PUB - Publikationen an der Universität Bielefeld

Research has shown that approximately 10% of medical diagnoses are wrong. As a direct consequence, appropriate medical treatment may be delayed or even absent, leading to an increased burden on patients, increased costs for medication, or even harm and death. As cognitive biases contribute to roughly three-quarters of these diagnostic errors, a lot can be gained from increasing a physician’s reasoning quality during the diagnostic process. Clinical decision support systems (CDSS), leveraging recent advances in artificial intelligence (AI) and insights from eXplainable AI (XAI), aim at providing accurate predictions and prognosis paired with corresponding ex-post explanations that make the reasoning of the system accessible to humans. Viewing explanations as involving the interactive construction of shared belief, we propose to move from diagnostic decision support to reasoning support which, in its true sense, needs to tailor the timing and content of generated explanations to the state of the reasoning process of physicians to meet their information needs and effectively mitigate the influence of cognitive biases. We claim that, given the uncertain and incomplete information inherent to medical diagnosis, the most effective way not to fall prey to cognitive reasoning errors is to establish and maintain proper justification for each decision throughout the diagnostic process. This paper contributes (1) a conceptual model and desiderata for AI-based interactive reasoning support that enhances reasoning quality through increased justification at every stage of the process, and (2) preliminary work on the development of the assistive, co-constructive differential diagnosis system, ASCODI, which provides reactive as well as proactive reasoning support to improve the justification of actions taken and decisions made during and after medical diagnosing. We also present selected use cases of ASCODI concerning its application in supporting the diagnosis of transient loss of consciousness and highlight their connection back to the theoretical concepts established.