IDDF2024-ABS-0124 Cutting-edge deep learning models with domain adaptation outperform traditional methods in predicting liver-related complications in metabolic dysfunction-associated steatotic liver disease

Background

To develop innovative risk models for predicting liver-related events including hepatic decompensation and hepatocellular carcinoma in patients with metabolic dysfunction-associated steatotic liver disease (MASLD).

Methods

The training cohort included adult patients with MASLD from a territory-wide database in Hong Kong between January 2000 and July 2021. Five modern domain adaptation (DA) methods on fully connected neural networks were evaluated using the area under the time-dependent receiver operating characteristic curves (AUROCs) and compared with the FIB-4 index, NAFLD outcomes score (NOS), and a Fine-Gray model. The validation cohort comprised adult patients with type 2 diabetes (T2D) and probable MASLD, identified using previously developed NAFLD ridge score. We excluded patients with liver-related events before MASLD diagnosis or follow-up <6 months. This study was supported by the Health and Medical Research Fund (Reference number: 19202141).

Results

Among 25,166 patients with MASLD in the training cohort (mean age 56.9 years, 54.3% females, 0.7% cirrhosis), 272 (1.1%) developed liver-related events during 133,816 person-years (PYs). During 4,386,544 PYs among 411,395 patients in the validation cohort (mean age 61.8 years, 49.3% females, 0.4% cirrhosis), 5,984 (1.5%) developed liver-related events. Among the five DA methods, maximum classifier discrepancy (MCD) (AUROC [95% CI] 0.822 [0.814-0.829]) and confidence regularised self-training (CRST) (0.825 [0.817-0.832]) performed best in validation (IDDF2024-ABS-0124 Figure 1). The AUROC of the Fine-Gray model decreased from 0.804 in training to 0.681 in validation, demonstrating the advantage of DA in preserving model accuracy in a less definite MASLD population. Similarly, the AUROC of NOS and FIB-4 dropped to 0.649 and 0.645 in validation. Among the 19 factors, including common laboratory tests, comorbidities, and demographics in MCD and CRST, the eight leading factors were cirrhosis, diabetes, platelets, aspartate aminotransferase, gamma-glutamyl transferase, international normalised ratio, dyslipidaemia, and albumin. MCD labelled 78.6% of patients with T2D and MASLD as low risk, achieving a 99.2% negative predictive value for excluding liver-related events in 15 years.

Conclusions

Our novel models, integrating common clinical parameters, effectively identify low-risk individuals for liver-related events among patients with MASLD and patients with T2D and probable MASLD.