Learning to Improve Clinical Decisions and AI Safety by Leveraging Structure

Abstract

The availability of large collections of digitized healthcare data along with the increasing power of computation has allowed machine learning (ML) for healthcare to become one of the key applied research domains in ML. ML for health has great potential in providing clinical decision-making support that can improve quality of care and reduce healthcare spending by easing clinical operations. However, the successful development of ML models in healthcare is contingent on data that is complex, noisy, heterogeneous, limited in labels and highly sensitive. In this thesis, we leverage the unique structure present in medical data along with the availability of external knowledge to guide model predictions. Additionally, we develop differentially private (DP) training techniques using gradient structure to mitigate privacy leakage.

In this thesis, we develop methods on different medical modalities such as multivariate physiological signals of ICU patients, patient discharge summaries, biomedical scientific articles, radiology reports, chest radiography imaging and spoken utterances. We tackle tasks such as forecasting patient states, relationship extraction, disease prediction, medical report generation and differentially private model training. We begin the thesis by offering open source data processing and modeling frameworks, move towards improved interpretability of model predictions to develop clinician trust and finally investigate differentially private ML techniques to protect user data.

First, we show that the use of aggregated feature representations based on clinical knowledge offers model robustness against evolving hospital systems. Second, we leverage external knowledge in the form of clinical concept extraction to significantly improve relationship extraction. Third, we leverage the rich information from reports associated with chest radiographs to develop highly accurate disease severity prediction models using contrastive learning. Fourth, we showcase that the report generation task offers competitive disease prediction capabilities, label efficiency and improved interpretability. Finally, we introduce novel methods for improved privacy-utility-compute tradeoffs for DP pre-training of large speech models. We highlight DP as an important component of model safety, necessitating its development in conjunction with AI safety approaches that will be pertinent in healthcare and beyond.