10104845Assessing the Feasibility of a Neurosymbolic Methodology for Bias Mitigation of Cardiac Arrest Early Warning in General WardsClosedCollaborative R&DInnovate UKClinical early warning scores (EWS) are widely used by hospital care teams to identify early signs of patient deterioration, to facilitate timely intervention. EWS intend to prompt early actions, such as increased nursing attention or informing clinicians for more intensive care. Traditional EWS, such as the commonly used NEWS2 score in NHS hospitals, issue warnings based on predetermined thresholds of vital signs, like arterial blood pressure or heart rate. Despite their prevalence, these tools suffer from low specificity, leading to high false alarm rates and contributing to 'alarm fatigue' among healthcare practitioners. Alarm fatigue arises when clinical workers are frequently exposed to EWS alarms, resulting in desensitisation, delayed, or no response, leading to delayed treatment and patient harm. Importantly, traditional EWS, like NEWS2, exhibit varying sensitivity values for different adverse events, being more effective at predicting mortality than cardiac arrest. This inconsistency poses a significant technological hazard in hospitals, with excessive, misleading, or missed alerts having serious consequences for both patients and clinical personnel. Machine learning enables computers to identify patterns within large amounts of data, and deep learning (DL), a subset of machine learning, mimics human neurons to achieve its goals. DL models trained on large amounts of Electronic Health Records (EHRs) have been used to enhance deterioration prediction of traditional EWS. However, clinical adoption of DL EWS (DEWS) is limited due to challenges in gaining healthcare worker trust. DL models function as black-boxes, relying on complex mathematics to detect patterns within data. However, they are prone to carrying forward the various sources of bias common to EHR data, e.g. bias reflecting clinical judgement. Furthermore, DEWS successful operation requires substantial amounts of data, rendering such models less accurate in the less intensively monitored general medical wards. These challenges contribute to the clinical reluctance to embrace DEWS. This proposal aims to address the bias and robustness issues of DEWS by using well-established clinical knowledge and guidelines to complement the statistical power of deep learning. The idea is to embed the ability to recognise and reason about clinical knowledge in a data-driven DL framework to identify sources of bias and to steer the predictions made by a DEWS when bias is detected. Such alignment with clinical guidelines will also help improve the accuracy of DEWS prognoses where data is insufficient due to sparse (less frequent) monitoring. In essence, this project will design and implement much-needed fair, robust and reproducible computational EWS.