Data-Efficient Machine Learning on Administrative Healthcare Records

Abstract

The Australian Government Department of Health and Aged Care has access to large-scale, complex healthcare data, including Electronic Health Records and administrative claims data. These comprehensive datasets hold significant potential for accurately predicting patient journeys through the healthcare system, optimising resource planning, and informing evidence-based policymaking.

Despite challenges related to complexity, sparsity, irregularity, and scale, healthcare data have been effectively leveraged by existing machine learning models for various tasks. However, deep learning models, in particular, often require extensive labelled training data and substantial computational resources, which can be impractical or prohibitively expensive in real-world government applications.

To address these challenges, this doctoral research presents a suite of data-efficient machine learning strategies specifically tailored to diverse scenarios within healthcare government settings, encompassing unsupervised paradigms, knowledge-sharing techniques, and efficient learning frameworks.

Specifically, the thesis first introduces a visualisation-supported interactive deep metric learning approach to enhance labelling efficiency in cohort discovery for population healthcare policymaking. It then presents a machine teaching framework to improve students' labelling efficiency by enabling teachers to recommend high-impact samples. Finally, a prompt learning-based solution is proposed to efficiently adapt foundation models and well-labelled source tasks to target tasks, particularly in few-shot healthcare applications.