Exploring differential privacy in CNNs, LSTMs, GRUs, and RNNs for heartbeat detection from multimodal data

Abstract Machine learning techniques in healthcare applications have seen rapid growth. However, the use of sensitive health data raises significant privacy concerns. In this paper, we present the application of various deep learning models (CNN, LSTM, GRU, and RNN) to classify heartbeat abnormalities while preserving privacy through differential privacy. We achieve this by adding Gaussian noise to the gradients during stochastic gradient descent training, ensuring that individual patient data cannot be identified or traced from the model’s results. We trained and evaluated these models on the multimodal MIT-BIH polysomnographic dataset. The data was preprocessed using noise reduction filters, heartbeat segmentation through frequency-based sampling, and resampling. Our results show that, even with differential privacy constraints, the GRU model achieved the highest accuracy of 99.5%, followed by CNN (99.12%), LSTM (98.89%), and RNN (79.60%). These findings provide practical guidance for selecting effective and privacy-preserving deep learning models for heartbeat abnormality detection in real-world healthcare scenarios.