Objective:

To assess the predictive value of social determinants of health (SDOH) in estimating risks of disease progression to type 2 diabetes (DM2) and cardiovascular disease (CVD).

Approach:

• Data Creation: Created a dataset of primary and secondary care determinants.
• Model Comparison: Compared models using only biomedical risk factors with those incorporating both medical and social factors.
• Focus on SDOH: Evaluated the influence of SDOH on model performance, particularly in 5-year risk prediction.
• Benchmarking: Used survival models as benchmarking analyses alongside traditional statistical approaches and machine learning algorithms.

Key Findings:

• Traditional models often exclude SDOH, which may lead to underestimation of CVD risk in diverse populations.
• Machine learning algorithms, such as XGBoost, can better handle complex, high-dimensional data compared to traditional models.
• Incorporating SDOH into predictive models may enhance their accuracy and applicability across different demographic groups.

Interpretation:

The study aims to clarify the extent to which SDOH provide incremental predictive value beyond established biomedical risk factors in routine care data.

Limitations:

• Traditional models may perform comparably or better in low-dimensional contexts.
• Machine learning models require sufficient meaningful predictors and events for effective training.

Conclusion:

The study seeks to motivate the use of XGBoost for analyzing the added value of SDOH in clinically meaningful risk prediction settings.