When AI Is Not Enough: Reducing Diagnostic Errors with Radiologist Oversight

What the paper says

Artificial intelligence (AI) is becoming increasingly prevalent, particularly in healthcare, in which it is shaping the future of decision-making processes. In radiology, AI has revolutionized diagnostics by enabling rapid analysis of patient imaging. However, the consequences of AI misdiagnoses can be significant. For example, an incorrect result can unnecessarily flag a healthy patient for treatment, whereas a missed detection may fail to identify a serious condition that requires immediate intervention. To mitigate such risks, most diagnostic systems combine AI analysis with radiologist review: AI first classifies cases, and then, radiologists review and confirm or modify the initial diagnosis of AI. Effective radiology scheduling must account for the likelihood and cost of false negatives and false positives as well as AI characteristics, such as sensitivity and specificity. To address the limitations of AI predictions, we develop a multiserver queuing model with separate queues for suspected positive and suspected negative cases. Using a fluid approximation, we derive an index-based policy, a modified version of the [Formula: see text] rule, to optimally schedule and allocate resources, taking into account AI characteristics and potential misclassifications. Our proposed policy naturally incorporates the anchoring effect, causing radiologists to devote more time to misclassified cases. As the anchoring effect is incorporated into the classes’ indexes, it may change the classes’ prioritization and significantly influence overall system performance. Furthermore, to prevent excessive waiting times, even for patients diagnosed as negative, we extend our model to incorporate diagnosis-based service-level requirements established by hospitals and regulators. Numerical results demonstrate the effectiveness and superiority of our policy compared with a widely used benchmark, underscoring its potential to improve diagnostic accuracy and efficiency. History: This paper has been accepted for the Service Science Special Issue on the Impact of AI on Service Design and Delivery. Funding: Partial financial support was received from Israel Science Foundation (ISF) [Grant 277/21] and The Bernard M. Gordon Center for Systems Engineering at the Technion. Supplemental Material: The online appendix is available at https://doi.org/10.1287/serv.2024.0234 .