Predicting Liver Stiffness Approximately 12 Months Ahead Using Serial Transient Elastography

Background

Liver stiffness measurement (LSM) is widely used to assess chronic liver disease, but a single measurement does not capture longitudinal change. We evaluated whether serial LSMs and routine clinical indicators could predict continuous LSM approximately 12 months later.

Methods

This retrospective longitudinal study included patients with at least two historical LSMs and a follow‐up LSM 270–450 days after the index date. The LSM closest to 365 days was the outcome, and each patient contributed the earliest eligible task. Patients were separated chronologically into a train set (index date through 2022), test set (2023), and validation set (2024). Nested expanding temporal cross‐validation selected the model, which was then fixed and compared with longitudinal benchmarks. Patient‐level bootstrapping provided 95% confidence intervals (CIs).

Results

The analysis included 715 patients (train, 434; test, 178; validation, 103). Random forest was selected. In the test set, its mean absolute error (MAE) was 3.622 kPa (95% CI, 2.828–4.769), R ² was 0.456 (0.240–0.550), and calibration slope was 1.043 (0.612–1.475). In the validation set, MAE was 2.989 kPa (2.149–4.149), R ² was 0.432 (−0.141–0.575), and calibration slope was 0.951 (0.513–1.598). In the test set, the model reduced MAE by 8.443 kPa (4.833–12.605) versus linear‐trend extrapolation; however, the last observed LSM performed better, with a paired last‐observed‐LSM‐minus‐model MAE difference of −0.658 kPa (−1.168 to −0.187). Historical LSM summaries were the dominant features.

Conclusions

The locked model showed reproducible out‐of‐time performance and near‐ideal calibration point estimates but did not outperform the last observed LSM. Independent multicenter validation is required before clinical translation.

Trial Registration: ClinicalTrials.gov identifier: NCT06123858