A precision-constrained framework for evaluating noninvasive biomarkers in MASLD and beyond

In patients with Metabolic Dysfunction-associated Steatotic Liver Disease (MASLD), accurate and low-cost non-invasive risk stratification remains a major unmet need. We developed a clinical records-based neural network integrating patient history, routine laboratory tests, and ultrasound imaging features. Here we show that in an internal test set (n = 209), the model achieved receiver operating characteristic area under the curve (ROC-AUC) values of 0.85 vs 0.82 (F ≥2), 0.90 vs 0.86 (F ≥3), 0.96 vs 0.89 (F = 4) compared with Fibrosis-4 (FIB-4). To address limitations of ROC-AUC, we applied RP-AUC0.5−0.7, a recall-precision metric focused on clinically relevant precision range, showing improved performance over FIB-4. External validation (n = 194) shows reduced liver biopsy failure rate from 86.6% to 50.0% for at-risk metabolic dysfunction-associated steatohepatitis (MASH) prediction. Our work presents a low-cost neural network improving FIB-4, introduces RP-AUC0.5−0.7 for biomarker comparison, and provides a generalizable framework for evaluating screening biomarkers in clinical care and drug trials. The study proposed a precision-constrained framework using routine clinical data and ultrasound to improve noninvasive risk stratification in metabolic dysfunction-associated steatotic liver disease, outperforming FIB-4, reducing liver biopsy burden.