A Computational Framework for Precision Medicine
The project proposes the inception of novel Artificial Intelligence frameworks to model and learn the complex temporal interactions embedded within big data repositories of health and medicine, using those to drive the next-generation of personalised health recommendation and clinical decision support systems.
The current data deluge in medicine, as a result of the digitisation of patient care, presents a massive opportunity to uncover algorithmic insight and high-quality indicators of recommended and personalised treatments. However, the highly dimensional, heterogeneous, noisy and irregular nature of available data has rendered the task of uncovering true interactions embedded within a patient’s treatment timeline a computationally challenging task. Although state of the art machine learning models, exemplified by deep neural networks, have shown tremendous success in domains such as image and speech recognition, their use in medical application is limited by the fact that they are unexplainable black-box models, making any predictions made by such models of little value to medical healthcare practitioners.
This project is built around the idea of using explicit knowledge representation of medical data as temporally constrained graphs to develop novel explainable graph-based deep learning models. Aided by the semantics power presented by knowledge graphs, such models will unearth the sophisticated temporal interactions embedded within a patient’s treatment timeline in a robust representation that is resistant to noise and irregular observations, and will enable the explanation of the outcomes generated by machine learning models, resulting accurate, sound and explainable tools. In a nutshell, the project aims to build a model for patient digital phenotype.
The model will be developed and evaluated to predict the onset of sepsis in secondary care hospitals. Sepsis, defined by a life-threatening response to infection and potentially leading to multiple organ failure, is one of the most significant causes of worldwide morbidity and mortality. Sepsis is implicated in 6 million deaths annually, with costs totalling $24 billion in the USA alone.
Early identification of sepsis is a crucial factor in improving outcomes. Yet, machine learning (ML) algorithms aiming to recognise sepsis onset from vital signs data have shown mixed results reflecting the heterogeneity of sepsis, populations and methodologies.
- Develop and evaluate a knowledge representation platform to enable modeling the multimodal, unevenly sampled, temporal and sparse patient EHRs.
- Develop and evaluate an interpretable sepsis risk predictor using machine learning models which make use of the knowledge representation framework developed.
- Evaluate the models effectiveness in making personalized prediction of sepsis risks before onset to healthcare practitioners.
Novelty and Importance:
The models to be developed are part of a new wave of “explainable AI”, which is a field of study that is fast emerging as a result of the continued unsuitability of AI models in medicine and healthcare due to their black-box nature. The project will focus on explainability and performance, assessing tradeoffs as the project matures. The project will also yield a much needed sepsis prediction tool, building on the supervisors expertise and ongoing work in knowledge representation, machine learning and sepsis prediction.
Access to the Electronic Health Records of King’s College Hospital. This will be done through an existing governance framework to issue a research passport through a KCH honorary contract.
explainable AI, knowledge representation, machine learning, sepsis prediction, personalised medicine