Early detection of type 2 diabetes mellitus using machine learning-based prediction models
Problem
Detecting undiagnosed type 2 diabetes early, when existing screening tests rely on simplified regression-based scoring formulas.
Why it matters
Undiagnosed diabetes leads to severe complications and high healthcare costs.
What the AI did
Machine learning models (Glmnet, RF, XGBoost, LightGBM) were trained on electronic data to predict fasting plasma glucose, evaluated over 100 bootstrap iterations on 6-month data batches.
Results
With 6 months of data, simple regression had lowest average RMSE 0.838, then RF 0.842, LightGBM 0.846, Glmnet 0.859, XGBoost 0.881; Glmnet improved most (+3.4%) with more data.
Impact
ML matched but did not outperform simple regression initially, improving only as data grew.
Record
| Venue | Scientific Reports |
|---|---|
| Published | 2020 |
| Work type | article |
| Authors | Leon Kopitar, Primož Kocbek, Leona Cilar, Aziz Sheikh, Gregor Štiglic |
| Citations | 366 (OpenAlex, at import time) |
| References | 50 |
| Model or method | Glmnet, RF, XGBoost, LightGBM |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1038/s41598-020-68771-z |
| Open access | gold |
| Keywords | type 2 diabetes, machine learning, prediction, screening, RMSE, bootstrap |
Importance rating
significant — thousands to millions affected
Early detection of undiagnosed T2DM matters clinically, but abstract reports no clinically relevant improvement over simple regression; scale of affected population not stated.