ORCID

Abstract

Background: Chronic liver disease (CLD) is increasing globally, driven by metabolic and alcohol-related aetiologies. Late diagnosis is common due to a prolonged, subclinical fibrotic phase preceding advanced disease (ACLD). Current pathways and cross-sectional fibrosis tests have poor sensitivity in low-prevalence settings and limited population-scale availability. Predicting future ACLD in primary care offers opportunities for earlier diagnosis and prevention, but few suitable clinical prediction models (CPMs) exist.Aim: To review existing general-population liver prediction models and develop CPMs for incident ACLD in unselected primary care populations using routinely collected data.Methods: A systematic review and PROBAST+AI appraisal of fourteen CPMs (median external Harrell's C = 0.76, IQR 0.72–0.80) revealed limitations including spectrum bias, poor calibration, and over-reliance on liver function tests (LFTs). An ACLD phenotype algorithm was derived from SNOMED/Read codes and validated by retrospective case-note review (sensitivity 91.7%, specificity 95.1%). A retrospective cohort of 3.85 million adults without prevalent ACLD from the Optimum Patient Care Research Database (OPCRD) was followed for a median of 7.5 years, with 13,645 incident ACLD events. Patients were stratified by laboratory test availability; Cox proportional hazards, XGBoost Survival, random survival forests, and DeepHit models were compared under cause-specific competing-risks frameworks, with multiple imputation and native missingness handling.Results: In a geographic test cohort, non-laboratory models predicted ACLD effectively (Harrell's C = 0.760–0.781). Adding LFTs and gamma-glutamyl transferase increased discrimination to C = 0.833–0.854 (Tier B) and 0.888–0.903 (Tier C). An XGBoost model with native-missingness handling achieved C = 0.803 and a 5-year time-dependent AUC 0.851, integrating laboratory data where available.Conclusions: ACLD risk can be accurately stratified across primary care populations without additional blood tests. Implementing these models could shift from reactive secondary-care diagnosis to automated case-finding and proactive prevention.

Awarding Institution(s)

University of Plymouth

Supervisor

Matthew Cramp, Ashwin Dhanda, William Henley

Keywords

chronic liver disease, cirrhosis, clinical prediction model, machine learning, Systematic Review, Primary care, risk assessment, Big Data

Document Type

Thesis

Publication Date

2026

Embargo Period

2026-09-24

Deposit Date

September 2026

Creative Commons License

Creative Commons Attribution-NonCommercial 4.0 International License
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License

Download

Share

COinS