New ML framework predicts NAFLD risk with 91.3% accuracy, outperforming deep learning
A gradient boosting model with conformal prediction achieves 0.912 AUROC and distribution-free coverage guarantees.
Researchers have developed a machine-learning framework that could transform screening for non-alcoholic fatty liver disease (NAFLD), which affects 25% of adults globally. The method, developed by Xinze Zhang, pairs gradient-boosted decision trees with conformal prediction to issue calibrated, distribution-free coverage guarantees on individual risk estimates. Using a multicenter cohort of 2,187 patients from Guangzhou, China (plus 412 for external validation) and 78 candidate features, the model achieved an AUROC of 0.912 on internal data and 0.891 externally—surpassing deep neural networks, TabNet, support vector machines, and logistic regression.
A key innovation is the use of mutual-information-based stability selection to identify a compact, clinically interpretable subset of six features: waist circumference, ALT, GGT, triglycerides, fasting glucose, and BMI. The conformal prediction sets achieved 91.3% empirical coverage at the 90% nominal level, providing provable statistical guarantees. The model enables a three-tier risk stratification, with the high-risk subgroup showing a 12-month disease progression rate 4.7 times that of the low-risk tier. This work addresses a critical gap in population-level NAFLD screening by offering both high accuracy and distribution-free confidence intervals.
- Model achieves 0.912 AUROC internally and 0.891 externally, beating deep neural networks and logistic regression.
- Conformal prediction provides 91.3% empirical coverage at 90% nominal level with distribution-free guarantees.
- High-risk subgroup shows 4.7x faster 12-month progression rate using six clinically interpretable features.
Why It Matters
This ML framework could enable widespread, reliable NAFLD screening using standard blood tests and simple measurements.