Intelligent and customizable conversational systems for clinical communication training

Abstract

The aging of the population and the outbreak of pandemics place the challenges on global healthcare. With advances in artificial intelligence and big data, intelligent healthcare improves the efficiency of the medical system and reduces the workload of the practitioners greatly. In this context, we aim to facilitate autonomous, low-cost, and customizable clinical communication training by developing intelligent techniques. Effective clinical communication is essential for delivering safe and high-quality patient care, especially under the scenarios that the healthcare system faces high pressure. Training on standardized clinical communication helps to organize the conversation in a structured and focused way that ensures clinical staff get timely and appropriate responses without missing any important information. Traditional classroom teaching on clinical communication requires substantial human and medical resources, and more importantly, lacks enough high-fidelity practice. Therefore, the primary function of this intelligent training system is to simulate the dialogue among the clinicians. We develop a task-oriented multiturn chatbot, which can play various roles to practice conversations with clinical staff. The key research problem addressed here is the detection of sentence-level intents referring to the context of clinical communication standards. Compared to the existing works on intent detection, the sample dialogues for standardized clinical handover are insufficient. Moreover, these dialogues are inherently sequential and their intents are interrelated. Given this feature, we propose Intent-aware Long-Short Term Memory (IA-LSTM) to incorporate context information into intent detection. In the experiments, IA-LSTM outperforms all baseline methods of intent detection on clinical handovers. Moreover, the proposed intent-aware mechanism can be expanded to other deep learning models, thereby improving their performances. The second piece of work we have delivered is a timely assessment model, which can automatically evaluate the performance of individual clinical staff in a conversation. The research problem addressed here is the accurate recognition of the conversation content by integrating the information from both the domain knowledge and the learning examples. Based on the biomedical ontology for general purpose, we construct a specific knowledge graph for the clinical communications. A novel method called Knowledge-infused Prompt Tuning is proposed to infuse the external knowledge into prompts. The empirical validation in real world application shows that the proposed method not only achieves superior performance, but also proves more robust with limited data or complex components. The third function we have developed for intelligent clinical communication training is a friendly platform that enables users to define new training tasks by themselves. The key research problem addressed here is the customizable conversational system with insufficient training data. We propose a novel data augmentation methods for user-defined scenarios, such as the clinical handover under the COVID-19. Based on the pre-trained conversational system with user-defined knowledge, the proposed Data Augmentation with User-Defined Knowledge (UDK-DA) significantly boosts the performance of the clinical training system with only a few samples. By integrating the aforementioned modules, we develop Heallo, an intelligent, autonomous, and customizable conversational system for clinical communication training. Now Heallo has been incorporated into junior staff training programs at local hospitals and largely benefits the promotion of intelligent healthcare.