Modeling Uncertainty in Computer Vision Based Gross Motor Function Assessment of Children with Cerebral Palsy

Assessing motor impairment is essential for understanding disease progression and tailoring treatment. Traditionally, this assessment relied on manual evaluations. We are exploring computer vision, utilizing monocular video captured on personal smartphones. However, machine learning-based assessment faces significant challenges due to limited labeled data and label quality, introducing uncertainties in the model. Here, we propose modeling aleatoric uncertainty with a two-head neural network, enabling uncertainty estimation alongside Gross Motor Function Classification System (GMFCS) scores. Our training method, involving data combination and a loss function consisting of consistency loss and regression loss, contributes to improved performance. Experimental results show that the network outputs uncertainty positively correlated with GMFCS Level estimation error. Setting confidence thresholds allows for filtering out incorrect estimations and achieving higher accuracy.