Data Sheet 1_A clinically interpretable machine learning model for early detection of diab

1,475 diabetic patient records from three community health centers in China were used to develop a machine learning model for early diabetic retinopathy detection. The model, developed by Juncheng Tong, achieved an AUROC of 0.770 on a held-out test set of 298 patients. The dataset, last updated in 2026, uses routinely collected health variables like urine glucose for risk prediction.

Use Cases

• Benchmarking machine learning models for clinical prediction based on routinely collected health data mentioned in the description
• Analyzing feature importance for diabetic retinopathy risk factors based on SHAP analyses described
• Studying model calibration and decision curve analysis in a community health setting as detailed in the results

Strengths

• Includes performance metrics for three ML models (GLMNET, XGBoost, Ranger) with specific AUROC and AUPRC values
• Based on a cohort of 1,475 patients with a defined development/test split (n=1,177 / n=298)
• Model is designed for clinical interpretability using SHAP and uses routinely collected variables

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count for the underlying data is unknown, which may limit suitability assessment
• The model showed moderate discrimination (AUROC 0.770) and imperfect calibration, requiring external validation

Provenance

Source

figshare

Collection Method

Retrospective study using data from three community health centers

Time Range

null

Freshness

Last updated 2026-05-04 04:24:12; freshness should be verified

Geography

China