Explainable models of disease progression in ALS: Learning from longitudinal clinical data with recurrent neural networks and deep model explanation

Abstract Background and Objectives Deep neural networks recently become a popular tool in medical research to predict disease progression and unveil its underlying temporal phenotypes. While being well suited to study longitudinal clinical data and learn disease progression models, its application in clinical setting is challenged by the importance of model explainability. In this work, we integrate model learning using recurrent neural networks with deep model explanation using SHAP to study disease progression in Amyotrophic Lateral Sclerosis (ALS). Methods We propose and evaluate different deep neural networks to predict disease progression in a large cohort of Portuguese ALS patients. We learn models of disease progression targeting the prediction and explanation of respiratory decline to respiratory failure, as measured by clinical administration of non-invasive ventilation (NIV). Afterward, we explain the learnt models using SHAP and inspect the outcome with clinical researchers, targeting the identification of highly influencing features in model prediction, and putative features to be discarded in an augmented model, due to their small influence. Results When used to predict breathing capability of ALS patients in different time windows, our recurrent neural networks with LSTMs (Long Short-Term Memory) showed mean squared errors below 0.01 and 0.02 in train and test, respectively. This enables their effective use to predict the need for non-invasive ventilation, and potentially other clinically relevant endpoints, while providing clinical insights regarding disease progression to respiratory insufficiency. It was exciting to find that this study supports previous results showing that neck weakness is related to disease outcome and respiratory decline in ALS. Conclusions Our study to learn and explain a predictive model for ALS shows the potentialities of using deep learning from longitudinal clinical data together with deep model explanation to achieve accurate prognostic prediction and model interpretability, while drawing insights into disease progression and promoting personalized medicine.