Predictive Machine Learning and Digital Health for Improving Patient Outcomes

Research question: Construct a real-time AI platform connected to hospital data streams for interconnected use in the Emergency Department, main hospital, and at discharge. Demonstrate patient benefit and improvements in hospital provision via the development of novel interconnected exemplar tools. Background: Ubiquitous acquisition of physiology, blood tests, scans, diagnoses, prescriptions, and equipment-location data results in very large repositories of data that mostly go unused in routine care. AI-based systems can exploit such real-time streaming data, also using repositories of historical data and data from unobtrusive patient-worn wearables, to improve patient care through the predictive assessment of health status. Available to this Research Professorship are some of the world's largest (non-imaging) datasets of their kind, exceeding 3 million admissions. Conventional methods of data analysis are inappropriate for handling such massively-multivariate and -multimodal data; recent advances in hospital-oriented AI are required to construct robust system at this new, large scale of data. Aims and objectives: Seven exemplar predictive tools will be created for use, focusing on (1) improving ED flow and patient care, (2) improving the treatment of deteriorating patients in the main hospital to reduce 'avoidable deaths', (3) supporting the patient-discharge process. For each of the seven tools, objectives are co-design of candidate models with clinicians and patients; blinded studies (2 months duration) of the prototypes via NHS partners; evaluation and revision of models with clinicians and patients; unblinded studies (6 months duration) of the revised prototypes; dissemination of results. Study durations and sizes are indicative and will be designed with collaborators during the programme. Methods: Novel 'end-to-end' deep learning architectures will be compared against conventional AI methods (XGBoost, etc.) and medical statistical benchmarks (logistic regression, Cox models as appropriate). The complexity of datasets comprising several million patient admissions, each being longitudinal and massively multivariate, requires new, explanatory methods to construct models optimally; such methods include 'continual' and 'sequential' learning, using single complex models across multiple interconnected tasks, and in estimating the relative 'informativeness' of individual examples within the dataset. Timelines for delivery: Exemplars will be developed and prototyped (years 1-2), deployed for study (years 2-3), iterated (years 3-4), and redeployed for study (years 3-5), with IP registration and applications for regulatory approvals during the programme. Anticipated impact: Varying by tool, with the main aim of demonstrating that a platform of interconnected predictive tools can both improve patient outcomes and improve the 'flow' of patients within hospitals. Impact metrics will be decided upon and tracked with clinical partners and reviewed regularly, including model efficacy (with respect to comparators), and subsequent changes in patient-care metrics (e.g., accuracy of discharge predictions, and any effect on length-of-stay / bed-occupancy / patient flow). Subsequent scale-up with industrial partners within and after the programme. Co-funding opportunities to build on the developed platform with clinical collaborators in various specialties, and in Primary Care. Dissemination via high-impact scientific and medical journals and conferences; annual workshops and public exhibitions; outreach via media, schools, and governmental-advisory roles, supported by senior appointment as a Research Chair in 'AI for Healthcare'.