O12\u2003Artificial intelligence-based risk prediction for death after emergency laparotomy using multi slice contrast enhanced computerised tomography

Abstract

\n \n \n Emergency laparotomy has a considerable mortality risk, with more than one in ten patients not surviving to discharge. Preoperative risk prediction using clinical tools is well established, however implemented variably. Preoperative CT is undertaken almost universally and contains granular data beyond diagnostics, including body composition, disease severity and other abstract features with the potential to enhance risk prediction. In this study we established the value of features extracted in an automated fashion from pre-operative CT in predicting 90-day post-surgery mortality.\n \n \n \n Anonymised CTs were collated from patients undergoing emergency laparotomy at ten hospitals in Southern England (2016–2017). For each case, axial portal venous abdominal/pelvic series were analysed using a pre-trained neural network, with each image converted into a matrix of numerical features. An elastic-net regression model to predict 90-day mortality was trained using these features and evaluated by bootstrapping with 1000 resampled datasets.\n \n \n \n A total of 136,709 images from 274 cases were available for analysis with a mean of 503 per case. Mortality within 90\u2009days occurred in 34 cases (12.4%) with an average NELA mortality prediction of 8.5%. On internal (bootstrap) validation, the elastic net model derived from CT yielded excellent performance (AUC 0.903 95%CI 0.897–0.909), significantly in excess of the NELA risk calculator (AUC 0.809 95%CI 0.736–0.875), with a broader prediction range (0.01%-89.71%).\n \n \n \n Artificial intelligence techniques applied to routinely performed cross-sectional imaging predicts emergency laparotomy mortality with greater accuracy than clinical data alone. Integration of these automated tools may be possible in the future.\n \n \n \n Automated analysis of CT can accurately predict risk of mortality after emergency laparotomy.\n