1609P A machine learning-based prognostic model to predict survival in patients with advanced cancer admitted to an acute palliative care unit

Although accurate prognostication is very important to assist with patients with advanced cancer at the end-of-life (EOL), current prognostic models have several limitations. We aimed to develop a machine learning (ML)-based prognostic model to predict survival at 2 weeks, 1 month, and 2 months in patients with advanced cancer referred to palliative care specialists. A total of 2,104 consecutive patients admitted to the acute palliate care unit (APCU) in a single tertiary cancer center between April 2015 and December 2019 were selected for analysis. Variables for developing the ML algorithm included age, gender, performance status, cancer type, body mass index, laboratory findings, and Edmonton Symptom Assessment System scores. Analysable patients (N=2,084) were divided into training set (N=1,667) and validation set (N=417), and the performance of the ML algorithm model was evaluated by accuracy, sensitivity, precision, specificity, area under the receiver-operating characteristic curve (AUROC), and area under the precision-recall curve (AUPRC). Among 2,084 patients, the median age was 67 years, 865 (41.5%) patients were female, and the most common cancer type was lung cancer (20.4%). As for the final performance of the ML algorithm, the ensemble model using gradient boosting, random forest, and logistic regression was the best in terms of AUROC and AUPRC. The AUROC and AUPRC for predicting 2-week, 1-month, and 2-month survival were 0.84, 0.79, 0.79, and 0.66, 0.77, 0.92, respectively. Because our data cannot reflect all characteristics of patients with imminent death, AUPRC was relatively low in 2-week mortality model. More precise results could be obtained by deep learning with multilayer perceptron using soft labelling (AUPRC 0.90). Our study demonstrated ML-based prognostic models capable of predicting short-term survival of 2 weeks, 1 month, and 2 months in patients with advanced cancer admitted to an APCU. Integrating these models in the real-world may help patients nearing the EOL to make timely clinical and personal decisions.