Early Detection of Infant Cerebral Palsy Risk based on Pose Estimation using OpenPose and Advanced Algorithms from Limited and Imbalance Dataset

Detection of the risk of cerebral palsy existance in infant phase is critical during human development. The fidgety movements of infant during this phase plays an important role in indication of normal or abnormality of balanced and coordination. Previous researches have shown the possibility of abnormality detection using infant pose estimation. However, in particular for predicting the risk of cerebral palsy (CP) based on the estimation of the infant's movement poses, it is not optimal in its classification due to the rarity of dataset sources. This research aimed to develop a classifier based on OpenPose and advanced algorithms, including a Long Short-Term Memory (LSTM) network, 1-dimensional Convolutional Neural Network (CNN) combined with LSTM, and Gated Recurrent Unit (GRU), to predict the likelihood of cerebral palsy in infants, where amount of data is limited and there is an imbalance in categories. Such dataset was obtained from Chambers et al. and divided into 'at-risk' and 'healthy' categories. This research evaluates the performance of different algorithms in classifying infants with cerebral palsy and those without. After perfecting the model, ID CNN combined with LSTM outperformed other models with an accuracy of 0.96. Meanwhile, GRU achieved an accuracy of 0.83, and LSTM achieved an accuracy of 0.77. This research also highlights the potential of using OpenPose and advanced algorithms to accurately predict and prevent cerebral palsy in infants, providing valuable insights for future research in this area.