Modelling a Good Delivery of Bad News

The development of soft skills and specifically good delivery of bad news gain increasing importance in all healthcare disciplines. These skills improve the communication between healthcare professionals, patients and their families. Good delivery of bad news is defined and taught using qualitative and subjective means. Quantitative voice and language attributes could provide an automated practice and education tool for healthcare professionals and improve their delivery of bad news. We investigated acoustic and verbal features in a database recorded by healthcare professional simulating delivery of bad news. The recordings were rated by other healthcare professionals and labelled as “good” or “bad”. Prosodic features were extracted directly from the recordings and provided speech tone attributes. Automated speech recognition was applied to compute the speech pace feature. A bidirectional long short term memory network was trained on the features and labels. The classification model trained on the tone features yielded an accuracy of 81.8%. The model trained on the combined tone and pace features yielded an accuracy of 90.0%. This proof of concept implies a feasibility for a fully automated practice tool that could quantify good delivery attributes and train and improve the skills of healthcare professionals in their delivery of bad news.