Multimodal AI & Glucose Spikes: Diabetes & Prediabetes
Unlocking Glycemic Risk: A Multimodal XGBoost Approach for Type 2 Diabetes Classification
Table of Contents
Abstract: Early identification of individuals at risk for Type 2 Diabetes (T2D) is crucial for effective intervention and management.This study leverages the power of gradient boosting decision trees (XGBoost) to build robust binary classifiers capable of distinguishing between T2D and normoglycemic individuals. We explore the predictive capabilities of various data modalities, including demographics, anthropometrics, continuous glucose monitoring (CGM) data, food intake patterns, gut microbiome composition, genomic details, Fitbit-derived activity data, and electronic health record (EHR) conditions. By systematically evaluating different feature configurations and employing rigorous validation techniques such as leave-one-person-out cross-validation and bootstrap percentile-based confidence intervals, we demonstrate the efficacy of XGBoost in capturing complex, non-linear relationships within the data. Feature importance is elucidated through Shapley Additive Explanations (SHAP) to provide interpretable insights into the drivers of T2D classification. Furthermore, a multimodal classification model is developed and applied to predict glycemic risk in prediabetic individuals, offering a novel approach to complement traditional HbA1c measurements.
Introduction
Type 2 diabetes (T2D) is a growing global health concern characterized by hyperglycemia resulting from insulin resistance and impaired insulin secretion. It’s progressive nature and association with numerous comorbidities necessitate early detection and intervention strategies. While traditional diagnostic methods like fasting plasma glucose and HbA1c are valuable,they may not fully capture the dynamic physiological changes that precede overt T2D. The advent of wearable sensors,advanced omics technologies,and thorough electronic health records (EHRs) has opened new avenues for a more nuanced understanding of T2D pathogenesis and risk stratification.
machine learning,particularly tree-based ensemble methods,has shown notable promise in analyzing complex biological data for disease prediction. Gradient Boosting Decision Trees (XGBoost) stands out due to its efficiency, scalability, and ability to handle non-linear relationships, making it an ideal candidate for dissecting the intricate interplay of factors contributing to T2D.This study aims to develop and validate XGBoost-based classifiers using a diverse array of data modalities to accurately distinguish between individuals with T2D and normoglycemic controls. We further investigate the contribution of individual features to the classification process and explore the potential of a multimodal approach to predict glycemic risk in at-risk populations.
Methods
Participant Cohorts and Data Collection
This study utilized data from the PROGRESS cohort, comprising individuals diagnosed with T2D and normoglycemic controls.Additionally, prediabetic individuals from the PROGRESS cohort and both normoglycemic and prediabetic individuals from the HPP cohort were included for multimodal model evaluation. Data modalities encompassed:
Base Features: Age, sex, and Body Mass Index (BMI).
Glucose Spike Variables: Derived from continuous glucose monitoring (CGM) data, capturing postprandial glycemic excursions.
Food Intake Variables: Quantifying dietary patterns and meal composition.
Microbiome Variables: Characterizing the composition and diversity of the gut microbiome.
Genomic Variables: Including relevant genetic markers associated with T2D risk.
Fitbit variables: Activity levels, sleep patterns, and heart rate data from Fitbit devices.
EHR Condition Variables: Diagnoses, medications, and laboratory results extracted from electronic health records.
Binary Classification Model Progress
Binary classifiers were developed using XGBoost to differentiate between T2D and normoglycemic individuals. The model was chosen for its ability to capture possibly nonlinear relationships among variables and its relative simplicity and data efficiency compared to other nonlinear models.
Feature Configuration and Tuning: Different configurations of input features were considered, including:
‘Base’ features (age, sex, BMI)
‘Spikes’ (glucose spike variables)
‘food’ (food intake variables)
‘Microbiome’
‘Genomics’
‘Fitbit’
‘EHR’
The model’s hyperparameters were tuned separately for each input configuration using a preliminary grid search cross-validation loop with five splits.
performance Assessment: Model performance was assessed using a leave-one-person-out (LOPO) cross-validation scheme. In this approach,the data of each participant were excluded from training and used solely for testing,providing a robust evaluation of generalization. The receiver operating characteristic (ROC) curve and the corresponding area under the curve (AUC) were used as primary performance metrics.statistical Importance and Confidence Intervals: A bootstrap percentile-based method with 10,000 iterations was employed to estimate 95% confidence intervals for the AUC values.