Gait and balance are no longer considered solely as physical activity, but rather as multifaceted complex processes involving the integration of motor, perceptual, and cognitive functions (Montero-Odasso et al., 2012; Ahn et al., 2023). More specifically, executive functions, which encompass attentional control, cognitive flexibility, psychomotor processing, inhibition, and goal setting, share a comprehensive neural network with motor abilities and gait control (Leisman et al., 2016). This network includes the prefrontal cortex, medial temporal lobe, and nigrostriatal system. Additionally, it encompasses anatomical structures like ventricles, cerebellum, white matter tracts, and the parietal lobes (Scherder et al., 2007; Wennberg et al., 2017; Crook et al., 2020).

In the past, healthcare professionals and researchers used to conduct gait assessments and cognitive assessments as distinct evaluations for older adults (Montero-Odasso et al., 2012). However, recent findings from clinical practice, epidemiological research, and clinical trials have provided growing evidence that gait and cognition are interconnected in the elderly (Woollacott and Shumway-Cook, 2002). Indeed, a substantial body of research consistently indicates a connection between gait abnormalities and early signs of cognitive decline, even among cognitively healthy individuals (Mielke et al., 2013; Savica et al., 2017). This relationship underscores the intricate interplay between physical and cognitive health. Gait disorders include slow, unstable, shifting, staggering, and/or asymmetrical walking as a result of neurological, musculoskeletal, and/or other acquired medical conditions (Verghese et al., 2008; Pirker and Katzenschlager, 2017). For instance, it was found that Mild Parkinsonian Signs (MPS) were a strong indicator for the development of dementia in the future. In both unaltered and statistically adjusted examinations, individuals exhibiting MPS were twice as prone to experiencing the onset of dementia when compared to their counterparts who did not exhibit these neurological indications (Louis et al., 2010). Changes in gait can guarantee potential early indicators of underlying cognitive issues, providing valuable insights into an individual’s neurological well-being before more apparent cognitive decline symptoms manifest (Ahn et al., 2023). Gait disorders can be assessed through visual inspection or quantitatively (e.g., speed, stride length, swing, and stance time), depending on how they affect the observed gait abnormalities. While the latter requires technological equipment that can be used to further differentiate individuals based on their cognitive status, the former is a helpful and reliable tool in routine therapeutic practice (Verghese et al., 2008).

For example, there is substantial evidence indicating that gait irregularities have the potential to forecast a gradual deterioration in cognitive function as assessed by the Digit Symbol Substitution Test (DSST), which evaluates executive functions, psychomotor speed, and attention (Kaufman, 1983), as demonstrated in numerous studies (Rosano et al., 2005; Inzitari et al., 2007; Mielke et al., 2013). Moreover, gait abnormalities are predictive of a reduction in both divided attention and cognitive flexibility, as assessed by the Trail Making Test part B (TMT-B) (Mielke et al., 2013). Furthermore, the presence of gait abnormalities is longitudinally linked to a decline in global cognitive assessments (Alfaro-Acha et al., 2007; Taniguchi et al., 2012). Gait disorders have been identified as one of the contributing factors to the development of dementia (Verghese et al., 2008).

In recent years, there has been a growing focus on the milder end of the cognitive staging, which encompasses the spectrum from normal aging to Alzheimer’s Disease (AD). It is increasingly recognized that there exists a transitional phase between normal aging and the clinical diagnosis of very early-stage AD. This intermediate stage has been referred to by various terms, including Mild Cognitive Impairment (MCI) (Petersen, 2004, 2011). In recent years, there has been a growing focus on amnestic MCI (aMCI), which refers to individuals experiencing memory loss beyond what is typical for their age and education. These individuals are also at a higher risk of developing AD compared to those with non-amnestic MCI (Petersen, 2011). Increased gait variability was recently linked to dementia, especially AD dementia and PD dementia, according to multisite research of 500 older persons with various neurodegenerative diseases. Although rhythm and postural control domains were similarly linked to dementias and MCI, only gait variability could reliably identify and categorize people with AD (Pieruccini-Faria et al., 2021).

Moreover, in a study (utilizing accelerometers to analyze gait patterns in individuals with normal cognition (NC), MCI, and AD, it was observed that the AD group exhibited significantly lower velocity and step length compared to both MCI and NC subjects. Notably, the inclusion of dual-task testing, where subjects were instructed to count backward from 100 to 0 while undergoing gait assessment, proved to be crucial in distinguishing between MCI and NC individuals (Muurling et al., 2020). These findings underscore the early onset of gait dysfunction in the AD spectrum and underscore the value of gait assessments as a potential biomarker for predicting transitions.

A recent longitudinal retrospective study underlined that the presence or the severity of gait and/or balance disturbances was associated with an increased risk of AD in a sample of aMCI. Therefore, community-dwelling older adults with aMCI may need frequent evaluation by nurses to detect possible risk factors for cognitive decline, particularly in cases involving gait and/or balance issues (Ahn et al., 2023). Early evidence by Tuena et al. (2023) highlighted that gait abnormalities identified through routine neurological gait examinations are linked to distinct trends in cognitive test performance over time in individuals with aMCI. Specifically, when compared to the group with normal gait, those with abnormal gait experienced a more rapid decline in attention (DSST) and overall cognitive function (Mini-Mental State Examination [MMSE]) tests. Significantly, TMT parts A (TMT-A) and TMT-B exhibited unique declines over time in the abnormal gait group, not observed in the normal gait group. Crucially, the presence of gait disorders [hazard ratio (HR) = 1.70] and declines in the performance of three gait-related cognitive tests (MMSE, HR = 1.09; DSST, HR = 1.03; TMT-B, HR = 1.01) were significantly associated with a higher risk of developing AD dementia in the overall aMCI Alzheimer’s Disease Neuroimaging Initiative (ADNI) population.

In recent years, machine learning (ML) models have been developed to make prognostic/diagnostic predictions to improve AD clinical practice (Battista et al., 2020). Machine learning models can be used to improve our understanding of the conversion from MCI to AD dementia by predicting clinical progressions and identifying individuals at risk. These models utilize different features (clinical, biological, and neuroimaging data, biomarker positivity, neuropsychological tests, and comorbidity information) to estimate the probabilities of progression (Pang et al., 2023). By incorporating longitudinal information encoded in efficient markers, machine learning frameworks can differentiate between progressive and non-progressive MCI subjects, aiding in developing personalized strategies for preventing or slowing the progression of dementia (Battista et al., 2020; Bucholc et al., 2023). By analyzing large datasets and applying machine learning algorithms, these models can identify patterns and relationships that may not be apparent through traditional statistical methods. Interpretable machine learning algorithms can also be used to develop predictive algorithms for individual conversion to dementia, considering complex patterns and interactions between variables (Chun et al., 2022). These machine learning approaches can contribute to early intervention and the selection of appropriate treatments for individuals at high risk of developing dementia (Dhakal et al., 2023).

Based on these premises, we further explored the predictive value of the gait profile and gait-associated neuropsychological measures that have been identified in previous work (Tuena et al., 2023). The aim of this study was to evaluate the potential of a neuropsychological assessment combined with a gait objective exam in predicting the transition from aMCI to AD dementia through machine learning algorithms. In the present study, we developed an ML model to predict conversion from aMCI to AD using relevant features (MMSE, DSST, TMT-B, and gait profile). We investigated the role of these three neuropsychological tests and gait profiles in the automatic classification of aMCI subjects who converted or did not convert to AD within 36 months. We developed an ML model to assess the accuracy of predicting conversion by using our features.

At baseline, the aMCIc group already showed an initial decline in cognitive measures, in particular, MMSE and DSST scores were lower and the execution time for the TMT-B was higher compared to aMCIs. In contrast, there were no significant differences in the sociodemographic variables and gait type.

In this paper, we sought to explore the predictive power of four predictors: gait (assessed by visual examination during neurological exam), MMSE, DSST, and TMT-B. These variables were previously identified (Tuena et al., 2023) as critical for the conversion from aMCI to AD within a follow-up period of 36 months.

To our knowledge, this is the first research that using previously published results, uses ML to study the impact of gait and gait-related cognitive measures to forecast the diagnosis of AD in aMCI. Our results show that the best algorithm was SVM and that on baseline assessment it was able to provide a 70% predictive accuracy of future AD diagnosis within 36 months from the first visit. This is true, especially for identifying cases (i.e., aMCI patients that during the 36-month follow-up converted to AD). That is to say that our model and best algorithm can be used as a quick, low-cost, and preliminary assessment for aMCI individual at high risk of developing AD, which might require more expansive and invasive testing (e.g., neuroimaging, genetic testing, cerebrospinal fluid testing). Additionally, this would enable to program in advance of secondary prevention strategies and therapies. Importance variable analysis showed that MMSE was the most important predictor during training and testing of the model, whereas gait was crucial for the testing phase of the model alone.

Many authors have used the ADNI database to identify predictors to forecast AD diagnosis using ML algorithms (Massetti et al., 2022; Rye et al., 2022; Franciotti et al., 2023; Sarica et al., 2023).

The features that have been used in the literature in this classification can be clustered into four groups: neuropsychological measures only (Battista et al., 2020; Massetti et al., 2022), neuropsychological measures and biomarkers (Chun et al., 2022; Massetti et al., 2022), neuropsychological measures and neuroimaging (Rye et al., 2022; Franciotti et al., 2023), and finally neuropsychological measures, biomarkers, and neuroimaging (Sarica et al., 2023). The study (Massetti et al., 2022) used all ADNI stages for training and compared the combination of various features to find the best performing model also stratifying it by age range, specifically for the 75–78 range and relying only on the neuropsychological assessment they obtained an accuracy of 0.72 at testing. Our accuracy score (0.70) in testing is in line with them; in particular, we demonstrate that is possible to obtain good results with a restricted set of predictors linked to AD development. Indeed, the sensitivity of our models (see Tables 2, 3) is higher than the one reported (0.5) by Masetti and co-authors for neuropsychology tests alone (Massetti et al., 2022). It is important to note that multidimensional features datasets yield higher accuracies (e.g., 0.87; Sarica et al., 2023) and should be preferred. Yet, a meta-analysis (Battista et al., 2020) that evaluated ML approaches used in AD illustrated the ranges of performance metrics of studies that predicted conversion from MCI to AD based on neuropsychological tests: accuracy = 0.61–0.85, sensitivity = 0.50–0.91, specificity = 0.48–0.91, AUC-ROC = 0.67–0.93. Moreover, the work by Battista et al. (2020) found that the best tests to predict the conversion to AD in MCI included MMSE and TMT-B in addition to memory tests. However, as found in a study (Tuena et al., 2023) memory tests are not affected by gait, and we preferred to focus on the cognitive measures linked to gait. Our results show that it is possible to have comparable accuracy, sensitivity, and specificity using a restricted set of neuropsychological gait-related predictors that are linked to the condition of interest. Nevertheless, a great challenge of ML research is to improve specificity parameter for aMCIc and aMCIs. This was noted in our results and also in the meta-analytic work by Battista and co-authors (Battista et al., 2020).

Feature weights in an SVM model indicate the relative importance of distinctive features in ML model decision making. Although these weights are not statistically significant values in the traditional sense, nevertheless they are valuable tools for model interpretation. For example, in neuroimaging studies, features are often considered as brain voxels to identify the brain regions that contribute most to the diagnosis or course of a disease (e.g., AD-dementia) and show them graphically by weight-map (Bron et al., 2015). Our analysis of the SVM model weights shows that MMSE was the predictor that contributed most positively to our classification problem and that gait contributed only to the testing phase. Because the test set represents a different portion of the data (30%) than the training set, it is possible that features take on different importance in the testing phase. This suggests that gait type might play an important role in the conversion phase on a different sample than the one used for training the model. In line with the previous study (Tuena et al., 2023), we found that the most important predictor was MMSE in the training and testing phase and that the type of gait, within this feature cluster, contributes to the prediction of conversion. One longitudinal study (Chun et al., 2022) used ML to forecast dementia diagnosis in aMCI using sociodemographic, neuropsychological, and APOE variables as features. Among the features they found important for their best algorithm was MMSE. We showed that gait could be a critical predictor to improve ML performance for future AD diagnosis. This finding is also in line with embodiment theories in aging, which suggest that sensorimotor and perceptual impairments could hamper higher-order cognitive functions and be linked to neurodegenerative processes (Kuehn et al., 2018).

ML techniques can be a powerful method to overcome the limits of classic statistical approaches (Tuena et al., 2020). Indeed, despite p-value being a widely used and consolidated parameter, it lacks predictive information; crucially, ML methods enable to generalize findings of diagnostic, prognostic, and predictive studies to new observations and inform theory and practice (Dwyer et al., 2018). Therefore, it is suggested to validate research studies with both classical and ML statistical approaches. Following this line of reasoning, we provided scientific evidence that previously identified predictors of AD diagnosis can have promising predictive diagnostic value.

However, our study has some limitations that we must acknowledge. First, the results can be only generalized to aMCI patients with inclusion/exclusion criteria of the ADNI1 cohort. Second, the performance of the classification accuracy is still not high enough, nevertheless considering that we used only four quick-to-administer predictors to forecast AD diagnosis within 36 months from the first visit, our results are encouraging. Third, regarding the classification of gait type is numerically unbalanced, within aMCI, and this may have negatively affected the evaluation of permutation importance. Fourth, the categorization of the gait type from the ADNI database is dichotomous categorical. Future research could improve these results by adopting quantitative (e.g., accelerometer) gait measures to provide ML with numeric patterns. Then, the model could be tested on other public and private MCI databases to assess its generalizability. Lastly, more complex algorithms such as neural networks and deep learning algorithms could be tested to improve accuracy performance.

To conclude, in this ML study we provided additional evidence that previously identified predictors with classic statistical approaches can have promising and encouraging results using artificial intelligence. Moreover, we showed that in addition to established neuropsychological tests, gait, and in general the motor system, could contribute to cognitive deterioration and AD diagnosis, endorsing theories that suggest a link between mind and body.

The original contributions presented in this study are included in this article/Supplementary material, further inquiries can be directed to the corresponding author.

The studies involving human participants were reviewed and approved by the institutional review boards of the participating institutions in the Alzheimer’s Disease Neuroimaging Initiative (ADNI). This study will be conducted in accordance with good clinical practice guidelines, the Declaration of Helsinki, US 21CFR Part 50–Protection of Human Subjects and Part 56–Institutional Review Boards, and in compliance with state and federal HIPAA regulations. Written informed consent was obtained from all subjects and/or authorized representatives and study partners before protocol-specific procedures were carried out. The Ethics committees/institutional review boards that approved the ADNI study are: Albany Medical Center Committee on Research Involving Human Subjects Institutional Review Board, Boston University Medical Campus and Boston Medical Center Institutional Review Board, Butler Hospital Institutional Review Board, Cleveland Clinic Institutional Review Board, Columbia University Medical Center Institutional Review Board, Duke University Health System Institutional Review Board, Emory Institutional Review Board, Georgetown University Institutional Review Board, Health Sciences Institutional Review Board, Houston Methodist Institutional Review Board, Howard University Office of Regulatory Research Compliance, Icahn School of Medicine at Mount Sinai Program for the Protection of Human Subjects, Indiana University Institutional Review Board, Institutional Review Board of Baylor College of Medicine, Jewish General Hospital Research Ethics Board, Johns Hopkins Medicine Institutional Review Board, Lifespan–Rhode Island Hospital Institutional Review Board, Mayo Clinic Institutional Review Board, Mount Sinai Medical Center Institutional Review Board, Nathan Kline Institute for Psychiatric Research & Rockland Psychiatric Center Institutional Review Board, New York University Langone Medical Center School of Medicine Institutional Review Board, Northwestern University Institutional Review Board, Oregon Health and Science University Institutional Review Board, Partners Human Research Committee Research Ethics, Board Sunnybrook Health Sciences Centre, Roper St. Francis Healthcare Institutional Review Board, Rush University Medical Center Institutional Review Board, St. Joseph’s Phoenix Institutional Review Board, Stanford Institutional Review Board, The Ohio State University Institutional Review Board, University Hospitals Cleveland Medical Center Institutional Review. Board, University of Alabama Office of the IRB, University of British Columbia Research Ethics Board, University of California Davis Institutional Review Board Administration, University of California Los Angeles Office of the Human Research Protection Program, University of California San Diego Human Research Protections Program, University of California San Francisco Human Research Protection Program, University of Iowa Institutional Review Board, University of Kansas Medical Center Human Subjects Committee, University of Kentucky Medical Institutional Review Board, University of Michigan Medical School Institutional Review Board, University of Pennsylvania Institutional Review Board, University of Pittsburgh Institutional Review Board, University of Rochester Research Subjects Review Board, University of South Florida Institutional Review Board, University of Southern, California Institutional Review Board, UT Southwestern Institution Review Board, VA Long Beach Healthcare System Institutional Review Board, Vanderbilt University Medical Center Institutional Review Board, Wake Forest School of Medicine Institutional Review Board, Washington University School of Medicine Institutional Review Board, Western Institutional Review Board, Western University Health Sciences Research Ethics Board, and Yale University Institutional Review Board.

CT: Conceptualization, Formal analysis, Methodology, Writing –review and editing. CP: Formal analysis, Writing –original draft, Writing –review and editing. CS-B: Writing –original draft. MS-B: Supervision, Writing –review and editing. GR: Supervision, Writing –review and editing.