Deep Learning for Amyotrophic Lateral Sclerosis Progression: A Multi-Task Learning Approach for Simultaneous Prediction of ALSFRS and FVC



Poster Number: M206


Hamza Turabieh, PhD, University of Missouri, Jeffrey Statland, MD, University of Kansas Medical Center, Xing Song, PhD, University of Missouri

Machine learning has recently received more attention in its ability to discover risk factors and make predictions about disease progression, especially for complex conditions such as ALS. However, all the established work only focused on prediction of a single prognostic marker (i.e., single-task learner) in isolation of other important prognostic markers. In this study, we developed a multi-task screener-learner model based on Boruta features selection and Convolutional Neural Network (Boruta-MT-CNN) to simultaneously predict rates of ALSFRS and FVC declines.

We utilized the public PRO-ACT dataset (August 2022 release) with 13 tables from 23 trials, covering 11,600 ALS patients and up to 881 clinical and biological markers, including ALSFRS questionnaire details, laboratory data, demographics, and more. Among them, we selected a cohort of 531 patients with complete set of all 321 features. We combined Boruta feature selection algorithm (“screener”) and CNN-based multi-task learning model (“learner”) to predict ALSFRS and FVC declining rates from month 3 to month 12 since disease onset. The model featured two shared layers (128 and 64 neurons) and two task-specific layers (8 and 1 neuron) for each task.

The median (IQR) age of the selected cohort is 56 [48, 63], with 63.65% males and 4.3% non-white. The Boruta-MT-CNN model demonstrated comparable performance achieving root-mean-square-error (RMSE) values of 0.588 for ALSFRS prediction and 0.0285 for FVC prediction, in comparison with single-task models (0.570 for ALSFRS-only model, 0.0280 for FVC-only model).

The Boruta-MT-CNN showed significant promise in creating a more effective, resilient and wholistic prognostic model for ALS patients. Future work is to further validate this model on external datasets as well as extract intra- and inter-task risk factors to better support clinical decision making.