UC San Diego
Human-Centered Machine Learning for Healthcare: Examples in Neurology and Pulmonology
- Author(s): Ramesh, Vishwajith
- Advisor(s): Weibel, Nadir
- Cauwenberghs, Gert
- et al.
Machine learning (ML) in healthcare has enabled the automatic detection of diseases from medical images or sensors with high accuracy, often outperforming domain experts. Unfortunately, there is a large variance between how such diagnostic aids perform in research settings and in the real-world. This is due to challenges unique to healthcare that, if unaddressed, limit the usefulness of ML-based software when deployed in hospitals. For example, in human subjects research and clinical trials, subject recruitment and data acquisition are involved processes for both patients and healthcare providers; there are several regulatory and cybersecurity requirements to satisfy to ensure that patient care is not compromised in the pursuit of big data. Without abundant data to train ML models, it can be difficult to elicit good performance that also generalizes well on unseen data in clinical practice. Moreover, ML tools in hospitals cannot function independently but must integrate with existing workflows. There are ethical considerations with respect to how these tools influence the decision making of clinicians and whether they encourage an over-reliance on predictions.
In this dissertation, we discuss these and other concerns in the context of three focus areas: stroke, respiratory disease, and Parkinson’s disease. We present machine and deep learning pipelines for weakness detection in stroke patients from video, respiratory disease classification from audio of coughs, and gait assessment in Parkinson’s disease with body sensors. In our efforts, we were cognizant of the technical and human challenges of healthcare. We developed models that not only performed well but also could be trained and rigorously evaluated in a data-conscious way. Our ML solutions ranged from simple leave-one-out approaches to data augmentation with generative adversarial nets. Lastly, we show how ML can more effectively aid medical diagnosis when paired with human-centered design. We describe a clinical decision support system for acute stroke, focusing on the development of an intuitive user interface that balances neurologist assessments with the symptom predictions of our models. This dissertation details novel, human-centered ML techniques for disease diagnosis in neurology and pulmonology, highlighting several lessons learned to benefit the field of machine learning in healthcare at large.