Abstract Background: Rapid response systems identify deteriorating hospitalized patients and prevent unexpected mortality by providing immediate interventions. But there are also difficulties in costs and human resources for clinical implementation. Machine learning is an application of artificial intelligence that provides systems the ability to automatically learn and improve from experience without being explicitly programmed. We try to develop a machine learning based early warning system to predict intubation and cardiopulmonary resuscitation (CPR) using vital signs and laboratory values. Methods: The training set was composed of all patients admitted to a cancer hospital named Chonnam National University Hwasun Hospital for two years (event group 277 vs. non-event group 3576 cases), and the test set was made from all patients admitted to non-cancer hospital named Chonnam National University Hospital for four months (event group 61 vs. non-event group 2334 cases). The time point at which the patient's clinical deterioration was recognized in the ward was defined as ‘detection time’, and the time point at which the patient suffered airway intubation or CPR was defined as ‘event time’. The algorithm is developed using five vital signs (blood pressure, heart rate, body temperature, respiratory rate and oxygen saturation) and laboratory values. We compared the efficacy of developed models which were ensemble model, recurrent neural network (RNN), low gradient boosting (LGB), extreme gradient boosting (XGB) and random forest with established modified early warning score (MEWS) using receiver operating characteristic area under curve (ROC AUC) and precision recall AUC (PR AUC) analyses. Results: The ROC AUC value of the ensemble model using five vital signs was 0.9219 and the PR AUC was 0.0732. After adding laboratory values, the ROC AUC values of the ensemble, RNN, LGB, XGB, random forest improved to 0.9964, 0.9814, 0.9950, 0.9862 and 0.9962. The PR AUC values were also increased to 0.2326, 0.2140, 0.3252, 0.2095 and 0.1985, respectively. The developed models showed higher performance than MEWS (ROC AUC: 0.7253, PR AUC: 0.0468) Conclusion: Machine learning based early warning system using vital signs and laboratory results was feasible. And the performance was better than traditional MEWS model because the ensemble model showed highest value in ROC AUC and LGB showed highest value in PR AUC. Citation Format: Bo-Gun Kho, Min-Seok Kim, Tae-Ok Kim, Cheol-Kyu Park, Young-Chul Kim, Soo-Hyung Kim, In-Jae Oh. Development of machine learning based early warning for rapid response system in a single cancer center [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PO-054.
Read full abstract