Machine Learning to Predict Mortality in Older Patients With Cancer: Development and External Validation of the Geriatric Cancer Scoring System Using Two Large French Cohorts

Abstract

PURPOSE Establishing an accurate prognosis remains challenging in older patients with cancer because of the population's heterogeneity and the current predictive models' reduced ability to capture the complex interactions between oncologic and geriatric predictors. We aim to develop and externally validate a new predictive score (the Geriatric Cancer Scoring System [GCSS]) to refine individualized prognosis for older patients with cancer during the first year after a geriatric assessment (GA). MATERIALS AND METHODS Data were collected from two French prospective multicenter cohorts of patients with cancer 70 years and older, referred for GA: ELCAPA (training set January 2007-March 2016) and ONCODAGE (validation set August 2008-March 2010). Candidate predictors included baseline oncologic and geriatric factors and routine biomarkers. We built predictive models using Cox regression, single decision tree (DT), and random survival forest (RSF) methods, comparing their predictive performance for 3-, 6-, and 12-month mortalities by computing time-dependent area under the receiver operator curve (tAUC). RESULTS A total of 2,012 and 1,397 patients were included in the training and validation set, respectively (mean age: 81 ± 6 years/78 ± 5 years; women: 47%/70%; metastatic cancer: 50%/34%; 12-month mortality: 43%/16%). Tumor site/metastatic status, cancer treatment, weight loss, ≥five prescription drugs, impaired functional status and mobility, abnormal G-8 score, low creatinine clearance, and elevated C-reactive protein (CRP)/albumin were identified as relevant predictors in the Cox model. DT and RSF identified more complex combinations of features, with G-8 score, tumor site/metastatic status, and CRP/albumin ratio contributing most to the predictions. The RSF approach gave the highest tAUC (12 months: 0.87 [RSF], 0.82 [Cox], 0.82 [DT]) and was retained as the final model. CONCLUSION The GCSS on the basis of a machine learning approach applied to two large French cohorts gave an accurate externally validated mortality prediction. The GCSS might improve decision making and counseling in older patients with cancer referred for pretherapeutic GA. GCSS's generalizability must now be confirmed in an international setting.