Survival Prediction with Machine Learning: Addressing Unequal Probability of Selection in Survey Data

<p class="p1" style="text-align: left;">Accurate all-cause mortality prediction is essential for assessing population health and guiding targeted public health interventions. Traditional survival models, such as the Cox proportional hazards model, are widely used for statistical inference but may potentially struggle with prediction due to their limited ability to capture nonlinear relationships and complex interactions. Machine learning approaches offer a promising alternative by providing greater flexibility in modeling complex survival data. However, their application to survey-based epidemiological studies is challenging, particularly when accounting for unequal probability of selection. Motivated by the Hispanic Community Health Study/Study of Latinos (HCHS/SOL), this work explores the methodological challenges of applying machine learning to survival prediction under such conditions. We review existing mortality prediction models, discuss machine learning methods suitable for handling unequal selection probabilities, and propose a structured workflow for implementing these approaches. By addressing key methodological considerations, this work lays the foundation for improving survival prediction in studies with unequal sampling designs and supports the development of more generalizable risk prediction models.