Leveraging Decision-Focused Learning to Optimize Stroke Patients' CT Scanning Regimens

Abstract

The United States’ healthcare sector currently faces a variety of challenges, including rising costs, overcrowding, a lack of personalized medical care, and physician burnout. Incorporating optimization, machine learning, and artificial intelligence within healthcare systems (often referred to as healthcare analytics) has the potential to address many of these shortcomings. We apply machine learning and optimization to the process of sequentially planning computed tomography (CT) scans across multiple patients within a neurology intensive care unit. We model this as a rolling horizon value maximization integer program where the value of scanning each patient at each time step within the planning horizon is uncertain and must be predicted from data. We evaluate the model’s performance on the basis of its event alignment (i.e., if it schedules patient scans appropriately). Motivated to predict these scanning values such that they produce optimal scheduling decisions, we construct a neural network regressor we term RewardNet and train it using the decision-focused learning (DFL) SPO+ loss function. We found that the DFL approach consistently outperforms other regressor loss functions and a multi-class classifier baseline in both offline and online simulation-driven out-of-sample performance tests and admits an interpretable feature importance structure. Nevertheless, although the rolling horizon oracle drastically outperforms the historical clinical decision baseline, our DFL-driven model falls short of this metric. This motivates further research to develop a DFL-driven model that can surpass the clinical baseline and eventually serve as a clinical decision support tool.