Background
Making accurate diagnoses in teams requires complex collaborative diagnostic reasoning skills, which require extensive training. In this study, we investigated broad content-independent behavioral indicators of diagnostic accuracy and checked whether and how quickly diagnostic accuracy could be predicted from these behavioral indicators when they were displayed in a collaborative diagnostic reasoning simulation.
Methods
A total of 73 medical students and 25 physicians were asked to diagnose patient cases in a medical training simulation with the help of an agent-based radiologist. Log files were automatically coded for collaborative diagnostic activities (CDAs; i.e., evidence generation, sharing and eliciting of evidence and hypotheses, drawing conclusions). These codes were transformed into bigrams that contained information about the time spent on and transitions between CDAs. Support vector machines with linear kernels, random forests, and gradient boosting machines were trained to classify whether a diagnostician could provide the correct diagnosis on the basis of the CDAs.
Results
All algorithms performed well in predicting diagnostic accuracy in the training and testing phases. Yet, the random forest was selected as the final model because of its better performance (kappa = .40) in the testing phase. The model predicted diagnostic success with higher precision than it predicted diagnostic failure (sensitivity = .90; specificity = .46). A reliable prediction of diagnostic success was possible after about two thirds of the median time spent on the diagnostic task. Most important for the prediction of diagnostic accuracy was the time spent on certain individual activities, such as evidence generation (typical for accurate diagnoses), and collaborative activities, such as sharing and eliciting evidence (typical for inaccurate diagnoses).
Conclusions
This study advances the understanding of differences in the collaborative diagnostic reasoning processes of successful and unsuccessful diagnosticians. Taking time to generate evidence at the beginning of the diagnostic task can help build an initial adequate representation of the diagnostic case that prestructures subsequent collaborative activities and is crucial for making accurate diagnoses. This information could be used to provide adaptive process-based feedback on whether learners are on the right diagnostic track. Moreover, early instructional support in a diagnostic training task might help diagnosticians improve such individual diagnostic activities and prepare for effective collaboration. In addition, the ability to identify successful diagnosticians even before task completion might help adjust task difficulty to learners in real time.