Computer Science ETDs

Publication Date

Spring 2-10-2023

Abstract

Positive and Unlabeled (PU) learning problems abound in many real-world applications. In healthcare informatics, diagnosed patients are considered labeled positive for a specific disease, but being undiagnosed does not mean they can be labeled negative. PU learning can improve classification performance, and estimate the positive fraction, α, among unlabeled samples. However, algorithms based on the Selected Completely At Random (SCAR) assumption are inadequate when the SCAR assumption fails (e.g., severe cases overrepresented), and when class imbalance is substantial. This dissertation presents and evaluates new algorithms to overcome these limitations. The proposed methods outperform the state-of-art for α-estimation, enhance classification performance, and provide well-calibrated classification on synthetic and benchmark datasets to support good decision thresholds. Furthermore, as verified through chart review, the proposed methods can detect uncoded self-harm events in electronic health records, and accurately estimate their prevalence, with demonstrated pharmacovigilance applications in mental health informatics.

Language

English

Keywords

positive and unlabeled learning, PU learning, noisy labels learning, machine learning, healthcare informatics, SCAR, SNAR, PULSNAR

Document Type

Dissertation

Degree Name

Computer Science

Level of Degree

Doctoral

Department Name

Department of Computer Science

First Committee Member (Chair)

Christophe G. Lambert

Second Committee Member

Abdullah Mueen

Third Committee Member

Trilce Estrada

Fourth Committee Member

Tudor I. Oprea

Project Sponsors

Patient‐Centered Outcomes Research Institute, NIH National Institute of Mental Health

Download

Share

COinS