Machine learning and computational statistics for risk forecasting of breast cancer screening noncompliance: A case study

Details

Supervisors

Vandenberghe, Vincent ; Federico, Andreis

Faculty

Faculté des sciences économiques, sociales, politiques et de communication

Degree label

Master [120] en sciences économiques, orientation générale

Abstract

The cancer care field has historically been focused on improving post-detection treatment and outcomes. But the real defining factors for modern breast cancer policymaking are the timeliness, reliability and population compliance of preventive screening, more than the treatment itself. In this work we present a range of statistical models that can be used to forecast the risk of preventive screening noncompliance for each patient in a novel dataset with more than 300.000 unique patients and 1.800.000 unique screening events. As a key research decision, we hold the model accountable for its decisions by making it able to provide complete and statistically grounded explanations of its individual forecasts, in order for the policymaker or doctor to be able to step in and assess the model's choices without any formal technical training. Finally, we compare different models and select a number of sociodemographic variables with strong predictive power, paying special attention to the policymaker's cost vs benefit point of view and the model's effectiveness in a simulated scenario. The resulting ensemble model retains state-of-the-art performance, while being able to provide a full breakdown of its choices. Assessed performance on simulated data suggests our work holds significant promises for real-world applications.