Comorbidity-free longevity: Variations in predictive importance across social strata — The Association Specialists
  1. Schedule
  2. Oral Presentations - Conversations starters - population health & wellbeing
  3. Comorbidity-free longevity: Variations in predictive importance across social strata

Comorbidity-free longevity: Variations in predictive importance across social strata (23949)

Joseph Van Buskirk 1 , Philip Clare 1 2 3 , Saman Khalatbari-Soltani 1 2 , Edward Jegasothy 1 , Melody Ding 1 2
  1. University of Sydney, University Of Sydney, NSW, Australia
  2. Prevention Research Collaboration, University of Sydney, Sydney, NSW, Australia
  3. National Drug and Alcohol Research Centre, University of New South Wales, Sydney, NSW, Australia
Background

Despite the growing recognition of social and structural determinants on health, quantitative studies often focus on individual factors, with fewer studies investigating how upstream factors might modify the relationship between individual factors and health. Machine learning techniques can be used to model complex interactions and associative pathways among variables not easily identified by traditional methods. We modelled relationships between a range of social, intermediary, and individual level variables and comorbidity-free longevity (CFL).

Method

We used an ensemble learning algorithm to assess time to presence of significant comorbidities (as defined by the Charlson Comorbidity Index), or death, in the UK Biobank baseline cohort (n=498,766). This approach fits and weights multiple candidate algorithms and returns a final ensemble model with the highest predictive performance. We ranked variables by predictive importance, overall and within socioeconomic deprivation quintiles.

Results

Models performed well, achieving integrated Brier scores of 0.05 or under, and time-dependent AUC curves ranging from 0.70 to 0.85. Material circumstances (e.g. housing tenure and nature of employment) were more important in predicting CFL in higher deprivation quintiles, while income itself was more important for prediction in lower deprivation quintiles. Risk factors such as smoking status were more important in predicting CFL in higher deprivation quintiles, while specific anthropometry measures (e.g. body fat percentage and weight) were more important in lower disadvantaged quintiles.

Discussion

This data driven approach suggests differences in the relationship between social determinants and CFL across disadvantage quintiles. Importantly, income itself may be less important for prediction of CFL in higher deprivation quintiles than other factors that help to achieve and maintain stability in material conditions. Given information gaps related to data driven approaches, such as the lack of clear explanatory pathways, these findings warrant further scrutiny using more rigorous epidemiological methods.