Skip to main content
Back to timeline
medRxivSource publication:

Letting the model learn its own target: LIFT-MASLD reaches a 0.836 predictive correlation in an external cohort

Synopsis

Prognostic models for metabolic dysfunction-associated steatotic liver disease (MASLD) usually fix an endpoint first and then improve the predictor. LIFT-MASLD reverses that order: SCORE learns a weighted phenotype from eight hepatic measures that is most predictable from baseline characteristics, with sparse regression selecting the baseline variables. The phenotype reached a predictive correlation of 0.860 in the HALO cohort, above 0.763 for predicting future liver stiffness from all baseline variables, and 0.836 in the NIDDK external cohort without refitting.

Source-provided article image: Accurate Prediction of Hepatic Disease Severity in MASLD Using a Learned Phenotype
Figure 1 · medRxiv · Page 8

Interpretation

The prediction target itself can be learned: instead of prespecifying a single hepatic endpoint, the model learns a phenotype that weights several hepatic measures and then predicts that phenotype. Earlier MASLD prognostic models generally name a future endpoint such as liver stiffness or fibrosis stage, then raise accuracy by adding predictors or model flexibility. In the HALO-MASLD cohort of 332 participants and 456 follow-up observations, nested five-fold cross-validation gave a predictive correlation of r = 0.860 (95% CI [0.824, 0.896]), against r = 0.763 for predicting future liver stiffness from all 99 baseline predictor columns.

The learned phenotype concentrates weight on platelet count, bilirubin, alkaline phosphatase and liver stiffness, while albumin, AST and ALT receive zero weight. SCORE maximizes the correlation between the phenotype and its baseline prediction under nonnegative weights summing to 8, and replaces ordinary least squares on the predictor side with broken adaptive ridge regression. Of the 99 baseline predictor columns, only six had nonzero coefficients in the final predictor: four measured baseline characteristics, namely liver stiffness, bilirubin, alkaline phosphatase and platelet count, plus two missingness indicators.

Perspective

The demonstrated scope is the single-center prospective HALO-MASLD cohort of 332 participants and 456 follow-up observations, plus external validation without refitting in the NIDDK NAFLD Adult Database 2 cohort of 520 participants and 732 follow-up observations; the model is released as a web calculator at https://liftmasld.github.io.

The predictor uses a linear structure; the analysis includes only two subsequent follow-up visits, so whether the model-defined phenotype tracks outcomes such as hepatic decompensation, hepatocellular carcinoma, transplantation or liver-related mortality over longer follow-up remains to be determined; HALO is a predominantly White single-center cohort; and MRE liver stiffness was unavailable in NIDDK, so the corresponding association could not be replicated externally.