Document resource
Background We aimed to develop a prediction model for LREs in patients with T2DM, given the close relationship between diabetes and liver disease progression.Methods This retrospective cohort study included patients with T2DM from Hong Kong (HK-T2DM cohort) and patients with T2DM and metabolic dysfunction-associated steatotic liver disease (MASLD) from an international multicentre cohort undertaking vibration-controlled transient elastography (VCTE) examination (VCTE-Prognosis-T2DM cohort). The HK-T2DM cohort was randomly divided in an 80:20 ratio to form the development and internal validation cohorts. The VCTE-prognosis-T2DM cohort was used for external validation. The DM-LRE score was derived from 11 features including cirrhosis and routine serum markers at baseline ( IDDF2026-ABS-0242 Figure 1. SHAP importance and honeycomb diagram of DM-LRE score). GBM demonstrated the highest concordance index (C-index) and time-dependent area under the receiver operating characteristic curve among 3 potential machine learning algorithms in predicting LREs. The score was applied in risk stratification and a two-step algorithm followed by liver stiffness measurement (LSM) using cutoffs at 80% sensitivity and specificity.Results 575,000 patients were included in the HK-T2DM cohort (mean age 61.9 years; 52.3% males), while 3,587 patients were included in the VCTE-Prognosis-T2DM cohort (mean age 56.0 years; 56.8% males). In the development and external validation cohorts, the C-indices of the DM-LRE score in predicting 5-year LRE were 0.752 (95% CI: 0.742-0.763) and 0.864 (95% CI: 0.806-0.926), respectively, outperforming the Fibrosis-4 index (FIB-4, C-indices: 0.674 and 0.830) and XgBoost and RandomForest models. The DM-LRE score displayed a comparable C-index to LSM in the external cohort (0.864 vs. 0.866).Patients were stratified into low-, intermediate- and high-risk categories by the DM-LRE score with distinct 5-year cumulative incidence (p<0.001, IDDF2026-ABS-0242 Figure 2. 5-year risk-stratified metrics and two-step algorithms using DM-LRE cut-offs at 80% sensitivity and specificity). For the two-step algorithms in the VCTE-prognosis-T2DM cohort, DMLRE demonstrated a comparable negative predictive value to FIB4 (5 years: 99.5% vs. 99.6%), while enabling a larger proportion of patients to be classified as low-risk at the primary step (68.6% vs. 58.3%) and ultimately (83.9% vs. 78.8%), with less than a 0.1% increase in the 5-year LRE incidence in the final low-risk group (IDDF2026-ABS-0242 Figure 2. 5-year risk-stratified metrics and two-step algorithms using DM-LRE cut-offs at 80% sensitivity and specificity).Conclusions The DM-LRE score could identify individuals at risk of LREs in patients with T2DM and facilitate early referral to specialty care.Abstract IDDF2026-ABS-0242 Figure 1Abstract IDDF2026-ABS-0242 Figure 2