Prediction of all-cause mortality in older adults in Colombia: real-world evidence from a middle-income country

Accurate mortality prediction in older adults is a critical component of precision public health and risk stratification in healthcare systems worldwide. In middle-income countries, however, real-world applications remain scarce due to limitations in data quality and methodological scalability. This study aims to develop and benchmark state-of-the-art predictive models for all-cause mortality among adults aged 80 years and older in Colombia's five major cities, leveraging a uniquely rich and comprehensive administrative health dataset. We constructed two predictive cohorts from high-dimensional, real-world data covering over 300,000 and 50,000 individuals for one- and two-year models, respectively. The dataset includes detailed records on thousands of prescribed medications, medical procedures, ICD-10 diagnoses, service provider and insurer identifiers, care modalities (outpatient, inpatient, emergency, home-based), and financial payment mechanisms, reflecting routine clinical practice across hundreds of healthcare providers and insurers. We implemented and compared five advanced modeling strategies: artificial neural networks, autoencoders, extreme gradient boosting (XGBoost), Bayesian regression, and TabTransformer (a self-attention architecture tailored for structured data). Model performance was assessed using area under the curve (AUC), F1-score, accuracy, and Brier score. Our research demonstrates that diagnostic codes, pharmacological profiles, and hospitalization patterns are the most predictive domains of late-life mortality, whereas operational variables (e.g., provider or insurer) add little incremental value. The highest discriminative performance was achieved with XGBoost (AUC = 0.92 for two-year, 0.89 for one-year predictions), followed by autoencoders (AUC = 0.85 and 0.87) and TabTransformer (AUC = 0.82 and 0.84). Neural networks showed moderate performance, and Bayesian regression faced convergence issues due to computational demands. Notably, some models achieved AUC values above 0.88, indicating strong discriminatory performance for mortality classification in older adults using real-world data, although F1 scores reflected moderate classification balance. This study demonstrates the feasibility and potential of deploying advanced machine learning techniques on richly granular administrative health data in a middle-income country. Our findings offer a robust methodological foundation for the early identification of high-risk older adults and provide actionable insights for health insurers, care providers, and policymakers within Colombia's General System of Social Security in Health. The integration of such models could inform targeted preventive strategies and resource allocation to improve outcomes in aging populations.