Article Text
PDF
Abstract
Background We aimed to develop a prediction model for liver-related events (LREs) in patients with type 2 diabetes mellitus (T2DM).
Methods This territory-wide retrospective study included adult patients with the diagnosis of T2DM in Hong Kong between 2007 and 2021. Excluded patients were pre-existing liver diseases (except metabolic dysfunction-associated steatotic liver disease) and those who developed LREs before their T2DM diagnosis. The primary endpoint was incident LREs in 5 years, defined as hepatic decompensation or hepatocellular carcinoma (HCC). A conventional statistical risk model was developed using Cox regression with either stepwise or LASSO variable selection. Four machine learning algorithms were tested using the same covariates selected by the stepwise Cox model. The final model was selected based on the time-dependent area under the receiver-operating characteristics curves (tAUCs). The designated model was applied to stratify the patients using cutoffs at 90% sensitivity and specificity.
Results 575,000 patients with T2DM were included (mean age 61.9 years; 52.3% males); the incidence rate of LREs was 1.279 per 1,000 person-years. The conventional model developed by stepwise-Cox regression included 9 covariates. Among the machine learning approaches evaluated, XGBoost demonstrated the highest discrimination in the validation cohort and was selected as the final model. The tAUCs for XGBoost in the validation cohort at 1, 3, and 5 years were 0.805, 0.782 and 0.759 respectively, which were superior to the performance of the Fibrosis-4 index and Aspartate aminotransferase-to-Platelet Ratio Index (IDDF2024-ABS-0367 Figure 1. Discrimination of Multiple Risk Models in predicting 5-year outcomes in the training and validation cohorts). In the validation cohort (n=115000), 36.8%, 53.5%, and 9.7% of patients were stratified into low-, mediate- and high-risk categories, respectively. The corresponding 5-year incidence rates per 1000 person-years were 0.285, 1.114, and 4.658, respectively. The age- and sex-adjusted hazard ratios for LRE were significantly elevated in the intermediate- (3.25; 95% CI 2.47-4.29) and high-risk (13.78; 95% CI 10.35-18.35) versus low-risk categories.
Discrimination of multiple risk models in predicting 5-year outcomes in the training and validation cohorts
Conclusions This novel risk model may identify individuals at risk of adverse liver outcomes among patients with T2DM.
FIB-4, Fibrosis-4 index. APRI, AST-to-Platelet Ratio Index. The legend included the Integrated time-dependent AUCs of models.