Machine learning model to predict mortality after discharge in hospitalized oncologic patients (pts) under active systemic therapy in the advanced setting: A multicenter cross-validation study.

Abstract

12121 Background: Prognostic factors for oncologic pts after surgery or curative systemic treatment have been described, including ECOG performance status, tumor staging and malnutrition. However, there is no solid evidence on which combination of variables best predicts mortality after hospitalization of metastatic cancer pts under active systemic treatment. Methods: Prospective multicentric study of pts hospitalized between 2020 and 2022 at the Oncology wards of Vall d’Hebron, Sant Pau and Mar Hospitals [PLANTOLOGY database] in Barcelona, Spain. Clinical factors such as ECOG, comorbidities, tumor characteristics, and laboratory results were collected at admission. Mental status (depression and anxiety) and QOL were assessed through the HADS and EORTC-QLQ30 questionnaires, respectively. Nutritional assessment was performed using the chair and hand grip tests. All variables were analyzed in uni- and multivariable regressions including a machine learning LASSO model to assess predictive discriminators of 30-day mortality after discharge. Missing data was imputed using Multivariate Imputation by Chained Equations. A bootstrap with 1000 iterations was used to validate the model and c-index. Results: Among 1,663 pts, 932 had advanced disease and were under oncologic treatment during the 6 months previous to urgent admission, our target population for model development and validation. Median age was 64 years, 51% had an ECOG > 1, median Charlson comorbidity index was 8 and 34% were under treatment in a clinical trial. The most frequent tumor types were lung (25%), colorectal (14%) and breast (12%) cancer. The most relevant factors associated with higher mortality at 30-day after discharge in LASSO model were high Charlson index, low neutrophil count, high LDH, poor ECOG status and progressive disease at admission (all p-values p < 0.05). The c-index corrected after bootstrap validation was 0.75. After adding the nutritional assessment, mental health status and QoL (subset of 606 with complete data), the predictive power of the model increased to a c-index corrected after bootstrap validation of 0.81. Our final prognostic model called the PRognostic Oncologic Plantology score (PROP) obtained a sensitivity of 0.75 and a specificity of 0.80 with an overall accuracy of 0.80. Only 10% of “low” PROP score pts (72% of the population) died within 30 days of discharge, as compared to 58% of early mortality in “high” PROP score pts. Conclusions: Our model, including clinical and analytical factors, predicts with accuracy the 30-day mortality of oncologic pts after discharge which is significantly improved with the addition of nutritional assessment and standardized questionnaires of QoL and mental status. The PROP score calculator will be built to help physicians adjust medical interventions for hospitalized cancer pts.