On data, modeling, and clinically aligned predictive models in healthcare
Abstract/Contents
- Abstract
- The application of data-driven models to healthcare settings has grown exponentially. The general pathway for creating these models involves leveraging data to construct a model that helps healthcare stakeholders in making "data-driven" decisions. To build trusted models for high-stakes decision-making problems using this pathway, it is crucial to proactively incorporate healthcare stakeholders. This dissertation studies the different elements of this pathway, particularly the importance of alignment with healthcare stakeholders. In the first two parts, this dissertation identifies gaps between modern machine learning (ML) pipelines and healthcare stakeholders, focusing on both the data and model phase, and studies the proactive use of clinical knowledge to create more robust, clinically usable models. For the data phase, we use clinical expertise to generate features from unstructured physician progress notes and analyze their impact on clinical risk prediction models in a prostate cancer case study. The results provide a proof of concept for the value of clinical expertise-guided feature generation. For the model phase, we provide a reproducible framework to examine the misalignment, or inconsistencies, of modern ML model behavior with clinical experiential learning, focusing on the impact of underspecification of ML pipelines. In a prostate cancer outcome prediction case study, we identify and address these inconsistencies via a constraint-guided alignment approach and additionally test the feasibility of a key requirement for a feedback-driven alignment approach. The constraint-guided alignment approach shows that a model can be better aligned with clinical experiential learning without compromising performance. This practical framework is complemented by a mathematical analysis, rooted in recent theory of overparameterized neural networks, to illustrate how model complexity, combined with a large training dataset, can lead to inconsistencies. In the final part, this dissertation transitions to illustrating the capabilities of a non-ML model to inform healthcare decision making by analyzing the effectiveness of face masks during the COVID-19 epidemic using a dynamic disease model. The findings illustrate the model's ability to provide timely insights for policymaking, effectively capturing how characteristics of face masks influence the epidemic's course during both the initial outbreak and subsequent resurgence. Overall, this dissertation, through interdisciplinary collaborations and new methodologies, advances our understanding of how to design data-driven models more effectively for greater trust and adoption in healthcare practices, and how such models can be used to inform healthcare decision making promptly and successfully.
Description
| Type of resource | text |
|---|---|
| Form | electronic resource; remote; computer; online resource |
| Extent | 1 online resource. |
| Place | California |
| Place | [Stanford, California] |
| Publisher | [Stanford University] |
| Copyright date | 2024; ©2024 |
| Publication date | 2024; 2024 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Vallon, Jacqueline Jil |
|---|---|
| Degree supervisor | Bayati, Mohsen |
| Thesis advisor | Bayati, Mohsen |
| Thesis advisor | Brandeau, Margaret L |
| Thesis advisor | Johari, Ramesh, 1976- |
| Degree committee member | Brandeau, Margaret L |
| Degree committee member | Johari, Ramesh, 1976- |
| Associated with | Stanford University, School of Engineering |
| Associated with | Stanford University, Department of Management Science and Engineering |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Jacqueline Jil Vallon. |
|---|---|
| Note | Submitted to the Department of Management Science and Engineering. |
| Thesis | Thesis Ph.D. Stanford University 2024. |
| Location | https://purl.stanford.edu/kr063hf4408 |
Access conditions
- Copyright
- © 2024 by Jacqueline Jil Vallon
Also listed in
Loading usage metrics...