5966 Predicting patient deterioration on PICU using artificial intelligence

Objectives Big data from the electronic health record (EHR) can be used to enhance patient safety, and to improve work-flow and communication. Paediatric intensive area units (PICU) are environments which are rich in EHR data and contain patients who are at high risk of sudden deterioration. We set out to utilise data from the EHR to predict deterioration using artificial intelligence (AI) techniques. Methods We utilised six years of EHR data (2120 patients) from Cambridge University Hospitals PICU between 5th April 2014 and 21st January 2020. From this we created a predictive model utilising machine learning techniques, namely a decision tree combined with an application of calculus and polynomials to allow for the analysis of trends in time-series data. Results Our algorithm is 100% sensitive and 50% specific (AUC 0.91) for predicting death within the following six hour time period, with 95% sensitivity and 70% specificity for a 12 hour time period (AUC 0.82). Conclusion We were able to accurately predict survival and non-survival at six and 12 hours into the future in PICU patients. The benefits of being able to do this are in part to spot the patient who is in danger. But it is also a means of describing numerically the intensity of the patients on the unit; this can allow better resource-management. And finally it is a communication aid for explaining the acuity of patients in PICU to both staff and carers. This result will be validated using data from Great Ormond Street PICU.