SynthA1c: Towards Clinically Interpretable Patient Representations for Diabetes Risk Stratification

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Diabetes Mellitus Research

Background:

• Early diagnosis of Type 2 Diabetes Mellitus (T2DM) is critical for effective management.
• Clinical office visits are brief, necessitating efficient patient screening methods.
• Medical imaging data is increasingly accessible for clinical applications.

Purpose of the Study:

• To investigate the use of image-derived phenotypic data for automated T2DM risk prediction.
• To develop a classifier model that flags high-risk patients without requiring blood laboratory measurements.
• To introduce a novel metric for evaluating model generalizability across diverse patient populations.

Main Methods:

• Leveraged neural networks and decision tree models for T2DM risk classification.
• Developed 'SynthA1c' latent variables to mimic hemoglobin A1c measurements.
• Employed data augmentation techniques to assess model performance on out-of-domain covariates.

Main Results:

• Achieved sensitivities as high as 87.6% in predicting T2DM risk using image-derived phenotypes and physical examination data.
• Demonstrated accurate prediction of diabetes risk through an AI-enabled, image-based approach.
• Introduced a generalizable metric for evaluating model performance on unseen data.

Conclusions:

• Image-derived phenotypes combined with physical examination data can accurately predict diabetes risk.
• AI and medical imaging offer a powerful tool for opportunistic T2DM risk stratification.
• This automated approach facilitates early identification of patients needing further diagnostic workup for T2DM.