A retrospective analysis was conducted on 280 diabetes patients admitted to our hospital from January 2019 to December 2024. Based on different post-discharge management strategies, the diabetes patients (n = 280) were divided into a conventional management group (n = 100) and a personalized management group (n = 180). All patients were managed and followed up for 6 months. According to the blood glucose control status 6 months after discharge, patients in the personalized management group (n = 180) were further categorized into a well-controlled group (n = 98) and a poorly controlled group (n = 82). The well-controlled group was defined as having a glycated hemoglobin (HbA1c) level ≤6.5%, while the poorly controlled group was defined as having an HbA1c level >6.5% (17).

Inclusion criteria included: ① Meeting the diagnostic criteria for Type 2 Diabetes Mellitus (T2DM) (18); ② Age ≥18 years; ③ Duration of diabetes <10 years; ④ Ability to perform daily activities independently and use a smartphone to access the internet; ⑤ Complete electronic health records and questionnaires without missing data. Exclusion criteria included: ① Pregnant or lactating women; ② Individuals with mental or psychological disorders; ③ Presence of malignant tumors or severe complications.

First, based on multi-source big data, 15 core features were extracted as input variables for the model. These 15 core features are Age, Gender, Body Mass Index (BMI), Baseline glycated hemoglobin (HbA1c), Baseline fasting blood glucose (FBG), Duration of diabetes, Smoking status, Alcohol consumption history, High-density lipoprotein cholesterol (HDL-C), Triglyceride (TG) level, Physical activity level, Diet compliance, Frequency of blood glucose monitoring, Medication adherence, Presence of comorbidities.

To develop the risk prediction model, we evaluated the performance of several machine learning algorithms using the same training dataset and input features. The algorithms tested included logistic regression, support vector machine (SVM), random forest, and XGBoost (gradient boosting).

Model performance was assessed using 5-fold cross-validation. Evaluation metrics included accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC). Among the tested models, XGBoost demonstrated the highest AUC (0.861), the best balance between precision and recall, and superior performance in detecting high-risk patients with poor glycemic control. As a result, the XGBoost model was selected for the risk prediction submodule. Then, using the Python 3.8 platform and the Scikit-learn library, multiple algorithm performances were compared to select the optimal model:

This study strictly adhered to the Declaration of Helsinki and the ethical guidelines for medical research. All procedures were approved by the hospital’s medical ethics committee. All patient data, including electronic medical records, laboratory test results, dynamic data collected via smartphone, and scale scores, were anonymized using a unique patient identifier (ID). The raw data were encrypted and stored on a dedicated hospital server, accessible only to authorized research team members (such as nurses from the physical examination department and statistical analysts) after permission verification.

All data in this study were analyzed using SPSS statistical software (version 29.0; developed by SPSS Inc., Chicago, IL, United States). Continuous variables that passed normality tests were expressed as mean ± standard deviation (M ± SD), and group comparisons were performed using independent samples t-tests. Categorical variables were expressed as percentages (%), and group comparisons were conducted using χ² tests. A p < 0.05 was considered statistically significant. Variables with a p < 0.05 difference between the case–control group under personalized management were used as independent variables, with poor blood glucose control as the dependent variable, to perform multivariate logistic regression analysis to identify independent risk factors for poor blood glucose control under personalized management.

Table 3 shows significantly higher self-care scores across all domains (general diet, exercise, blood glucose monitoring, foot care, special diet, and medication) in the personalized management group compared to conventional management after 6 months (all p < 0.05), with the largest improvement in special diet (p = 0.001).

Overall adherence to self-management practices, as measured by the SDSCA, was significantly higher in the personalized management group across all domains, with particularly notable improvements in dietary and blood glucose monitoring behaviors. In addition, within the personalized group, adherence indicators derived from APP usage, such as login frequency, dietary/exercise plan completion, and blood glucose monitoring rates, were significantly associated with glycemic control outcomes.

Figure 2 demonstrates no baseline differences in DSQL dimensions (psychological: p = 0.625; physiological: p = 0.608; treatment: p = 0.517; social: p = 0.638). Superior quality of life in the personalized group at 3 and 6 months across all DSQL dimensions (psychological, physiological, treatment effectiveness, social relationships; all p < 0.05). The most pronounced difference was in psychological function at 6 months (2.25 ± 0.37 vs. 2.12 ± 0.32, p = 0.002).

The superior glycemic outcomes observed in the personalized management group align with the growing body of research suggesting that tailored interventions outperform standardized approaches (22–24). The modest but statistically significant HbA1c reduction at 6 months, while below thresholds often considered clinically impactful in isolation (25), gains importance when viewed alongside the concurrent significant improvements in self-care behaviors and QoL. While the absolute differences observed in glycemic parameters (e.g., HbA1c reduction of 0.15%) may appear modest, they are consistent with the effect sizes reported in behavioral and lifestyle interventions delivered in real-world settings. According to (25), even a 0.1–0.2% reduction in HbA1c is associated with meaningful reductions in long-term complication risk, especially when sustained over time. Moreover, when these metabolic improvements are accompanied by significant enhancements in self-care behavior and quality of life—as observed in this study—they may reflect broader behavioral change and risk reduction. Therefore, although statistically modest, the results may carry clinical significance when considered in the broader context of chronic disease management. This pattern suggests the model may foster sustainable behavioral changes rather than merely achieving short-term metabolic shifts (23).

The largest improvement in the “special diet” domain of the SDSCA is particularly noteworthy. This likely stem directly from the model’s core functionality: dynamically generating individualized meal plans based on personal preferences (analyzed via app dietary logs), BMI, and activity levels. This contrasts sharply with conventional dietary advice, which often provides generic templates that fail to account for individual tastes and practical constraints, leading to poor long-term adherence (12). Our findings resonate with studies emphasizing the critical role of dietary personalization in T2DM management and the difficulty of sustaining generic recommendations (26, 27).

The significant improvements in exercise adherence, blood glucose monitoring, foot care, and medication taking, as captured by the SDSCA, further elucidate potential mechanisms for the glycemic benefits. The model’s ability to recommend contextually appropriate exercise types and intensities (considering age, comorbidities like hypertension, and habits) and to automate reminders likely reduced barriers to initiation and persistence. Similarly, the adaptive adjustment of glucose monitoring frequency based on recent fluctuations provided actionable feedback without unnecessary burden, potentially enhancing patient engagement and understanding of their glycemic patterns. These findings support the hypothesis that AI-driven models can overcome key limitations of episodic care by offering continuous, tailored support that adapts to individual needs and contexts in real-time, thereby bridging the gap between clinic visits (28, 29).

The superior quality of life outcomes across all DSQL dimensions (psychological, physiological, treatment, social) in the personalized group, particularly the marked improvement in psychological function, underscores a holistic benefit beyond metabolic control. This aligns with research linking personalized support, clear goals, and reduced regimen burden to improved psychological wellbeing in diabetes (30–32). This likely reflects reduced diabetes-related distress, a common barrier to effective self-management. The model’s features—providing specific, achievable plans, reducing uncertainty through adaptive feedback, and potentially increasing perceived control over the condition—may have alleviated anxiety and frustration associated with complex self-care regimens.

The case–control analysis within the personalized management group revealed crucial insights into factors differentiating those achieving good versus poor glycemic control despite receiving the same advanced intervention. The identification of baseline low HDL-C and alcohol consumption as independent risk factors for poor control under personalized management reinforces established pathophysiological knowledge. Low HDL-C is a well-recognized component of diabetic dyslipidemia and insulin resistance, potentially rendering glycemic control more challenging irrespective of management intensity (33). Alcohol consumption can directly impair glycemic stability through effects on gluconeogenesis, insulin sensitivity, and hypoglycemia risk, while also potentially interfering with adherence to dietary and medication plans (33–35). These findings highlight that even sophisticated management models must contend with fundamental biological and behavioral factors.

More novel and specific to the digital, personalized approach were the findings related to engagement and adherence metrics. Lower adherence to the generated blood glucose monitoring plans emerged as a strong independent risk factor for poor control. This underscores the critical importance of acting upon the personalized recommendations; the model’s insights are only valuable if patients engage with the monitoring process that provides the necessary data for adaptation and feedback (36). Similarly, lower completion rates of the personalized exercise plans were associated with worse outcomes, emphasizing that tailored recommendations still require patient action. The protective effect of higher exercise plan adherence reinforces this point (37). These results highlight the essential role of adherence, not only to medication but also to behaviorally driven components like dietary and exercise plans, within AI-supported chronic disease management. Improving adherence in real-world practice may require multi-pronged strategies, such as personalized reminders, motivational messaging, adaptive goal-setting, and identifying early signs of disengagement through app usage analytics. AI models could be expanded to include adherence prediction modules, enabling proactive re-engagement interventions tailored to individual behavioral patterns. Interestingly, while APP login frequency and dietary plan adherence/completion rate showed significant univariate associations with better control, they did not retain significance as independent predictors in the multivariate model. This suggests that active engagement with the specific self-care actions (monitoring, exercising) prescribed by the model is more directly consequential for glycemic outcomes than general app usage or even dietary adherence within this specific framework, where dietary plans were highly personalized. The finding that patients requiring fewer dietary plan adjustments (potentially indicating initial plan suitability or stability) tended toward better control further hints at the interplay between personalization accuracy, patient engagement, and metabolic outcomes.

These process indicators offer a valuable new dimension for understanding intervention effectiveness in the digital health era. They move beyond simply measuring if an intervention is delivered to assessing how it is used and acted upon by the individual. This granular understanding of engagement is crucial for optimizing future iterations of AI-driven models, potentially allowing for early identification of individuals at risk of disengagement (e.g., low login frequency, declining adherence) and triggering proactive, tailored support strategies within the model itself (37, 38).

Several limitations warrant consideration. Firstly, the single-center, retrospective design limits generalizability and introduces potential selection bias. The requirement for smartphone literacy and internet access excludes digitally marginalized populations. The 6-month follow-up period is relatively short for assessing the long-term sustainability of benefits and prevention of chronic complications. Details of the AI model’s internal validation, feature importance for the XGBoost risk prediction, and the specific rules/reinforcement learning parameters within the plan generation submodel require fuller disclosure and external validation for reproducibility. The “black box” nature of complex AI models also poses challenges for clinical interpretability and trust.

Despite these limitations, this study provides promising evidence for the real-world application value of AI and big data in enhancing T2DM management. By demonstrating improvements in glycemic control, self-care, and QoL, alongside insights into key success factors within the model (engagement, adherence to specific actions, biological/behavioral baselines), it offers a foundation for future research. Larger, multi-center, prospective randomized controlled trials with longer follow-up are essential to confirm these findings and assess cost-effectiveness. Future work should focus on enhancing model interpretability, integrating predictive analytics for early intervention on adherence lapses identified through process indicators, and ensuring equitable access to such advanced digital health solutions. Integrating these AI tools as supportive complements within established, human-centered healthcare workflows, particularly empowering nursing-led management as demonstrated here, appears to be a promising pathway forward. Additionally, the cost of implementing and maintaining AI-driven tools and mobile health applications remains a relevant concern for real-world scalability. In this study, the mobile app and AI modules were developed in-house and deployed without cost to participants during the study period. However, future implementation may require resources for software maintenance, user support, and data security infrastructure. Although digital solutions may reduce long-term costs by improving outcomes and reducing complications, a formal cost-effectiveness analysis is necessary. Future studies should examine the economic feasibility of integrating such systems into routine care, particularly in resource-limited settings.