Enhancing Antidiabetic Drug Selection Using Transformers: Machine-Learning Model Development.

Area of Science:

• Artificial Intelligence in Medicine
• Computational Health
• Diabetes Management

Background:

• Diabetes affects millions globally, with primary care physicians managing a significant patient load.
• Physicians often face challenges in selecting appropriate diabetes medications for individual patients.

Purpose of the Study:

• To develop a predictive model for diabetes drug prescriptions by endocrinologists.
• To create a decision-support system for non-specialists in choosing diabetes medications.
• To improve diabetes treatment outcomes through enhanced medication selection.

Main Methods:

• A transformer-based encoder-decoder model was developed to predict prescriptions for 44 diabetes drugs.
• Input data included patient age, sex, 12 laboratory test results, and medication history.
• Model performance was evaluated using electronic health records from 7034 patients, comparing 2, 5, and 10-year training data subsets.

Main Results:

• The model trained on 5 years of recent data achieved a micro-averaged ROC-AUC of 0.993 and macro-averaged ROC-AUC of 0.988.
• The model successfully predicted 43 out of 44 drugs with an ROC-AUC above 0.95.
• Performance surpassed the target of 0.95 ROC-AUC and outperformed the LightGBM model.

Conclusions:

• The AI model accurately predicts endocrinologist-prescribed diabetes drugs, demonstrating feasibility.
• This tool can assist non-specialists in making informed diabetes treatment decisions.
• Future work includes incorporating contraindications and expanding data to multiple institutions for generalizability.