Robust Intent Classification using Bayesian LSTM for Clinical Conversational Agents (CAs)

Conversational Agents (CAs) are software programs that replicate human conversations using machine learning (ML) and natural language processing (NLP). CAs are currently being utilised for diverse clinical applications such as symptom checking, health monitoring, medical triage and diagnosis. Intent classification (IC) is an essential task of understanding user utterance in CAs which makes use of modern deep learning (DL) methods. Because of the inherent model uncertainty associated with those methods, accuracy alone cannot be relied upon in clinical applications where certain errors may compromise patient safety. In this work, we employ Bayesian Long Short-Term Memory Networks (LSTMs) to calculate model uncertainty for IC, with a specific emphasis on symptom checker CAs. This method provides a certainty measure with IC prediction that can be utilised in assuring safe response from CAs. We evaluated our method on in-distribution (ID) and out-of-distribution (OOD) data and found mean uncer-tainty to be much higher for OOD data. These findings suggest that our method is robust to OOD utterances and can detect non-understanding errors in CAs.