Stratification and Survival Prediction for Amyotrophic Lateral Sclerosis Patients

Survival analysis is widely used in modelling the relationship between clinical features and survival outcomes, which can help designing ALS trials and developing better treatment for ALS patients. However, the power of conventional survival analysis may be limited when the subgroups with different survival distributions exist in the population. Integrating clustering information into survival analysis is a possible solution. Besides, previous studies on clustering ALS data seldom took both clinical features and survival information into consideration. Pooled Resource Open-Access ALS Clinical Trials (PRO-ACT) Database is the largest ALS dataset containing over than 10700 patient records. In this paper, we applied a deep probabilistic method that jointly accounts for clustering and survival objectives to this dataset. Utilizing clinical features and survival time, the model is able to infer the hidden clusters and conduct cluster-specific survival analysis. We demonstrated that the model identifies meaningful subgroups and yields promising performance over baseline methods in survival time prediction. Using a SHAP explainer, we identified clinically meaningful predictors that are beneficial to develop precision medicine. Due to a large proportion of missing data in the database, we showed that different choices of imputation methods are crucial for analysing ALS data and found that probabilistic principal component analysis (PPCA) gives the best performance.