Work

Interpretable Machine Learning Predicts Hepatic Fat Response to a Monounsaturated Fat-Enriched Diet in Steatotic Liver Disease

Year: 2026

Type: preprint

Abstract: Background: Metabolic dysfunction-associated steatotic liver disease (MASLD) varies in response to dietary monounsaturated fatty acids (MUFA). We conducted a secondary analysis of the Habitual Diet vs Avocado Trial (HAT; NCT03528031; registered 17 May 2018; n = 790) to test whether machine learning using routinely collected clinical variables can identify MUFA-responsive phenotypes. Avocado versus habitual diet did not significantly change hepatic-fat fraction (ΔHFF + 0.48 percentage points, 95% CI − 0.24 to 1.22; p = 0.20) or visceral adipose tissue volume after FDR correction. Prespecified interaction tests showed sex- and glycemic-status-specific responses (FDR-q ≤ 0.05). MASLD classification reached AUROC 0.83 ± 0.03 (XGBoost) but its expected calibration error (ECE = 0.089) exceeded the prespecified ≤ 0.05 target, indicating that probability outputs require post-hoc Platt or isotonic recalibration before any thresholded clinical use. Two related HFF prediction tasks were evaluated: (i) the primary trial-cohort endpoint of prespecified ≥ 5 percentage-point absolute hepatic fat change versus no such change (either direction; n = 790; 261 positive), for which discrimination was 0.67 ± 0.03 with baseline biomarkers and rose to ~ 0.75 ± 0.03 when baseline MASLD status was carried forward as an input feature; and (ii) a translation-focused reduction-only endpoint (≥ 5 pp absolute decrease; listwise complete predictors, n = 790; 74 positives), summarized in Fig. 7, for which XGBoost achieved held-out AUROC = 0.792. Gains under (i) reflect sequential staging plus baseline-outcome correlation; confusion matrices must use endpoint-specific true-class row totals. Reporting follows TRIPOD + AI. Internally validated; external validation is required before clinical decision-support deployment.

Source: Research Square

DOI: 10.21203/rs.3.rs-9794767/v1