PhD Thesis Proposal: Spencer Bertsch

"Decision-Aware Prediction and Optimization: Applications in Healthcare and Wildfire Management"

Abstract

Although modern machine learning algorithms are capable of generating data-driven predictions, care is needed when translating predictions into meaningful decisions. Real-world operational constraints, asymmetric error costs, and prediction uncertainty make it critical to align predictive models with downstream decisions. This thesis develops a progression from predictive modeling to optimization in healthcare and wildfire management. 

The first chapter examines two prediction problems designed to support trauma response and cardiovascular disease screening. The trauma response study is purely predictive, leveraging time series classification to identify patients with occult hemorrhage. In contrast, the cardiovascular disease screening application incorporates healthcare operational constraints into the objective function used during hyperparameter optimization. Results show that the model reduces the number needed to screen by 54% while retaining the required sensitivity, supporting screening decisions and encouraging wider adoption of Lp(a) testing.

Modern wildfire simulators provide essential fire spread predictions that support operational decision-making. Existing studies only calibrate wildfire simulators online and do not model spatiotemporal suppression during calibration. The second chapter presents a unified stochastic wildfire simulation and calibration framework that incorporates spatiotemporal suppression into offline and online calibration, accounting for the greater real-world cost of under-prediction. The proposed framework substantially reduces under-prediction error, with offline calibration yielding double the accuracy of an uncalibrated baseline and online data assimilation resulting in an additional 35% improvement. This calibrated wildfire spread prediction model supports suppression optimization decisions.

Current wildfire incident response systems struggle to suppress active wildfires as fire conditions worsen globally. The third chapter develops a rolling horizon wildfire suppression model to generate suppression crew assignments over sequential decision epochs. Fire spread prediction uncertainty is modeled using a series of single-stage stochastic programs, supporting time-critical decisions about safe and operationally feasible crew assignments and routing. This framework provides fire managers with optimized crew assignments that adapt to real-time uncertainty in weather, fire behavior, and crew budgets.

Ultimately, these applications show how prediction, calibration, and optimization must be tightly coupled to ensure the quality and utility of operations research solutions.

Thesis Committee

• Vikrant Vaze (Chair)
• Wesley Marrero
• Geoff Parker

Contact

For more information, contact Thayer Registrar at thayer.registrar@dartmouth.edu .