AI-Driven Clinical Decision Support System for Enhanced Diabetes Diagnosis and Management

Identifying type 2 diabetes mellitus can be challenging, particularly for primary care physicians. Clinical decision support systems incorporating artificial intelligence (AI-CDSS) can assist medical professionals in diagnosing type 2 diabetes with high accuracy. This study aimed to assess an AI-CDSS specifically developed for the diagnosis of type 2 diabetes by employing a hybrid approach that integrates expert-driven insights with machine learning techniques. The AI-CDSS was developed (training dataset: n = 650) and tested (test dataset: n = 648) using a dataset of 1298 patients with and without type 2 diabetes. To generate predictions, the algorithm utilized key features such as body mass index, plasma fasting glucose, and hemoglobin A1C. Furthermore, a clinical pilot study involving 105 patients was conducted to assess the diagnostic accuracy of the system in comparison to non-endocrinology specialists. The AI-CDSS showed a high degree of accuracy, with 99.8% accuracy in predicting diabetes, 99.3% in predicting prediabetes, 99.2% in identifying at-risk individuals, and 98.8% in predicting no diabetes. The test dataset revealed a 98.8% agreement between endocrinology specialists and the AI-CDSS. Type 2 diabetes was identified in 45% of 105 individuals in the pilot study. Compared with diabetes specialists, the AI-CDSS scored a 98.5% concordance rate, greatly exceeding that of nonendocrinology specialists, who had an 85% agreement rate. These findings indicate that the AI-CDSS has the potential to be a useful tool for accurately identifying type 2 diabetes, especially in situations in which diabetes specialists are not readily available.

💡 Research Summary

**
The paper presents the development and evaluation of an artificial‑intelligence‑driven clinical decision support system (AI‑CDSS) designed to assist clinicians in diagnosing type 2 diabetes mellitus (T2DM). Recognizing that primary‑care physicians often face difficulty in accurately identifying T2DM, the authors adopt a hybrid approach that merges expert‑driven clinical knowledge with data‑driven machine‑learning (ML) models, aiming to combine interpretability with high predictive performance.

Data and Pre‑processing
A total of 1,298 patient records (both diabetic and non‑diabetic) were collected. Each record includes demographic information (age, sex), anthropometric measures (BMI), laboratory values (fasting plasma glucose, HbA1c, lipid profile, renal function, insulin, C‑peptide, auto‑antibodies), and clinical observations (family history, blood pressure, physical signs). Missing values were imputed using the mean of available observations, and all continuous variables were standardized (z‑score). The dataset was split using stratified sampling into a training set of 650 patients (≈70 %) and a test set of 648 patients (≈30 %).

Expert‑Driven Knowledge Acquisition
Endocrinology specialists were consulted to codify current diagnostic guidelines into a decision‑tree structure called the Diabetes Clinical Knowledge Model (D‑CKM). The tree yields four possible outcomes: confirmed diabetes, pre‑diabetes, at‑risk, and no diabetes. Fourteen attributes (e.g., family history, physical activity, BMI, blood pressure, HbA1c, fasting glucose) constitute the nodes of the tree, providing a transparent, “white‑box” representation of clinical reasoning.

Machine‑Learning Model Development
Five ML algorithms were evaluated: Decision Tree (DT), Random Forest, J48, CHAID, and Classification and Regression Tree (CART). Feature selection employed Recursive Feature Elimination (RFE) and automatic relevance detection, identifying age, BMI, fasting plasma glucose, and HbA1c as the most informative predictors. Each algorithm was ranked based on three criteria—overall accuracy, number of extracted rules, and number of attributes used. CART achieved the highest composite score (accuracy = 89.8 %, ranking value = 0.798) and was therefore selected as the data‑driven component.

Hybrid Model Integration
The final AI‑CDSS combines the expert model (D‑CKM) and the ML model (CART) through a weighted linear fusion:

 H(x) = α·D‑CKM(x) + (1 − α)·hθ(x)

where α (0 ≤ α ≤ 1) controls the relative influence of the rule‑based and statistical components. This formulation preserves the interpretability of the expert tree while allowing the data‑driven model to capture patterns not explicitly encoded by clinicians.

Performance Evaluation
On the internal test set (n = 648), the hybrid system achieved an overall accuracy of 98.8 %, with sensitivity and specificity both exceeding 98 %. Agreement with endocrinology specialists was 98.8 %, indicating near‑perfect concordance.

A prospective clinical pilot involving 105 patients was then conducted. The AI‑CDSS correctly identified diabetes in 45 % of participants and demonstrated a Cohen’s kappa of 0.985 (98.5 % concordance) when compared with specialist diagnoses. In contrast, non‑specialist physicians achieved only 85 % agreement, highlighting the system’s potential to elevate diagnostic quality in settings lacking diabetes expertise.

Critical Assessment
The study showcases a well‑structured hybrid methodology that addresses a key limitation of many AI‑based CDSS: lack of transparency. By embedding expert rules, clinicians can trace the reasoning behind each recommendation, which is essential for trust and regulatory acceptance. The selection of CART as the ML backbone is appropriate given its own interpretability, further reinforcing the “white‑box” nature of the solution.

However, several limitations temper the enthusiasm for immediate clinical deployment. First, the dataset originates from a single institution and contains fewer than 1,300 records, raising concerns about external validity across diverse populations and health‑system contexts. Second, the reported near‑perfect accuracies (≈99 % for individual categories) may indicate overfitting, especially since cross‑validation details and external test sets are not provided. Third, the simplistic mean‑imputation for missing data ignores potential correlations among variables and could bias model estimates. Fourth, the pilot study’s sample size (105 patients) is modest, and the distribution of pre‑diabetes and at‑risk categories is not fully described, limiting insight into the system’s discriminative ability across the full disease spectrum. Finally, the paper does not elaborate on how the weighting parameter α was optimized, nor how conflicts between the expert tree and the CART predictions are resolved in practice—issues that could affect real‑world reliability.

Implications and Future Directions
Despite these caveats, the research provides compelling evidence that a hybrid AI‑CDSS can achieve high diagnostic performance while remaining interpretable. For primary‑care environments, especially in low‑resource settings where endocrinology expertise is scarce, such a tool could facilitate earlier detection of T2DM, enable timely lifestyle or pharmacologic interventions, and potentially reduce downstream complications. Future work should focus on:

1. Validating the system on multi‑center, multi‑ethnic cohorts to assess generalizability.
2. Implementing robust cross‑validation and external hold‑out testing to guard against overfitting.
3. Exploring more sophisticated imputation techniques (e.g., multiple imputation, model‑based methods).
4. Providing a transparent protocol for selecting and updating the α weight, possibly through Bayesian model averaging or reinforcement learning.
5. Conducting longitudinal studies to evaluate impact on patient outcomes, treatment adherence, and health‑care costs.

In summary, the paper contributes a thoughtfully engineered, hybrid AI‑CDSS that merges clinical expertise with machine learning, achieving impressive diagnostic concordance with specialists. While further validation is required, the approach holds promise for augmenting diabetes care in primary‑care settings worldwide.