Heterogeneous Clinical Data – Challenges and Opportunities
Hospitals and clinics are complex organizations, and the data they generate reflects that: patient information comes from many different sources and domains — radiology, controlling, medicine — all describing the same patients and their care, but rarely in compatible formats.
This thesis identifies four recurring challenges in working with this kind of heterogeneous clinical data: incompatible systems that require manual work to combine, multi-modal data and tasks that add complexity, large numbers of variables that can make methods infeasible or inefficient, and a general lack of interpretability that makes it hard for humans to understand the resulting models. It addresses these challenges through four contributions combining Business Intelligence and Machine Learning methods.
The first automates nursing management documentation using Business Intelligence methods, speeding up a process that is otherwise time-consuming and error-prone when systems don’t talk to each other. The remaining three contributions are new, interpretable, rule-based Machine Learning methods for tasks that are multi-modal and prone to lacking interpretability: predicting anxiety in palliative care patients from questionnaire data using Association Rule Mining metrics; estimating in-hospital mortality risk using the epidemiological odds ratio, to help providers allocate resources appropriately; and estimating a patient’s future risk of HIV infection based on prior hospital admissions, using incidence rate ratios — intended as a first risk assessment ahead of testing or pre-exposure prophylaxis counseling.
Together, these contributions show two kinds of payoff: the Business Intelligence methods give nursing managers better visibility into existing data through visualization, while the interpretable Machine Learning methods offer strong predictive performance that generalizes beyond healthcare, while still letting researchers trace the underlying correlations in the data.