Predicting survival and treatment compliance of cancer patients enrolled in medication access programs across 20 countries.

Abstract

e13514 Background: Cancer treatment access programs operate worldwide. This study applied classical and machine learning methods to access program administrative data to identify predictors of survival and treatment compliance of cancer patients. Methods: We performed a retrospective cohort study of patient records in a multi-country cancer medication access program database. The study examined the association between basic information collected during enrollment with the outcomes of survival and treatment compliance (defined as the proportion of prescribed medication cycles obtained on time). All patients enrolled between June 2016 and September 2025 were included. Survival was analyzed using Kaplan-Meier, Cox proportional hazards, and random survival forest. Compliance with treatment was analyzed using linear regression and an explainable boosted machine model. Analysis was performed in R and Python. The study protocol was developed prior to study execution. Data was obtained by querying the program database. Records with missing or inconsistent data were excluded. To account for confounding, all analyses included covariates capturing patient demographics and treatment context. Results: 18,556 patients met the inclusion criteria. 147 (0.8%) patients were excluded from the survival analysis due to missing (n = 116) or inconsistent (n = 31) data. 2,063 (11.1%) patients were excluded from the compliance analysis due to a missing compliance value (n = 1,928), missing data for other covariates (n = 116), or inconsistent data (n = 19). Final cohorts included patients from 16 low- and middle-income countries and 4 high-income countries. For the survival analysis, mean age at treatment start was 51.2 years (SD 18.1; sex ratio M:F = 0.85). Patients spent an average of 8.2 months (SD 8.6) in the access program, with 2,204 patients (12.0%) passing away while enrolled. All analyses revealed a strong association between survival and factors such as country, product, age, and program type while indicating a weak association for factors such as charity support and the specialty of a patient’s treating physician. For the compliance analysis, mean age at treatment start was 50.8 years (SD 18.1; sex ratio M:F = 0.82). Mean compliance was 78.9% (SD 35.7 pp). Both linear regression and an explainable boosted machine model showed country, product, and health insurance status as strongly associated with compliance, and other factors like sex, program type, and health facility type as having a weak association. Conclusions: This study identifies which data from access program administration is useful (and which is less useful) for predicting cancer patient survival and treatment compliance. Such predictions can help identify patients with higher mortality and non-compliance risk during enrollment and tailor interventions that improve these outcomes.