Portal de Periódicos da CAPES

1365: MACHINE LEARNING ALGORITHM TO PREDICT EMERGENCY DEPARTMENT VISITS IN PATIENTS ON ACTIVE CHEMOTHERAPY

2023; Lippincott Williams & Wilkins; Volume: 52; Issue: 1 Linguagem: Inglês

10.1097/01.ccm.0001003620.47154.09

1530-0293

Wedad Awad, Ammar Gharaibeh, Aseel Abusara, Razan Za’tra, Aysheh Al-Barawi, Karim Al-Sawalha, Lama Nazer,

COVID-19 diagnosis using AI

Introduction: Most patients with cancer receive their chemotherapy in the ambulatory setting. However, during their chemotherapy cycles, they may develop adverse events requiring visits to the Emergency Department (ED). We aimed to apply a machine-learning model to predict ED visits among ambulatory cancer patients on active chemotherapy. Such model would help clinicians proactively identify patients who are at high risk for adverse events that necessitate ED visits, as well as predict necessary resources. Methods: This was a retrospective study that included adult patients with cancer who received cancer-related treatment in the chemotherapy infusion clinics of a comprehensive cancer center. The electronic medical records were used to identify all patients scheduled for a clinic visit over 1 year and to identify those who had ED visits during the study period. Patient baseline characteristics, as well as the type of malignancy and chemotherapy, and laboratory results were extracted. We applied Extreme Gradient Boosting (XGBoost) to predict the ED visits following chemotherapy administration. The data was split into training and test cohorts. The SMOTE technique was utilized to handle the imbalance between the two groups. The model was evaluated using various metrics including precision, recall, F1-score, accuracy, and AUC-ROC. Results: We included 3,285 patients who received total 32,758 chemotherapy cycles. The average age was 53 + 14 (SD), 73% were females, and the majority had breast cancer (56%). ED visits were recorded for 1846 (57%) patients during 5168 (16%) of the administered chemotherapy cycles. Among the features utilized in the model, gender and smoking status had the highest correlation with ED visits. Using the XGBoost Gradient algorithm, we achieved the best performance with an accuracy of 0.88 and an AUC-ROC of 0.813. It demonstrated precision, recall, and F1-scores of 0.9, 0.96, and 0.93 for 'No ED visit' predictions, and 0.66, 0.44, and 0.53 for 'Yes ED visit' predictions, respectively. Conclusions: The generated machine-learning model had overall good performance to predict ED visits in patients on active chemotherapy in the ambulatory setting. External validation is necessary to further enhance the model performance and its clinical utility.