Towards Automatic Annotation of Clinical Decision-Making Style

Clinical decision-making has high-stakes outcomes for both physicians and patients, yet little research has attempted to model and automatically annotate such decision-making. The dual process model (Evans, 2008) posits two types of decision-making, which may be ordered on a continuum from intuitive to analytical (Hammond, 1981). Training clinicians to recognize decision-making style and select the most appropriate mode of reasoning for a particular context may help reduce diagnostic error (Norman, 2009). This study makes preliminary steps towards detection of decision style, based on an annotated dataset of image-based clinical reasoning in which speech data were collected from physicians as they inspected images of dermatological cases and moved towards diagnosis (Hochberg et al., 2014). A classifier was developed based on lexical, speech, disfluency, physician demographic, cognitive, and diagnostic difficulty features. Using random forests for binary classification of intuitive vs. analytical decision style in physicians’ diagnostic descriptions, the model improved on the baseline by over 30%. The introduced computational model provides construct validity for decision styles, as well as insights into the linguistic expression of decision-making. Eventually, such modeling may be incorporated into instructional systems that teach clinicians to become more effective decision makers.