A Unified Three-Stage Machine Learning Framework for Diabetes Detection, Subtype Discrimination, and Cognitive-Metabolic Hypothesis Testing

Technical Summary: A Unified Three-Stage Machine Learning Framework for Diabetes Detection, Subtype Discrimination, and Cognitive-Metabolic Hypothesis Testing

Problem Statement
Current machine learning (ML) approaches to diabetes mellitus prediction face three primary limitations: (1) Metric incompleteness, where studies often prioritize accuracy over clinically critical metrics like recall and ROC-AUC, potentially missing false negatives; (2) Lack of subtype discrimination, where Type 1 (T1DM) and Type 2 (T2DM) diabetes are frequently conflated into a single binary class despite distinct etiologies and management needs; and (3) Unexplored cognitive-metabolic links, specifically the "Type 3 diabetes" (T3DM) hypothesis, which posits a neurodegenerative pathway driven by central insulin resistance, remains computationally untested on public longitudinal data.

Methodology
The authors propose a reproducible, three-stage framework utilizing three distinct datasets: the NCSU Diabetes Dataset (mixed Indian cohort), the Pima Indians Diabetes Database (for benchmarking), and the Ohio Longitudinal Cognitive Dataset (n=373).

• Stage 1: Binary Diabetes Detection & Explainability

  • Data Preprocessing: Categorical symptoms were binarized; physiologically impossible zero-values were imputed with medians; outliers were filtered via IQR; and features were standardized.
  • Modeling: Five supervised classifiers (SVM-RBF, Logistic Regression, Random Forest, Extra Trees, Gradient Boosting) and a Stacking Ensemble were trained using stratified 5-fold cross-validation.
  • Evaluation: Models were assessed using six metrics: Accuracy, Balanced Accuracy, Precision, Recall, F1-score, and ROC-AUC.
  • Explainability: SHAP (SHapley Additive exPlanations) values were computed using a TreeExplainer on the top-performing tree ensemble to identify dominant predictive biomarkers.

• Stage 2: Unsupervised Subtype Clustering

  • Cohort: Restricted to confirmed diabetic cases (DO=1) from the NCSU dataset.
  • Features: Glucose, Insulin, and Age were selected based on physiological relevance to T1DM/T2DM differentiation.
  • Algorithm: K-Means clustering was applied with $k$ ranging from 2 to 8.
  • Validation: Optimal $k$ was determined using three internal validity indices: Silhouette Coefficient, Davies-Bouldin Index, and Calinski-Harabasz Score.

• Stage 3: T3DM Hypothesis Testing

  • Hypothesis 1 (H1): Tested for significant differences in Glycemic Control across three cognitive groups (Non-demented, Demented, Converted) using Kruskal-Wallis one-way ANOVA.
  • Hypothesis 2 (H2): Tested for a monotonic association between Glycemic Control and Cognitive Function (COG-FUNC) using Spearman rank correlation, with Holm-Bonferroni correction applied for multiple comparisons.

Key Results

• Binary Classification:

  • Performance: SVM-RBF and Logistic Regression achieved the highest ROC-AUC ($0.825 \pm 0.026$ and $0.825 \pm 0.034$, respectively). Random Forest achieved the highest raw accuracy ($0.762$) but the lowest recall ($0.556$), indicating a high rate of missed diabetic cases.
  • Test Set: On the held-out test set, SVM-RBF achieved a recall of $0.741$ and an ROC-AUC of $0.800$.
  • Explainability: SHAP analysis identified Glucose, BMI, and Age as the dominant predictive biomarkers. Features such as Polyphagia and Smoking showed marginal predictive contribution ($|SHAP| < 0.05$).

• Subtype Clustering:

  • Validation: A 2-cluster solution ($k=2$) was selected based on a Silhouette score of $\approx 0.116$ and clinical parsimony.
  • Profiles: Cluster 0 exhibited lower median Insulin and younger age (aligning with T1DM phenotypes), while Cluster 1 showed higher median Insulin and older age (aligning with T2DM phenotypes). The authors note the separation is moderate, reflecting the biological overlap between subtypes.

• Cognitive-Metabolic Association:

  • Group Differences: Kruskal-Wallis testing failed to find significant differences in Glycemic Control across cognitive groups ($p = 0.541$).
  • Correlation: A statistically significant positive Spearman correlation was found between Glycemic Control and Cognitive Function ($\rho_s = 0.208, p = 5.29 \times 10^{-5}$). This result survived Holm correction ($p_{Holm} = 1.06 \times 10^{-4}$), supporting a moderate positive association where better metabolic regulation co-occurs with higher cognitive function.

Significance and Claims
The paper claims to advance diabetes research along three axes:

1. Detection Performance: By prioritizing recall and ROC-AUC over raw accuracy, the framework addresses the clinical necessity of minimizing false negatives in screening.
2. Clinical Interpretability: The use of SHAP provides model-agnostic, post-hoc explanations that map directly to established clinical biomarkers, enhancing trust in ML outputs.
3. Multi-Type Scope: The framework moves beyond binary detection to offer an unsupervised proxy for subtype discrimination and provides the first computational, open-data corroboration of the T3DM hypothesis using statistical hypothesis testing on longitudinal cognitive data.

The authors emphasize that the framework is designed for reproducibility, with all code and data processing steps released as an annotated Jupyter notebook. They modestly note limitations, including the lack of ground-truth subtype labels for the clustering stage and the cross-sectional nature of the cognitive analysis, which limits causal inference regarding T3DM.