Accurate Prediction of Hepatic Disease Severity in MASLD Using a Learned Phenotype

Background & Aims: Most prognostic models for metabolic dysfunction-associated steatotic liver disease (MASLD) target prespecified hepatic endpoints, potentially limiting attainable accuracy. We developed and externally validated a model that learns the future hepatic phenotype most reliably predictable from baseline clinical characteristics. Methods: We analyzed 332 HALO-MASLD participants (456 follow-up observations). Sparse Canonical Outcome REgression (SCORE) learned a weighted phenotype from eight hepatic measures, coupled with broken adaptive ridge regression using 99 baseline predictor columns. Participant-level nested five-fold cross-validation compared prediction of the learned phenotype with future liver stiffness measurement (LSM), controlled attenuation parameter (CAP), and FIB-4 benchmarks. Models were externally validated without refitting in the independent NIDDK NAFLD Adult Database 2 cohort (520 participants, 732 follow-up observations). Biological validity and bootstrap stability were also assessed. Results: The learned phenotype emphasized lower platelet count and higher bilirubin, alkaline phosphatase, and LSM. In HALO, predictive correlation was r=0.860 (95% CI [0.824, 0.896]) and predictive R^2=0.671, versus r=0.763 for LSM and r=0.469 for CAP predicted from all baseline variables. All six prespecified comparisons remained significant after Holm correction (p_Holm < 0.05). The phenotype correlated with MRE stiffness (rho = 0.694) and biopsy fibrosis stage (rho = 0.567), and was higher with imaging evidence of portal hypertension (p < 0.001). Bootstrap relearning yielded mean phenotype correlation 0.947. In NIDDK, predictive correlation was r = 0.836 (95% CI [0.805, 0.863]; predictive R^2 = 0.680), and associations with fibrosis and portal hypertension replicated. Conclusions: Learning the prediction target produced a sparse, clinically interpretable hepatic phenotype more accurately predicted than conventional longitudinal MASLD endpoints in both cohorts. The resulting LIFT-MASLD model is available at https://liftmasld.github.io.

Authors

Institutions

Publication Details

Journal
medRxiv
Published
2026-11-17
DOI
https://doi.org/10.64898/2026.11.14.26368091
Primary Topic
Liver Disease Diagnosis and Treatment
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Accurate Prediction of Hepatic Disease Severity in MASLD Using a Learned Phenotype

Eric V. Strobl, Mark T. Miedel, Jaideep Behari, D. Lansing Taylor
6 links
medRxiv
Liver Disease Diagnosis and Treatment
preprint

Accurate Prediction of Hepatic Disease Severity in MASLD Using a Learned Phenotype

Eric V. Strobl, Mark T. Miedel, Jaideep Behari, D. Lansing Taylor
preprint en

Abstract

Background & Aims: Most prognostic models for metabolic dysfunction-associated steatotic liver disease (MASLD) target prespecified hepatic endpoints, potentially limiting attainable accuracy. We developed and externally validated a model that learns the future hepatic phenotype most reliably predictable from baseline clinical characteristics. Methods: We analyzed 332 HALO-MASLD participants (456 follow-up observations). Sparse Canonical Outcome REgression (SCORE) learned a weighted phenotype from eight hepatic measures, coupled with broken adaptive ridge regression using 99 baseline predictor columns. Participant-level nested five-fold cross-validation compared prediction of the learned phenotype with future liver stiffness measurement (LSM), controlled attenuation parameter (CAP), and FIB-4 benchmarks. Models were externally validated without refitting in the independent NIDDK NAFLD Adult Database 2 cohort (520 participants, 732 follow-up observations). Biological validity and bootstrap stability were also assessed. Results: The learned phenotype emphasized lower platelet count and higher bilirubin, alkaline phosphatase, and LSM. In HALO, predictive correlation was r=0.860 (95% CI [0.824, 0.896]) and predictive R^2=0.671, versus r=0.763 for LSM and r=0.469 for CAP predicted from all baseline variables. All six prespecified comparisons remained significant after Holm correction (p_Holm < 0.05). The phenotype correlated with MRE stiffness (rho = 0.694) and biopsy fibrosis stage (rho = 0.567), and was higher with imaging evidence of portal hypertension (p < 0.001). Bootstrap relearning yielded mean phenotype correlation 0.947. In NIDDK, predictive correlation was r = 0.836 (95% CI [0.805, 0.863]; predictive R^2 = 0.680), and associations with fibrosis and portal hypertension replicated. Conclusions: Learning the prediction target produced a sparse, clinically interpretable hepatic phenotype more accurately predicted than conventional longitudinal MASLD endpoints in both cohorts. The resulting LIFT-MASLD model is available at https://liftmasld.github.io.

medRxiv
University of Pittsburgh (US)
Good health and well-being
Liver Disease Diagnosis and Treatment
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon