AI-driven clinical risk prediction: Advances, bias evaluation and deployment challenges

Artificial Intelligence (AI) for Clinical Risk Prediction has developed into an effective means of predicting adverse events (AE), progression of disease, and patient outcomes through early identification across many different healthcare settings. The development of Machine Learning (ML) and Deep Learning (DL) technologies has enabled clinicians to utilize sophisticated machine learning techniques to extract complex and non-linear relationships in high-dimensional clinical data sets and to interpret that data into clinically meaningful guidelines. As such, AI-based Clinical Risk Prediction has been demonstrated to perform better than traditional statistical modelling tools in regard to Intensive Care Unit (ICU) monitoring, Hospital Readmission Prediction, Cardiovascular Risk Assessment, and Oncology prognosis. However, despite these successes, there exists a gap between current technology innovation and adoption in clinical practice. This review aims to discuss some of the current barriers that exist in the real-world adoption of AI-based Clinical Risk Prediction Technologies, while also discussing some of the tools and methodologies that are currently being utilized to address these. Furthermore, this review will outline some potential futures of AI-based Clinical Risk Prediction Technologies, including deployment pathways, performance degradation (model drift), other regulations and ethics, and how an interdisciplinary approach can catalyze the safe, fair, and clinically meaningful application of clinical risk prediction technologies at scale.