Document resource
Background Machine Learning (ML) models have demonstrated ability to recognise out-of-hospital cardiac arrest (OHCA) and to improve acuity classification in ambulance triage.We describe a ML model that predicts OHCA outcome using unstructured text volunteered by 999 callers. It could help guide the ambulance response to OHCA.Methods A retrospective observational study of outcome data from the South Western Ambulance Service OHCA Registry linked to the unstructured caller narrative recorded in the Computer Aided Dispatch (CAD). Outcomes were stratified into categories of ‘no out-of-hospital (OH) resuscitation’, ‘OH resuscitation unsuccessful’, ‘survival to hospital’ and ‘survival to 30 days’.Natural Language Processing techniques identified n-gram tokens of one, two and three words in a Bag-Of-Words approach from the narrative. Vectorised into numerical weights by Term Frequency Inverse Document Frequency, these served as feature inputs to the categories as outcome labels using the supervised learning ensemble method of Gradient Boosting.Results The period of 01/04/2020-31/03/2023 identified 37,532 OHCA records. There was significant class imbalance – 24,349 ‘no resuscitation’, 9546 ‘unsuccessful’, 2257 ‘survived to hospital’ and 1373 ‘survived 30 days’.Caller narrative was succinct with a mean number of words of 6.18, median 5, range 1-44. Entries were characterised by typing errors. ‘Breathing’ had 76 variant spellings.The model had an overall accuracy of 83% in predicting incident outcome from the unstructured call narrative. Precision, recall and f1 scores respectively ranged from 0.72-0.87, 0.73-0.93 and 0.76-0.9. Accuracy was highest for classifying survived 30 days.Conclusion The ML model can predict patient outcomes from call input data for patients in OHCA with high accuracy. It could highlight incidents with high potential for good patient outcome.