Deep Learning-Adjusted Monitoring of In-Hospital Mortality after Liver Transplantation

Background: Surgeries represent a mainstay of medical care globally. Patterns of complications are frequently recognized late and place a considerable burden on health care systems. The aim was to develop and test the first deep learning-adjusted CUSUM program (DL-CUSUM) to predict and monitor in-hospital mortality in real time after liver transplantation. Methods: Data from 1066 individuals with 66,092 preoperatively available data point variables from 2004 to 2019 were included. DL-CUSUM is an application to predict in-hospital mortality. The area under the curve for risk adjustment with Model of End-stage Liver Disease (D-MELD), Balance of Risk (BAR) score, and deep learning (DL), as well as the ARL (average run length) and control limit (CL) for an in-control process over 5 years, were calculated. Results: D-MELD AUC was 0.618, BAR AUC was 0.648 and DL model AUC was 0.857. CL with BAR adjustment was 2.3 with an ARL of 326.31. D-MELD reached an ARL of 303.29 with a CL of 2.4. DL prediction resulted in a CL of 1.8 to reach an ARL of 332.67. Conclusions: This work introduces the first use of an automated DL-CUSUM system to monitor postoperative in-hospital mortality after liver transplantation. It allows for the real-time risk-adjusted monitoring of process quality.