Predictive and Prescriptive Analytics for Optimizing Concussion Management Decisions

Abstract

As the volume and granularity of health data continue to increase, clinical decision-makers are faced with two key questions: (Q1) How can large clinical datasets be used to gain a patient-specific understanding of disease risk and disease progression? (Q2) How can a data-driven understanding of patient-specific disease risk and disease progression be combined with multiple stakeholders' perspectives to optimize medical decision-making? These challenges are especially pertinent to managing patients with concussion. Concussion, an emergent public health issue, affects millions of people in the United States each year. Characterized by wide-ranging symptoms and impairment in neurocognitive function, researchers believe that improving concussion management can mitigate the long-term consequences associated with the injury. In this dissertation, we answer (Q1) and (Q2) by analyzing three key aspects of concussion management: acute concussion assessment, diagnosis decisions, and return-to-play (RTP) decisions. Throughout this dissertation, we develop, parameterize, and validate our models using data from the Concussion Assessment, Research, and Education (CARE) Consortium --- a large dataset on concussion among collegiate athletes from 29 universities and military service academies across the United States.

In our analysis of acute concussion assessment, we design predictive models to assess the relationship between acute concussion and clinical assessments, individual risk modifiers (e.g., age, sex, number of previous concussions), and time of injury characteristics (e.g., loss of consciousness). This research provides valuable contributions in (1) quantifying the value of a multi-dimensional approach to acute concussion assessment and (2) identifying specific components of the Sport Concussion Assessment Tool which best identify acute concussion.

To analyze concussion diagnosis decisions, we formulate and solve the Two-Threshold Problem (TTP): a data-driven stochastic programming approach to determine optimal diagnosis decision thresholds with risk estimation models. Using the personalized risk estimation models from the first part of this dissertation as an input, we apply the TTP to acute concussion diagnosis and identify its implications for clinicians. The contributions of this research include (1) the development of a novel data-driven framework for optimizing diagnosis decisions, (2) an algorithmic approach to classifying the certainty in acute concussion diagnosis decisions (i.e., Unlikely, Possible, Probable, and Definite concussion), and (3) the characterization of athletes who are most difficult to diagnose, i.e., Possible and Probable concussions.

The final part of this dissertation analyzes the timing of RTP from concussion. We first formulate and solve a novel Behavior-Learning Multi-agent POMDP (BLM-POMDP): a multi-agent, stochastic dynamic programming model which incorporates the patient's and doctor's perspectives while accounting for uncertainty in the patient's health and symptom-reporting behavior. We then apply the BLM-POMDP to CARE Consortium data to estimate the value of incorporating patient behavior in RTP decisions. The contributions of this work include (1) the formulation and characterization of a novel dynamic programming framework which naturally models patient-doctor interactions in sequential treatment planning regimes and (2) the development and analysis of an optimal RTP policy which can be tailored to each athlete and outperforms current practice.

In summary, this dissertation combines data analytics and operations research to address major challenges in concussion management. Our modeling frameworks span the range of predictive models for risk estimation to data-driven sequential decision-making under uncertainty. While this research was motivated by and applied to concussion management decisions, this research can be adapted to a broader range of application areas where data, prediction, and decisions play a crucial role.