Machine Learning-Guided Cancer Screening: The Benefits of Proactive Care

Abstract

Problem definition: With the advance of data analytics, many disease prediction models have been developed with the intent of detecting diseases earlier and improving patient outcomes through earlier treatment. The operationalization of interventions and care based on these predictive models is critical to attaining these goals. We study the real-world effects of a machine learning-guided colorectal cancer screening program deployed at a health system in Pennsylvania. Methodology/results: Using a regression discontinuity design based on the predicted risk score for having cancer, we find that the program increases the likelihood of colonoscopy uptake in three and six months by 6.0 percentage points (214% increase relative to the control sample within the bandwidth) and 6.9 percentage points (117% increase), respectively. Importantly, we also find significant effects on mortality. We estimate that the program decreases 2-year mortality by 6.2 percentage points (43% decrease). Managerial implications: Our finding suggests that a machine learning-guided proactive cancer screening program could significantly improve patient outcomes in addition to achieving higher disease detection rates. We argue that establishing unbiased estimates of the impact of machine learning-guided screenings is critical for capacity planning of screening resources, such as colonoscopies.

Keywords: machine learning, colorectal cancer, screening, mortality, regression discontinuity