(W-084) Exploring the feasibility of using AI to identify patient characteristics predictive of histological endpoints in metabolic dysfunction associated steatohepatitis (MASH)

Lianzhe Xu, Ph.D, – Prin. Scientist, Merck & Co Inc; Annaswamy Raji, Ph.D, – Sr. Prin. Scientist, Merck & Co Inc; Ira Gantz, Ph.D, – Dist. Scientist, Merck & Co Inc; Islam Younis, Ph.D, – Sr. Dir, Merck & Co Inc; Maria Trujillo, Ph.D, – Sr Dir, Merck & Co Inc

Objectives: Despite several inherent limitations of liver biopsy including the invasiveness of the liver biopsy and sampling variability, the liver biopsy remains the gold standard for assessing the stage and severity of metabolic dysfunction associated steatohepatitis (MASH). There is a growing interest in using machine learning to predict changes in liver histology as machine learning algorithms can efficiently explore nonlinear correlations between biomarkers, patient characteristics, and histological changes such as fibrosis score [1]. In this study, we explored the ability of machine learning algorithms to distinguish between patients with fibrosis scores of 2 vs. 3 (F2 vs. F3) using clinical laboratory measurements, MASH biomarkers, and baseline patient characteristics.

Methods: We examined multiple classification techniques, including the Support Vector Classifier, Random Forest Classifier, and Gradient Boosting Classifier. Our dataset consisted of 273 data points selected from a population of patients with histology-proven MASH with F2 or F3. The input features for all algorithms comprised of baseline demographics, clinical laboratory measurements, and a selection of MASH-related biomarkers. To ensure an unbiased evaluation of the model, the data set was split 4:1 into a training set and a testing set. A grid searching strategy was used to identify the best combination of hyperparameters in each algorithm using 5-fold cross validations over 5 different iterations. The accuracy of the best model from each algorithm was compared using the testing data set to select the final model.

Results: Our results show machine learning techniques can use patient characteristics and non-invasive laboratory and biomarker data to distinguish between liver F2 vs. F3 with great accuracy (75% to 80%). Among all classification techniques tested, Support Vector Classifier showed the best performance with an area under the curve of receiver operating characteristic (AUCROC) of 0.83. In contrast, the most informative input feature (taken from feature ranking provided by random forest algorithm) as a predictor of fibrosis stages showed an AUCROC value of 0.64.

Conclusions: These findings suggest that machine learning techniques are better predictors of F2 vs. F3 stage of liver fibrosis than a single biomarker. The study illustrates the potential of machine learning techniques as a tool in predicting fibrosis scores in MASH patients. Additional work is needed to further validate these conclusions and explore the utility of machine learning techniques in predicting other stages of liver fibrosis. Ultimately, a model with these capabilities has the potential to reduce the high screen failure rates in MASH trials based on failure to meet criteria for liver histology.

-The authors acknowledge support from the MRL Postdoctoral Research Program

Citations: [1] Lee, J., Westphal, M., Vali, Y., Boursier, J., Petta, S., Ostroff, R., ... & Bossuyt, P. M. (2023). Machine learning algorithm improves the detection of NASH (NAS-based) and at-risk NASH: A development and validation study. Hepatology, 78(1), 258-271.