Raiki Yoshimura1, Shingo Iwami1
1Nagoya University
Title: Prediction of Disease Progression in Acute Liver Injury Patients Using Machine Learning and Mathematical Modeling Author: Raiki Yoshimura1, Shingo Iwami1 Institution: (1) Nagoya University, Japan Introduction/Objectives: Acute liver failure (ALF) is a severe condition with high mortality due to multiple organ failure[1]. With limited treatment options beyond liver transplantation and insufficient organ availability, rapidly identifying transplant candidates is crucial[2]. While previous models predict transplantation need at admission, few examine disease dynamics using time-series data. We developed a model to predict temporal disease progression in 320 acute liver injury (ALI) patients using first week clinical data from Kyushu University Hospital, combining mathematical modeling with machine learning to predict disease progression during early phase[3]. Methods: We employed Random Forest Classification to predict severity (transplantation requirement)[4], generating ROC curves via K-fold cross-validation. The SHAP algorithm identified high-contributing predictive factors[5, 6]., with SMOTE addressing data imbalance. For patient stratification, we applied Dynamic Time Warping to calculate dissimilarity between Prothrombin Time Activity Percentage (PT%) time series[7], determining cluster numbers through the elbow method. We developed a mathematical model for PT% dynamics, where parameter g represents the rate of PT% value increase from liver coagulation factor production, and parameter D represents the decrease rate due to degradation. We estimated these parameters using non-linear mixed effects modeling with stochastic approximation expectation-maximization and empirical Bayesian methods. Random Forest Regression was then used to predict these parameters from admission data. Results: Using chronological blood test data (Days 0-7), our random forest model predicted transplantation requirement with AUROC=0.85 at day 0 and AUROC=0.97 at day 7. This improved accuracy over time indicates data captured meaningful disease state changes. The SHAP algorithm identified PT% as the most contributive factor among all biomarkers tested, confirming its clinical significance in ALI progression. Time-series clustering of PT% data revealed six distinct patient groups with different clinical trajectories. Groups G3/G4 showed initially low PT% values that gradually improved over time, representing patients responsive to standard treatments. In contrast, groups G5/G6 maintained persistently low PT% values throughout hospitalization, correlating strongly with poor outcomes and transplantation requirement. These distinct trajectories were not apparent from single time-point measurements, highlighting the value of our dynamic approach. Our mathematical model successfully reconstructed individual PT% dynamics with high fidelity (R²: 0.92). The subsequent random forest regression model using only admission data could predict these mathematical parameters with sufficient accuracy to forecast individual PT% trajectories over the critical first week. Prediction accuracy varied between patient groups, with G3/G4 showing better predictability than G5/G6, suggesting potential differences in underlying pathophysiological mechanisms. Conclusion: We identified PT% as a critical predictor of ALI/ALF outcomes and developed a model to forecast individual disease trajectories from admission[3]. This approach enables early identification of patients likely to require transplantation versus those who may recover with standard care. From a drug development perspective, understanding PT% dynamics could guide the creation of therapeutics targeting coagulation factor production and lead to more efficient clinical trials through better patient stratification. Future refinements incorporating sequential data updates could further personalize treatment approaches and identify optimal intervention windows.
1. Maiwall, R., et al., Acute liver failure. Lancet, 2024. 404(10454): p. 789-802. 2. Stravitz, R.T., et al., Future directions in acute liver failure. Hepatology, 2023. 78(4): p. 1266-1289. 3. Yoshimura, R., et al., Stratifying and predicting progression to acute liver failure during the early phase of acute liver injury. PNAS Nexus, 2025. 4(2): p. pgaf004. 4. Breiman, L., Random Forests. Machine Learning, 2001. 45(1): p. 5-32. 5. Wang, R., et al., Deep learning-based identification of eyes at risk for glaucoma surgery. Sci Rep, 2024. 14(1): p. 599. 6. Huang, W., et al., Machine learning models for early prediction of potassium lowering effectiveness and adverse events in patients with hyperkalemia. Sci Rep, 2024. 14(1): p. 737. 7. Chiba, H.S.a.S., Dynamic programming algorithm optimization for spoken word recognition. IEEE Transactions on Acoustics, Speech, and Signal Processing, 1978. 86(1).
Reference: PAGE 33 (2025) Abstr 11762 [www.page-meeting.org/?abstract=11762]
Poster: Drug/Disease Modelling - Other Topics