Data in this study were obtained through face-to-face investigation and the Electronic Health Medical Record System (EHRS) of Sichuan Provincial People’s Hospital. All subjects were inpatients who had been screened according to the following criteria. Patients with T2D [the World Health Organization (WHO) (1999) criteria were adopted for diagnosis of T2D] were included and would be excluded when he or she visited a medical institution within 12 months, had adjusted their treatment plan within 12 months, did not use chemicals for hypoglycemic therapy, had used traditional Chinese medicine, Chinese herbal medicine, and acupuncture to control glycemia within the last 12 months, and had liver and kidney dysfunction. The patient’s private information, such as name, home address, and contact number were hidden during the research. Informed consent forms were obtained before the investigation.

Univariate analysis for continuous variables was performed using t-tests, variance analysis, or the Wilcoxon signed-rank test. The categorical variables were analyzed using the chi-square test or Fisher’s exact test. P-values less than 0.05 (p < 0.05) were considered statistically significant.

There were 32 input variables identified for this study, including demographic information, laboratory indicators, disease-related characteristics, medication information, and economics.

The outcome variables were poor glycemic control and whether complications occur. In this study, HbA1c <7% was considered to be good glycemic control and ≥7% was considered to be poor glycemic control (American Diabetes Association, 2020). The complications analyzed in this study were common chronic complications of T2D.

The variables with missing values >70%, the maximum percentage of records in a single category >90%, and the maximum number of categories >95% were excluded. The minimum coefficient of variation was set to 0.1, and the minimum standard deviation was set to 0. The Pearson method was used to evaluate the correlation between input variables and outcome variables. We set the cutoff value of variable importance to 0.9 (1−α).

The raw data were randomly split into a training set (80%) and an independent testing set (20%) by 8:2 after the variable screening. The model was built based on the training set, and the testing set was used only for the evaluation of the performance after the modeling stages. The grouping of the training set and the testing set was determined by the random seed value of the partition.

End-to-end models were built to predict outcome variables from the input variables. The data were processed using the following ML algorithms: artificial neural network (ANN), Bayesian network (BN), chi-squared automatic interaction detector (CHAID), classification and regression tree (CRT), quick unbiased efficient statistical tree (QUEST), and discriminate (D) and ensemble (XF) models. The XF models summarized the output of the best three models (assessed by AUC) and generated their outputs based on the voting principle.

The predictive performance of the final models was assessed by the following performance metrics: area under the receiver operating characteristic curve (AUC), negative predictive value (NPV), positive predictive value (PPV), and accuracy.

We explored the variable importance of each outcome variable derived from the best predictive model among all the tested models. Variable importance reflected the contribution of input variables to the outcome variables in specific models.

IBM SPSS Modeler 18.0 (Company Name) was used to build various models and SAS 9.21 (Company Name) was used to conduct hypothesis testing.

A total of 800 T2D patients were screened by the inclusion and exclusion criteria. 525 patients who had visited medical institutions in the past year, 49 patients who had hepatic and renal insufficiency, 43 patients who had adjusted their treatment plan and who did not use chemotherapy for hypoglycemic therapy in the last 12 months, and 18 patients who received hypoglycemic treatments other than chemical drugs in the last 12 months were excluded. The final cohort consisted of 165 patients (the screening process of patients is shown in Figure 1), including 97 male patients and 68 female patients. Seven types of complications were found in 83 cases (i.e., diabetic peripheral neuropathy (DPN), diabetic angiopathy (DA), diabetic nephropathy (DN), diabetic eye disease (DED), diabetic foot (DF), diabetic ketoacidosis (KE), and diabetic skin lesions (DD). Due to the small sample size and data imbalance, ketoacidosis (KE), diabetic skin lesions (DD), and diabetic foot (DF) were not included in this study.

Tables 1 and 2 list the results of univariate analysis of risk factors for complications and HbA1c in T2D patients, respectively. According to Table 1, the duration of T2D was a significant factor affecting DN (p < 0.0001), DPN (p = 0.0022), DA (p = 0.0015), and DED (p = 0.0082), and the duration of unadjusted hypoglycemic treatment was a risk factor of DN (<0.0001), DPN (<0.0001), DA (<0.0001), DED (<0.0001), and KE (<0.0284). Genetic history of diabetes was a risk factor for DPN (p = 0.037) and DO (p = 0.0189). According to Table 2, the number of hypoglycemic drugs (p < 0.0233) and the duration of T2D (p < 0.0020) were important factors affecting HbA1c. The percentage of patients with HbA1c under control declined with the prolonging of the duration of unadjusted hypoglycemic therapy.

Among the total of 32 input variables, 18 were excluded due to the low correlation with the characteristics of the outcome variable, and five were excluded due to data imbalance. There were nine input variables and five outcome variables that were retained for the development of the final models. The input variables were age, duration of diabetes (≥1 year), duration of unadjusted hypoglycemic treatment (≥1 year), number of insulin species, total cost (total expenditure during hospitalization) of hypoglycemic drugs, and number of hypoglycemic drugs (which were computed as continuous variables) and gender, genetic history of diabetes, and dyslipidemia (which were computed as categorical variables). The outcome variables included the onset of DPN, DA, DED, and DN and the control status of glycemia.

Sixteen best-performing algorithms with the highest AUCs were selected for modeling of four complications, and two best-performing algorithms were selected for HbA1c. Ten independent replicate results were generated for each model by changing the data split of a dataset. This was achieved by modifying random seeds of the “partition” node. A total of 180 models were obtained. The modeling steps of DED are shown in Figure 2, and the ROC curve for the model with the highest AUC of each complication is shown in Figure 3.

The PPV, NPV, accuracy, and AUC for different ML algorithms by the testing set are shown in Table 3. Among the 18 evaluated models, most models performed well. XF performed best among all the predictive models of DN and DA, with AUCs of 0.902 ± 0.040 and 0.889 ± 0.059. D performed best among all the models of DPN and DED, with AUCs of 0.859 ± 0.050 and 0.832 ± 0.086. The best model for HbA1c was considered to be BN, with the highest AUC of 0.825 ± 0.092.

Figure 4 shows the variable importance of DN, DA, DED, DPN, and HbA1c derived from the best-performing ML algorithms. It also shows the relative importance of the variables with the top three most important variables of complications being the duration of T2D, the duration of unadjusted hypoglycemic treatment, and types of insulin. The top three most important variables of HbA1c were the number of hypoglycemic drugs, types of insulin, and total cost. The most important variables of DN, DA, DED, and DPN were age, duration of T2D, types of insulin, and duration of unadjusted hypoglycemic treatment, respectively. A novel predictive variable, the duration of unadjusted hypoglycemic treatment (during this time, the patient’s hypoglycemic treatment regimen remains unchanged, and relevant follow-up monitoring has not been performed), of T2D was identified from this study. We can predict the probability of complications in T2D patients through the duration of the hypoglycemic regimen.

This study employed ML algorithms to screen for cases likely to have diabetic complications and poor glycemic control among nonadherent T2D patients and provided potential risk prediction tools for both outcomes. Eighteen models were evaluated, and the risk factors for complications and poor glycemic control were identified, with the most important risk factors being the duration of T2D, the duration of unadjusted hypoglycemic treatment, types of insulin, number of hypoglycemic drugs, and total cost of hypoglycemic therapy. The prediction models we established in this study obtained acceptable performances. According to previous reports, under-monitoring and delay of treatment are major challenges to diabetes management (Ross et al., 2011; Khunti et al., 2013; Khunti et al., 2016). The findings of this study are important because early screening may strengthen glycemic control and reduce the risk of diabetic complications through timely monitoring of glycemia and treatment intensification (Holman et al., 2008; Colagiuri and Davies, 2009; Griffin et al., 2011).

ML algorithms have been widely used in medical fields recently (Cichosz et al., 2015; Kavakiotis et al., 2017; Contreras and Vehi, 2018; Lan et al., 2019). It is the key technology of big data analysis, which provides new ways for clinicians to solve medical problems (Hui et al., 2016). Recent advances in ML algorithms have improved the accuracy of diagnosis and prediction in outcomes, in some cases even surpassing the performance of clinicians (Beam and Kohane, 2018). ML-based prediction models for classification or prediction of future health states are being developed (Emanuel and Wachter, 2019).

A large number of studies have reported on the prediction model of diabetic complications and glycemia. In the study of Hsin-Yi Tsao et al., data mining techniques were used to create prediction models of diabetic retinopathy, with results that indicated that insulin therapy and duration of diabetes are the most important risk factors of diabetic retinopathy, which was consistent with this study (Tsao et al., 2018). Compared with previous research, our study also found a new risk factor, the duration of unadjusted hypoglycemic treatment, for DED. Konstantia Zarkogianni et al. developed a risk prediction model for T2D cardiovascular complication (Zarkogianni et al., 2018). As with most predictive models, the prediction results are difficult to interpret. In our study, the prediction results were interpretable due to the use of decision trees. Dennis H Murphree et al. built several ML models to predict good HbA1c control (<7.0%) among T2D patients, which showed the potential for applying ML to solve problems in medical fields (Tsao et al., 2018). Consistent with prior research studies (Murphree et al., 2018; Tsao et al., 2018; Zarkogianni et al., 2018; Aminian et al., 2020), the findings of this study showed high AUCs.

Previous studies have explored the characteristics of patients with medication nonadherence from different perspectives. A systematic review analyzed the relationship between medication nonadherence and the health outcomes in the elderly (Walsh et al., 2019) and showed that medication nonadherence may be significantly associated with all-cause hospitalization and mortality in old people (Walsh et al., 2019). Instead of the senior group, the subjects of our study were T2D patients, and the prognosis of T2D we predicted was HbA1c and diabetic complications. In a cross-sectional study, the author explored the main predictors of poor adherence among T2D patients (Demoz et al., 2020). Another study, a previous article by Dr. Wu, assessed multiple ML algorithms and predicted the medication nonadherence risks of patients with T2D (Wu et al., 2020). The two articles above were studies on the influencing factors of medication nonadherence, and the predicted outcome is the compliance of patients. Both articles are quite different from our work. In our study, statistical and ML methods were used to predict risk factors of HbA1c and potential complications of nonadherent T2D. The research method and outcomes were quite different. New research ideas were provided for the influencing factors and prediction models of T2D progression.

This study had some strengths. Instead of a generic cohort, a highly specific one, nonadherent T2D patients, was used. This was the first study to use ML models to explore the health outcomes of nonadherent T2D. Besides, the internal validation of these models was conducted using the following method. The raw data were randomly grouped ten times by modifying the seed value of the “partition.” In this way, independently repeated experiments were conducted, and the bias that may occur when datasets are randomly grouped was avoided. This method is also better than bootstrapping (Milea et al., 2020), which may increase the weight of some data. Moreover, the dataset we used for prediction contained clinical information that has not been studied before, which is the duration of unadjusted hypoglycemic treatment.

However, there were some notable limitations to this study. This was a single-center, small-sample study, and the performance of the final models was not compared with that of the established clinical reference tools, which limits the reliability of the verification results. Nevertheless, the influencing factors were analyzed through conventional statistical calculations, and the results of the univariate analysis were consistent with the prediction models. In the future, a large-scale, forward-looking, and multicenter study is needed for further external validation.