HKS Affiliated Authors

Excerpt

2025, Paper: "Machine learning (ML) algorithms are increasingly used to enhance clinical care for a wide range of medical conditions. However, clinical decision-making often involves nuanced, contextual factors that ML models cannot easily account for, suggesting unrealized potential in combining human expertise with ML. This raises critical questions about how human expertise can complement ML models and the extent to which such collaboration could enhance performance in practice. Methodology/results: We propose a human-algorithm “centaur” model to harness synergies between human experts and ML. Through rigorous theoretical analysis, we motivate our model’s design and characterize its advantages over standalone ML and human-only approaches in predictive tasks. We introduce a framework for operationalizing centaur models in practice and validate our approach using a case study with a leading U.S. hospital focused on predicting 30-day readmissions in transplantation patients. Our findings consistently demonstrate that the centaur model can outperform both human experts and the best ML algorithm. Managerial Implications: We show that even when the accuracy of human prediction is low, creating a centaur model can yield substantial performance improvements by systematically enhancing ML algorithms with human insight. Furthermore, the predictive accuracy of the centaur model improves proportionally with the degree of divergence between the human-only and ML-only estimation policy, highlighting the value of capturing complementary information from both sources. Finally, our empirical results reveal that (1) while ML models excel in identifying non-linear and dynamic risk patterns, human experts tend to rely on linear assessments driven by primarily static, time-invariant features; (2) on average, human experts tend to over-estimate risks compared to ML models; and (3) centaurs can potentially impact clinical practices by addressing a major barrier in implementing pure ML algorithms: their lack of alignment with human experts’ clinical intuition, and hence, providing advice that can receive high weights by practitioners."