Predicting Adverse Outcomes in At-Risk Populations with Machine Learning Methods

• Author / Creator

  Sharma, Vishal

• Predicting Adverse Outcomes in At-Risk Populations with Machine Learning Methods
  by
  Vishal Sharma

  A thesis submitted in partial fulfillment of the requirements for the degree of
  Doctor of Philosophy
  in
  Public Health - Epidemiology
  School of Public Health
  University of Alberta

  © Vishal Sharma, 2022

  Abstract
  Canada has a publicly funded health system and heterogeneous population. There are segments of this population which account for substantial health care utilization and adverse outcomes. Machine learning (ML) approaches can assist public health intervention programs to mitigate health system costs and improve patient outcomes, particularly for segments within Canadian society that qualify as at-risk. The specific ones to be studied in this PhD are prescription opioid users, older adults taking benzodiazepines, and people with heart failure (HF). Identifying high risk individuals within these segments using ML methods trained on administrative health data as well as assessing prediction performance and value to inform health system planners are the objectives of this PhD program. This PhD studied outcomes related to admissions and deaths and presented findings on potential cost savings of ML assisted programs.
  The main findings in this thesis were:

  1. Machine-learning classifiers, especially incorporating hospitalization and physician claims data, have better predictive performance compared to guideline or prescription history only approaches when predicting 30-day risk of adverse outcomes pursuant to an opioid dispensation. Prescription monitoring programs and health departments with access to administrative data can use ML classifiers to effectively identify those at higher risk compared to current guideline-based approaches,
  2. Despite predicting readmissions in patients with HF better than the LaCE, even the best ML model trained on administrative health data (XGBoost) did not provide substantially informative prediction performance as it only generated a moderate shift from pre to post-test probability. Health systems wishing to deploy such a tool should consider training ML models with additional data. Adding other techniques like Natural Language Processing, along with ML, to use other clinical information (like chart notes) might improve prediction performance,
  3. Developing ML models using only administrative health data may not provide health regulators with sufficient informative predictions to use as decision aids for potential interventions, especially if considering daily or quarterly classifications of benzodiazepine risks in older adults. ML models may be informative for this context if yearly classifications are preferred. Health regulators should have access to other types of data to improve ML prediction, and
  4. Prescription drug monitoring programs can use ML classifiers to identify patients at risk of adverse outcomes from opioids and potentially reduce health-care costs by intervening on high-ranked predictions. Better access to available administrative and clinical data could improve the prediction performance of ML classifiers, especially if probability thresholds are important, and thus expand opioid stewardship efforts and further reduce costs. In conclusion, the findings suggest that ML methods may demonstrate value in opioid stewardship programs with limited benefits in predicting adverse outcomes in older adults taking benzodiazepines and readmissions in people with HF. Health systems wishing to integrate ML into their program planning may benefit from additional sources of data to train ML models. Data governance, bias and ML transparency are key issues requiring future research.

• Subjects / Keywords

  • public health
  • machine learning prediction

• Graduation date

  Spring 2023

• Type of Item

  Thesis

• Degree

  Doctor of Philosophy

• DOI

  https://doi.org/10.7939/r3-sszq-0f29

• License

  This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.