Using Machine Learning to Predict Mortality in Older Patients With Cancer: Development and External Validation of the Geriatric Cancer Scoring System (GCSS)

Background: Establishing an accurate prognosis remains challenging in older cancer patients because of the population’s heterogeneity. Using machine learning, we developed and externally validated a new predictive score (the “Geriatric Cancer Scoring System”, GCSS) to refine individualized prognosis for older patients with cancer during the first year following a geriatric assessment (GA). Methods: Data were collected from two French prospective multicentre cohorts of cancer patients aged 70 and over referred for GA: ELCAPA (training set, n=2012) and ONCODAGE (validation set, n=1397). Candidate predictors included baseline oncological and geriatric factors and routine biomarkers. Predictive models were built using Cox regression, single decision tree (DT), and random survival forest (RSF) methods, comparing their predictive performance for 3-, 6- and 12-month mortality by computing externally validated Harrell’s C-indices. Findings: At 12 months, respectively 875 (43%) and 219 (16%) patients in the training and validation sets had died (mean age at baseline: 81±6 and 78±5, respectively; women: 47% and 70%; metastatic cancer: 50% and 34%). Tumour site/metastatic status, cancer treatment, weight loss >3 kg (previous 3 months), 5 or more prescription drugs, Eastern Cooperative Oncology Group-Performance Status ≥2, Activities of Daily Living score ≤5, impaired timed up-and-go test, low creatinine clearance and elevated CRP/albumin were identified as independent predictors in the Cox model. DT and RSF identified more complex combinations of features, with the G-8 score, the tumour site/metastatic status and the CRP/albumin ratio contributing most to the predictions. The RSF approach gave the highest C-index (3 months: 0.91 [RSF], 0.88 [Cox], 0.87 [DT]; 12 months: 0.86 [RSF], 0.82 [Cox], 0.81 [DT]) and was thus retained as the final model. Interpretation: The GCSS based on a random forest machine learning approach gave an accurate externally validated 3-, 6- and 12-month mortality prediction. The GCSS might improve decision-making processes and counselling in older cancer patients. Funding Information: The ELCAPA study was funded by the French National Cancer Institute (Institut National du Cancer, INCa); Canceropole Ile-de-France; and Gerontopole Ile-de-France (Gerond’If). The ONCODAGE project was funded by the French National Cancer Institute (Institute National du Cancer); and SIRIC BRIO (Site de Recherche Integree sur le Cancer – Bordeaux Recherche Integree Oncologie; grant: INCa-DGOS-Inserm 6046). Declaration of Interests: None declared. Ethics Approval Statement: Two French prospective multicentre cohorts were considered in the present study. The study protocols were approved by the appropriate independent ethics committees (ELCAPA: CPP Ile-de-France I, Paris, France; ONCODAGE: CPP Sud-Ouest et Outre Mer III).