CapsLSTM-Based Human Activity Recognition for Smart Healthcare With Scarce Labeled Data

Abstract

Scarcity of labeled data in sensitive research areas such as healthcare limits the performance of artificially intelligent (AI) models. The effort required in labeling the acquired healthcare data for carrying out a classification task is a concern faced by healthcare researchers. In this work, we design a capsule-long short-term memory (LSTM) model, abbreviated as CapsLSTM, capable of classifying human activities with scarce labeled activity data. The proposed CapsLSTM model is used to recognize multiple human activities sensed by the accelerometer and the gyroscope sensors embedded in smartphones leveraging the spatio-temporal information. We validate the proposed framework based on two human activity recognition (HAR) databases, namely, UCI-HAR and MotionSense. The CapsLSTM model yields close test accuracies for different fractions of the training data unlike the other models such as LSTM, 1-D convolutional neural network (1D-CNN), convolutional LSTM (ConvLSTM), or CNN-LSTM. For the state-of-the-art models, the test accuracies decrease significantly with the decrease in training data. Our proposed model classifies the human activities using a minimum of 20% labeled training data from each database. There is less decrease in accuracies from that obtained using 70% of training data in comparison to the existing models. The classification performance of the proposed CapsLSTM model for a small fraction of training data proves the effectiveness and reliability of the same in a data-scarce situation.