Simple 17-Point Geriatric Score Predicts Death Risk in Older Cancer Patients

A new prognostic tool built from a standardised geriatric assessment could change how doctors decide on cancer treatment for people over 70. Researchers in France have developed and externally validated a mortality prediction score based on the Geriatric Core Dataset, known as G-CODE, that estimates an older patient’s risk of dying within six, twelve, and twenty-four months. The work, published in the British Journal of Cancer, addresses one of the most stubborn gaps in modern oncology: the near-total absence of reliable evidence on how to treat cancer in the very population that accounts for most cases.

The scale of that gap is striking. In 2012, adults aged 65 or over accounted for 48 percent of cancer cases, a proportion forecast to climb to 58 percent by 2035. Yet older adults remain chronically under-represented in oncology clinical trials, which means clinicians often extrapolate treatment decisions from studies conducted almost exclusively in younger, fitter patients. Ageing, moreover, is profoundly heterogeneous. Two patients of the same chronological age can differ dramatically in physical reserve, comorbidity burden, cognition, nutrition, and social support, all of which influence how well they tolerate therapy and how long they live.

The instrument behind the new score, the Geriatric Core Dataset, was itself born of consensus. A national and international Delphi process involving a multidisciplinary expert board in geriatric oncology distilled the essentials of a comprehensive geriatric assessment into ten items spanning seven domains: social situation, function, mobility, nutrition, cognition, mood, and comorbidity. The items include whether a patient lives alone and has a primary caregiver; activities of daily living and short instrumental activities of daily living scores; the timed up-and-go mobility test; weight loss over the previous six months and body mass index; the Mini-Cog cognitive screen; a mini Geriatric Depression Scale; and the updated Charlson Comorbidity Index. The goal was to make geriatric data collection in clinical trials brief, standardised, and comparable across centres.

To turn that dataset into a predictive instrument, the team drew on two prospective French cohorts of patients aged 70 or older with solid tumours or haematological malignancies, all referred for geriatric assessment before first-line treatment or a change in treatment. The Elderly Cancer Patients cohort, ELCAPA, from the Greater Paris area, served as the training set with 950 patients contributing complete follow-up data. The Geriatric Oncology Marseille cohort, GOM, provided an independent validation set of 766 patients. The populations differed in clinically meaningful ways: validation patients were younger, more likely to have localised disease, and far more likely to have lung or head and neck cancers, while training patients were more often living alone and carried a heavier comorbidity burden. That diversity was deliberate, because a score that only works in the population where it was built is of little use to anyone else.

The statistical machinery was rigorous. The researchers estimated crude hazard ratios for each G-CODE item against six-month mortality, then entered all items into multivariate Cox proportional hazards regression, using a stepwise selection process based on the Akaike information criterion to identify the optimal model. Regression coefficients were rescaled and rounded to integers using Cole’s algorithm, producing a score that is easy to compute at the bedside. Recursive partitioning analysis was used to find the best cut-points for the comorbidity index, which lacks an established threshold, yielding three classes of increasing risk. Missing data, which reached 38.5 percent for the comorbidity index in the training set, were handled with a random forest imputation algorithm, and a complete-case sensitivity analysis confirmed the results.

What emerged is a 17-point score built from nine of the original ten items. Five predictors carried the greatest weight in forecasting six-month death: a heavy comorbidity burden, particularly an updated Charlson index above 8, which contributed 3 to 5 points; an abnormal timed up-and-go test, indicating impaired mobility, worth 2 points; loss of independence in instrumental activities of daily living, also 2 points; the absence of a caregiver, 2 points; and weight loss exceeding 10 percent over the previous six months, again 2 points. Notably, simply living alone dropped out of the final model, suggesting that the presence of concrete support, rather than cohabitation itself, is what matters for survival.

Performance was solid and, crucially, held up outside the population in which the score was born. On its own, the G-CODE score achieved time-dependent areas under the receiver operating curve of 71.2 percent for six-month mortality, 68.4 percent at twelve months, and 68.1 percent at twenty-four months in the validation cohort. When combined with a baseline model including age, cancer type, metastatic status, and treatment, discrimination rose to between 78 and 82 percent, adding roughly four percentage points over the clinical model alone in both cohorts. Recursive partitioning of the continuous score defined three risk groups, below 7 points, 7 to 8 points, and above 8 points, and Kaplan-Meier curves showed a clear, graded relationship between higher risk categories and worse survival at every time point, with log-rank p-values below 0.0001. Calibration was excellent in the training set and remained good in validation, with a calibration slope of 0.99 at six months.

How does it compare with existing tools? The Geriatric Cancer Scoring System, the only other validated score for three-, six-, and twelve-month mortality, achieved higher discrimination, around 89 percent at six months, but at the cost of eighteen items including laboratory values such as the C-reactive protein to albumin ratio and creatinine clearance. The G-CODE score trades some accuracy for simplicity and standardisation, relying only on bedside geriatric measures. It is also, according to the authors, the first prognostic tool to demonstrate the value of the short four-item instrumental activities of daily living scale for predicting six-month mortality, and the first validated score to extend prediction out to twenty-four months, a horizon that matters when half of the patients in these cohorts died within two years.

The clinical implications cut both ways. For trialists, the score offers a standardised, consensus-based way to stratify older participants by frailty, something international guidelines have long demanded but trials have rarely delivered. For clinicians without easy access to a full geriatric assessment, the score can flag the impairments that matter most, inform conversations with patients and families, guide oncology decision-making, and trigger a referral for comprehensive assessment and targeted geriatric interventions. Predicting short-term mortality is precisely what is needed to avoid both over-treatment of patients unlikely to benefit and under-treatment of those who might.

The authors are candid about limitations. Both cohorts enrolled patients already flagged as potentially frail by a screening score or oncology assessment, so the findings need confirmation in the broader older population. In the training set, the comorbidity index had to be reconstructed from proxy variables because some required comorbidities were not recorded, a potential source of classification bias, although sensitivity analyses and the strong performance in the validation cohort, where the index was complete, argue against serious distortion. The tool also covers the most important geriatric domains but is not exhaustive. Even so, the study represents a rare commodity in geriatric oncology: a simple, externally validated, consensus-derived score that quantifies, in a single number between 0 and 17, how a patient’s ageing body, not just their tumour, will shape their survival.

Subject of Research: Development and external validation of a geriatric assessment-based mortality prediction score for older patients with cancer

Article Title: Using the Geriatric Core Dataset (G-CODE) to predict mortality in older patients with cancer: development and external validation

Article References: Couderc, A.-L., Martinez-Tapia, C., Mathoulin-Pelissier, S., Laurent, M., Bringuier, M., Gay, P., Caillet, P., Bertona, J., Greillier, L., Canovas, J., Soubeyran, P., Rostoft, S., Canouï-Poitrine, F., & Paillaud, E. (2026). Using the Geriatric Core Dataset (G-CODE) to predict mortality in older patients with cancer: development and external validation. British Journal of Cancer. https://doi.org/10.1038/s41416-026-03641-5

Image Credits: AI Generated

DOI: 10.1038/s41416-026-03641-5

Keywords: geriatric oncology, G-CODE, mortality prediction, geriatric assessment, prognostic score, older adults, cancer, ELCAPA cohort, frailty, external validation, Cox regression, comorbidity