfunction. Ann N Y Acad Sci. (2023) 1519:94–117. 10.1111/nyas.14926

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 92.

  RoennebergTAllebrandtKVMerrowMVetterC. Social jetlag and obesity. Curr Biol. (2012) 22:939–43. 10.1016/j.cub.2012.03.038

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 93.

  Diabetes UK. Data from: Ethnicity and type 2 diabetes (2025). Available online at: https://www.diabetes.org.uk/about-diabetes/type-2-diabetes/diabetes-ethnicity (accessed April 17, 2025).

• 94.

  HalvorsrudKNazrooJOtisMHajdukovaEBBhuiK. Ethnic inequalities in the incidence of diagnosis of severe mental illness in England: a systematic review and new meta-analyses for non-affective and affective psychoses. Soc Psychiatry Psychiatr Epidemiol. (2019) 54:1311–23. 10.1007/s00127-019-01758-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 95.

  DemashkiehMHardyRShahPEllahiBAmenyahSOsei-KwasiH, et al. Dietary patterns in ethnic minority groups: data analysis of vegetable intake from ‘understanding society’ (the uk household longitudinal study). Proc Nutr Soc. (2024) 83:E267. 10.1017/S0029665124005056

  • CrossRef
  • Google Scholar

• 96.

  Mental Health Foundation. Data from: Long-term physical conditions and mental health (2025). Available online at: https://www.mentalhealth.org.uk/explore-mental-health/a-z-topics/long-term-physical-conditions-and-mental-health (accessed April 17, 2025)

• 97.

  KivimäkiMBattyGDSingh-ManouxAFerrieJETabákAGJokelaM, et al. Body-mass index and risk of obesity-related complex multimorbidity: an individual-level analysis of 120,813 adults from 16 cohort studies. Lancet Diabetes Endocrinol. (2022) 10:253–63. 10.1016/S2213-8587(22)00007-9

  • CrossRef
  • Google Scholar

• 98.

  StirlandLEGonzalez-MontalvoJISantamaria-PelaezMRodríguez-ArtalejoFCoscoTDPrinceM, et al. Multimorbidity and long-term disability and physical functioning decline: a prospective cohort study. BMC Geriatr. (2022) 22:548. 10.1186/s12877-022-03548-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 99.

  RubinDB. Estimating causal effects of treatments in randomized and nonrandomized studies. J Educ Psychol. (1974) 66:688. 10.1037/h0037350

  • CrossRef
  • Google Scholar

• 100.

  RosenbaumPRRubinDB. The central role of the propensity score in observational studies for causal effects. Biometrika. (1983) 70:41–55. 10.1093/biomet/70.1.41

  • CrossRef
  • Google Scholar

• 101.

  AustinPC. Balance diagnostics for comparing the distribution of baseline covariates between treatment groups in propensity-score matched samples. Stat Med. (2009) 28:3083–107. 10.1002/sim.3697

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 102.

  McEwenBS. Protective and damaging effects of stress mediators. N Engl J Med. (1998) 338:171–9. 10.1056/NEJM199801153380307

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 103.

  McEwenBS. Physiology and neurobiology of stress and adaptation: central role of the brain. Physiol Rev. (2007) 87:873–904. 10.1152/physrev.00041.2006

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 104.

  BlackPH. The inflammatory consequences of psychologic stress: relationship to insulin resistance, obesity, atherosclerosis and diabetes mellitus, type II. Med Hypotheses. (2006) 67:879–91. 10.1016/j.mehy.2006.04.008

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 105.

  GodosJCurrentiWAngelinoDMenaPCastellanoSCaraciF, et al. Diet and mental health: review of the recent updates on molecular mechanisms. Antioxidants. (2020) 9:346. 10.3390/antiox9040346

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 106.

  FraynMLivshitsSKnäuperB. Emotional eating and weight regulation: a qualitative study of compensatory behaviors and concerns. J Eat Disord. (2018) 6:1–10. 10.1186/s40337-018-0210-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 107.

  BlackPH. The inflammatory response is an integral part of the stress response: implications for atherosclerosis, insulin resistance, type II diabetes and metabolic syndrome X. Brain Behav Immun. (2003) 17:350–64. 10.1016/S0889-1591(03)00048-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 108.

  Cleveland Clinic. Data from: Ghrelin (2022). Available online at: https://my.clevelandclinic.org/health/body/22804-ghrelin (accessed May 7, 2025).

• 109.

  SpiegelKLeproultRVan CauterE. Impact of sleep debt on metabolic and endocrine function. Lancet. (1999) 354:1435–9. 10.1016/S0140-6736(99)01376-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 110.

  García-AvilesJEMéndez-HernándezRGuzmán-RuizMACruzMGuerrero-VargasNNVelázquez-MoctezumaJ, et al. Metabolic disturbances induced by sleep restriction as potential triggers for alzheimer’s disease. Front Integr Neurosci. (2021) 15:722523.

  • Google Scholar

• 111.

  SingerMBulledNOstrachBMendenhallE. Syndemics and the biosocial conception of health. Lancet. (2017) 389:941–50. 10.1016/S0140-6736(17)30003-X

  • Pubmed Abstract
  • CrossRef
  • Google Scholar