Contribution of Structure Learning Algorithms in Social Epidemiology: Application to Real-World Data

Area of Science:

• Social Epidemiology
• Data Analysis
• Network Analysis

Background:

• Epidemiologists analyze complex, large datasets with growing analytical techniques like machine learning.
• Understanding variable relationships, causality, and network structures is critical in observational studies.

Purpose of the Study:

• To evaluate the contributions and limitations of structure learning (SL) methods in social epidemiology.
• To explore SL's application in identifying determinants of healthcare system access.

Main Methods:

• Applied SL techniques to the 2010 SIRS cohort data (N=3006) from Paris.
• Compared automated SL algorithms (with/without constraints) against a non-automated epidemiological method (expert network, logistic regression).
• Analyzed healthcare utilization as the outcome, with determinants including health status, demographics, and socio-economic factors.

Main Results:

• Both approaches identified similar interdependencies and relative strengths between variables.
• SL algorithms detected fewer associations with the outcome compared to the non-automated method.
• Purely data-driven SL approaches sometimes produced incorrectly oriented relationships.

Conclusions:

• SL methods are valuable for exploratory analysis and hypothesis generation in social epidemiology.
• Results from data-driven SL require validation against existing knowledge and further confirmatory analysis.
• SL can assist in mining novel databases for uncovering complex variable interactions.