Transforming the care of patients with delirium using machine learning and bedside electroencephalography

In this five-year project at Imperial College London, I will develop and validate Neurostate, a machine learning-based system that measures encephalopathy and seizure activity from portable electroencephalography (EEG) signals. I am a neurologist with a background in engineering, neuroscience, and experimental medicine, making me uniquely able to lead this research and drive Neurostate's future translation into NHS hospitals. Patients with abnormal states of consciousness-ranging from mild 'confusion' to 'coma'-are a clinical challenge because we cannot accurately measure conscious states. Behavioural assessments like the Glasgow Coma Scale, based on patients' responses to commands and painful stimuli, are subjective, crude, and confounded by patients' sensorimotor and language function. Clinically significant aspects of conscious state can therefore be missed. The most important example is delirium, an acute disturbance of consciousness. Delirium is a major problem for the NHS, affecting 25% of older hospitalised patients. Delirium is triggered by a range of insults, including infection and metabolic abnormalities, causing a diffuse disturbance in brain activity called encephalopathy. Delirium is associated with an increased risk of morbidity and mortality, increased hospital length-of-stay, and healthcare costs. However, 50% of delirium cases go unrecognised by healthcare professionals. In ~10% of patients, it is caused by ongoing non-convulsive seizure activity, which is also critically under-recognised. We need objective, accurate measures of both encephalopathy and seizure activity at the bedside. Clinical EEG, acquired and interpreted by specialists, does identify encephalopathy, and shows distinct changes in seizures. However, it needs expert staff, expensive equipment, is poorly available in NHS hospitals, and - currently - does not provide objective measures. Importantly, new easy-to-use portable EEG (pEEG) hardware means that EEG signals can now be acquired by non-specialists at the bedside. Machine learning (ML), which finds patterns in complex data, can turn these signals into objective, quantitative measures that index encephalopathy and seizure activity. I will develop a software pipeline for EEG signal processing and, using large-scale (>10,000) existing clinical EEG datasets, train ML models to identify encephalopathy and seizure activity. I will evaluate different ML architectures, including conventional, feature-based, approaches and contemporary deep learning models. I will enhance the best-performing ML models by linking the Imperial NHS Trust's EEG database (~5,000-10,000 EEGs) to clinical variables extracted from electronic health records in the same patients, allowing me to incorporate, e.g., patients' co-morbidities and medications. I will optimise the models for pEEG hardware, before validating the approach by prospectively collecting pEEG and other data in hospitalised patients with and without delirium (N=200), and in patients with seizures (N=30). Through patient, public, and professional involvement, I will explore how a Neurostate-based medical device could best be deployed in hospitals. Implementation of Neurostate in a medical device, providing accurate, informative, point-of-care measures obtainable by any healthcare professional, can transform the care of patients with delirium and seizures, enabling us to detect and treat these conditions earlier and more appropriately, and monitor them accurately. This can reduce complications, improve outcomes, shorten length-of-stay and reduce costs, benefitting the NHS and other healthcare systems worldwide.