Predicting ALS progression using Autoregressive deep learning models

Forecasting the progression of Amyotrophic Lateral Sclerosis (ALS) presents challenges due to the patients exhibiting different onset sites, progression rates, and survival times. Several recent studies have successfully applied machine learning to predict patient functional decline over time. However, the use of advanced techniques such as deep learning and temporal data modeling has yet to be explored in the field of ALS prognosis. In this study, we proposed a novel approach based on Autoregressive Multi-Step Multi-Output Time Series Forecasting to predict functional disability for the next 12 months, month-by-month, using patient data collected from the first three months. This study is the first to employ this approach to ALS prognosis to predict the functional decline over time. We extracted static and temporal features from the Pooled Resource Open-Access ALS Clinical Trials database. We developed and evaluated deep learning models using the Gated Recurrent Unit and Long Short-Term Memory algorithms. Our approach outperformed previous works with a significantly smaller set of input features, thus demonstrating greater effectiveness. With the promising results obtained, our approach could aid physicians in devising personalized treatment and resource planning or serve as an inclusion/exclusion criterion in clinical trials.