Will machines take over? Algorithms for human–machine collaborative decision making in healthcare

Abstract

Despite significant advancements in predictive artificial intelligence (AI), organizations continue to grapple with determining the most effective integration of AI into their operations, particularly in combining AI with human capabilities in task execution. This paper introduces a diagnostic system aimed at minimizing costs while maintaining patient safety in hospitals by efficiently allocating mammography interpretation tasks between AI algorithms and radiologists. The optimal diagnostic system employs algorithm-generated risk scores for mammograms to determine if additional assessment by a radiologist is necessary. It evaluates the costs of two approaches—automation, where AI completely replaces human interpretation, and delegation, where AI and radiologists share interpretation tasks—compared to the current expert-alone strategy. When AI performance does not exceed that of radiologists, the optimal design is a simple two-threshold policy: AI recommends no follow-up for low-risk cases, recommends follow-up for high-risk cases, and delegates ambiguous cases in between to radiologists; the thresholds are analytically derived and state-independent. This two-threshold policy is state-independent and that optimal thresholds are not contingent on the diagnostic system’s prior history. We demonstrate our system’s performance by back-testing against real-life scenarios, utilizing data from a mammography AI contest and real-world cost and performance metrics. Backtesting demonstrates potential cost savings of up to 20.9% compared to the expert-alone approach. Beyond radiology imaging, our work holds significant implications for the design of workflows in the AI era and human–machine collaboration contexts.