Evaluation of infants with spinal muscular atrophy type-I using convolutional neural networks

Abstract

Spinal Muscular Atrophy is the most common genetic cause of infant death. Due to its severity, there is a need for methods for automated estimation of disease progression. In this paper we propose a Convolutional-Neural-Network (CNN) model to estimate disease progression during infants’ natural behavior. With the proposed methodology, we were able to predict each child’s score on current behaviorbased clinical exams with an average per-subject error of 6.96 out of 72 points (<10% difference), using 30-second videos in leave-one-subjectout-cross-validation setting. When simple statistics were used over 30-second video-segments to estimate a score for longer videos, we obtained an average error of 5.95 (~8% error rate). By showing promising results on a small dataset (N = 70, 2-minute samples, which were handled as 1487, 30-second video segments), our methodology demonstrates that it is possible to benefit from CNNs on small datasets by proper design and data handling choices.