Identifying diagnostic indicators for type 2 diabetes mellitus from physical examination using interpretable machine...

Area of Science:

• Medical Informatics
• Machine Learning
• Diabetes Research

Background:

• Early identification of patients at risk for type 2 diabetes mellitus (T2DM) is crucial for preventing complications and reducing healthcare burden.
• Manual analysis of routine physical examination records is impractical for T2DM screening due to the condition's high prevalence.

Purpose of the Study:

• To develop interpretable machine learning models for T2DM diagnosis using physical examination indicators.
• To identify important diagnostic indicators, including age- and sex-related factors, for T2DM.

Main Methods:

• Three weighted diversity density (WDD)-based algorithms were developed for T2DM screening using physical examination indicators.
• Two of the WDD algorithms are designed to be tolerant of missing values, enhancing their applicability.
• A dataset of 43 physical examination indicators from 11,071 T2DM patients and 126,622 healthy controls was utilized.

Main Results:

• The algorithms identified the top 25% of indicators as directly or indirectly related to T2DM.
• Age- and sex-specific predictive markers were detected, revealing characteristic differences among T2DM patient groups.
• The models demonstrated strong performance in T2DM screening, achieving a maximum AUC of 0.9185.

Conclusions:

• Interpretable WDD-based algorithms were successfully employed to construct T2DM diagnostic models.
• The study identified age- and sex-related predictive markers, offering insights into group-specific T2DM characteristics.
• These findings highlight the potential of interpretable machine learning in T2DM risk stratification.