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 behavior-based clinical exams with an average per-subject error of 6.96 out of 72 points (<10 % difference), using 30-second videos in leave-one-subject-out-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 ( $\sim $ 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.