Insulin Resistance Prediction From Wearables and Routine Blood Biomarkers

TL;DR

Insulin resistance poses a key early risk for type 2 diabetes. This study develops a deployable, multimodal predictor that combines wearable-derived biomarkers with routine blood biomarkers to estimate IR via ground-truth HOMA-IR in a large US cohort. The best-performing predictor achieves $R^2\approx 0.50$ and $AUROC\approx 0.80$, with high sensitivity and specificity, particularly among obese and sedentary individuals, and generalizes to an independent validation cohort. An accompanying IR Literacy and Understanding Agent (IR Agent), based on a ReAct-enabled LLM, grounds IR predictions in patient data to provide safe, personalized metabolic health guidance. Collectively, the work demonstrates a scalable approach for early IR detection to enable timely lifestyle or therapeutic interventions and prioritizes clinical testing where needed.

Abstract

Insulin resistance, a precursor to type 2 diabetes, is characterized by impaired insulin action in tissues. Current methods for measuring insulin resistance, while effective, are expensive, inaccessible, not widely available and hinder opportunities for early intervention. In this study, we remotely recruited the largest dataset to date across the US to study insulin resistance (N=1,165 participants, with median BMI=28 kg/m2, age=45 years, HbA1c=5.4%), incorporating wearable device time series data and blood biomarkers, including the ground-truth measure of insulin resistance, homeostatic model assessment for insulin resistance (HOMA-IR). We developed deep neural network models to predict insulin resistance based on readily available digital and blood biomarkers. Our results show that our models can predict insulin resistance by combining both wearable data and readily available blood biomarkers better than either of the two data sources separately (R2=0.5, auROC=0.80, Sensitivity=76%, and specificity 84%). The model showed 93% sensitivity and 95% adjusted specificity in obese and sedentary participants, a subpopulation most vulnerable to developing type 2 diabetes and who could benefit most from early intervention. Rigorous evaluation of model performance, including interpretability, and robustness, facilitates generalizability across larger cohorts, which is demonstrated by reproducing the prediction performance on an independent validation cohort (N=72 participants). Additionally, we demonstrated how the predicted insulin resistance can be integrated into a large language model agent to help understand and contextualize HOMA-IR values, facilitating interpretation and safe personalized recommendations. This work offers the potential for early detection of people at risk of type 2 diabetes and thereby facilitate earlier implementation of preventative strategies.

Figures (6)

• Figure 1: Study Design and Data Summary. (A) Overview of physiological factors and associated lifestyle factors leading to insulin resistance, prediabetes and diabetes. (B) Illustration of our proposed modeling pipeline for predicting HOMA-IR, and interpreting the results with the Insulin Resistance Education and Understanding Large Language Model. (C) Correlation of blood biomarkers and lifestyle features (continuous values) with HOMA-IR. (D) Scatter plot of BMI and HOMA-IR values which signifies the relationship between higher BMI values and insulin resistance (measured through HOMA-IR). (E, F, G) Distribution of top three highly-correlated wearables features with HOMA-IR, namely Resting Heart Rate, Daily Step Counts, and Heart Rate Variability, for stratified insulin sensitivity groups (insulin sensitive, impaired insulin sensitivity, and insulin resistance). (H, I, J) Distribution of top three highly-correlated blood biomarkers (Triglycerides, HDL and LDL cholesterol) for stratified insulin sensitivity groups (same as E, F, G).
• Figure 2: Performance Evaluation of HOMA-IR prediction (Regression). (A) Comparison of HOMA-IR regression across input feature sets and models. (B, C, D, E) Scatter plots of predicted HOMA-IR values versus the true HOMA-IR models for selected feature sets. Areas of concern for true positive and false negative are highlighted as light green and rosy brown, respectively.
• Figure 3: Performance Evaluation of Insulin Resistance Prediction (Classification). (A) Performance of our binary classification model for various input features for identifying insulin resistant individuals (using MAE + L1-L2 learners), measured through Area Under the Receiver Operating Characteristic curve (AUROC), Sensitivity, Specificity, Precision, and Area Under the Precision-Recall Curve (AUPRC). (B) Visualization of the ROC curves for various feature sets across five cross validation folds. Average values are colors, with the gray areas around each line indicating the standard deviation across the five folds. (C) Visualization of the Precision-Recall curve for selected feature sets. Average values are colors, with the gray areas around each line indicate the standard deviation across the five folds.
• Figure 4: Interpretability and Stratification. (A) Sankey diagram showing the relative feature importance (SHapley Additive exPlanations [SHAP] values) for each of the proposed nonlinear XGBoost models for direct regression. (B, C) Qualitative evaluation of learned latent space’s interpretability of learning important features. The t-SNE reduced latent space shows that individuals with higher BMI and resting heart rate are clustered closely together in space, following our quantitative results of classifying high BMI and high-RHR individuals using these learned representations. (D, E) Distribution of individuals stratified by insulin resistance class and BMI classes (D) and insulin resistance versus physical activity classes as determined by number of daily steps (E). (F) Results of classification performance for various lifestyle stratification.
• Figure 5: Validation of Proposed Models on an Independent External Validation Cohort. (A) Overview of the study for the independent validation cohort. (B) Summary of the population characteristics of the external validation cohort. (C) Accuracy of selected ML models on the external validation cohort, including a side-by-side comparison with the results on the initial training cohort. We validated all of our trained models on this cohort except models with complete metabolic panel (CMP), since the external validation cohort did not include CMP in the blood tests.