Hybrid LSTM–Seq2Seq Models: Improved Patient Interaction for Healthcare Chatbots

Abstract

Healthcare chatbots play a critical role in improving communication between patients and healthcare providers by offering accurate and timely responses. A novel approach is proposed, which leverages a deep learning model that combines long short-term memory (LSTM) neural networks and a sequence-to-sequence (Seq2Seq) architecture to enhance text prediction accuracy in medical dialogue systems. The model leverages the capability of LSTM to capture long dependencies in sequential data alongside the contextual encoding of Seq2Seq, which improves predictive quality in dialogue responses. The encoder–decoder architecture, which utilizes tokenization and padding to standardize input sequences, contributes to the improvement in data processing. The validation accuracy of the model is 0.9766, with a loss of 0.0184. Specifically, the precision is 0.9961, the recall is 0.9981, and the F1 score is 0.9971. The capability of the model for sequence prediction is attributed to its robustness. Other methods of evaluation employing measures such as the Nash–Sutcliffe efficiency coefficient, correlation coefficient, and normalized root mean square error demonstrate that the model is superior to other machine learning algorithms utilizing linear regression and GP regression. Employing callback functions during training ensures the best-fit model is saved, which makes the method viable in different tasks described in the job descriptions.