The study that paved the way to PROTEGER: what the Oncogeriatric Telecommittee looks at when it decides on a treatment

Before PROTEGER there was a more basic question: when the Oncogeriatric Telecommittee of Hospital Digital decides whether an older person with cancer is a candidate for treatment, which variables really weigh on that decision? That was the aim of the work published in Procedia Computer Science and presented at the 29th International Conference KES 2025 by Constanza Contreras-Piña and Patricio Wolff, of the Industrial Engineering Department of the University of Chile, together with Dr. Gabriel Martínez (Hospital Digital, Ministry of Health, and Hospital Penco-Lirquén) and Dr. Gonzalo Navarrete (FALP and the University of Chile Clinical Hospital), both members of the Oncoger team.

What was done

The team worked with the records of 357 patients assessed by the Telecommittee between 2021 and 2023 in different parts of the country: demographic data, cancer type, geriatric scales (Barthel, SPPB, GDS-15, MMSE, Charlson, number of drugs, falls, BMI), the EVF frailty scale used at Hospital Digital, ECOG and the committee's final decision. On that basis they trained three deliberately simple, interpretable models —logistic regression, categorical Naive Bayes and a decision tree— to predict whether the committee considered the patient eligible for treatment, and used SHAP values to measure the weight of each variable. They then repeated the exercise excluding ECOG, the scale oncology usually uses to summarise functional status.

Main findings

The central finding is that the committee prioritises the geriatric-assessment indicators over traditional ECOG. The three variables with the most weight were the Barthel Index, frailty (EVF) and ECOG; when ECOG was removed from the model, the predictive ability dropped by just 0.02 to 0.03 AUC points and the weight of frailty increased. The decision tree trained without ECOG, with only two levels of depth, shows the logic legibly: patients with greater frailty and less independence on the Barthel tend not to be candidates for treatment; the less frail and more independent ones are.

A principal-component analysis explains why: Barthel, EVF and ECOG are highly correlated and measure similar aspects of frailty and function. Two components concentrate 94% of their variance (84% the first one), and the Barthel contributes a partially distinct dimension complementary to the other two. Reducing the dimensionality allowed, in the authors' words, «a more transparent model without compromising predictive accuracy».

Why it matters

The authors highlight two practical consequences. The first is for primary care: the Barthel and the EVF are more accessible to a general practitioner than ECOG, so these rules can help decide whom to refer to the Telecommittee and how urgently. The second is methodological: in resource-limited settings with few cases, simple, interpretable models, well built, offer a reasonable balance between accuracy and explainability, something indispensable when the decision directly affects a patient. The study acknowledges its limits —a modest sample, almost half the patients from one region (Antofagasta) and missing values in variables such as BMI, which themselves turned out to be an indicator of frailty— and calls for validating the results in larger, more diverse cohorts.

That validation is exactly what came next. The 357 patients in this study are PROTEGER's development cohort, which in 2026 added external validation in two tertiary centres (n = 272, AUC 0.84) and a national cohort of 629 patients in six public hospitals. The lesson of this first study —that geriatric assessment contributes more than age or ECOG to the decision— is the design principle of the platform.