PhD defence Shreyasi Pathak | Machine Learning in Healthcare: From Petri-dish AI to Reality-centric AI

 Machine Learning in Healthcare: From Petri-dish AI to Reality-centric AI

Shreyasi Pathak is a PhD student in the department Datamanagement & Biometrics. (Co)Promotors are prof.dr.ir. M. van Keulen from the faculty of Electrical Engineering, Mathematics and Computer Science, University of Twente and prof.dr. C. Seifert from the Marburg University Germany.

Deep learning has seen wide applications in clinical decision support systems. However, to be adopted in clinical practice, deep learning-based models have to be usable in the clinical workflow. Currently, most models developed for healthcare applications stay confined to the research phase and cannot be completely integrated in the clinical workflow. One of the reasons behind this is technological barriers. Most existing models are not developed while keeping the real-world settings in mind. These models are referred to as petri-dish AI, which are unusable as-is when put into practice.

This PhD thesis investigates multiple aspects behind development of reality-centric AI models for healthcare, i.e. models that are usable as-is in clinical practice. The development of reality-centric AI models are based on three pillars addressed in the three parts of this thesis : (I) models are built on real-world data, (II) model development is based on real-world settings, and (III) model output is useful in the real-world. These pillars are investigated using three clinical use cases: (a) breast cancer diagnosis using mammography images, (b) sleep stage prediction using electrophysiological signals, and (c) 30-day post-operative mortality prediction of elderly hip fracture patients using pre-operative and per-operative data.

Part I focuses on data curation and data sharing of real hospital data. It discusses various challenges and its solutions associated with data curation from hospital. It also proposes a model-to-data platform for sharing hospital data in a privacy-preserving manner for training models and provides access to a large mammography dataset through the platform.

Part II focuses on three model characteristics to develop reality-centric models, specifically: weak labels, multimodality and interpretability. Models with these characteristics achieve good performance, suggesting that model development should be done with real hospital settings in mind.

Part III focuses on explaining the model output through posthoc and antehoc interpretability techniques. It explores intrinsically interpretable (antehoc) prototype-based methods and proposes six metrics for evaluating the quality of prototype-based explanations based on domain knowledge.

A reality-centric approach is much needed for a continuous and effective progress of AI in healthcare, as models developed within this approach are capable of being used in clinical practice.