When living environment affects healthcare access and quality: a machine learning approach
Abstract
Rural and socially deprived areas often face limited access to healthcare services and professionals, leading to suboptimal care and poorer outcomes. The AIPAD project develops a scalable, data‑driven framework to quantify inequalities in peripheral artery disease (PAD) and build ML models that account for and reduce these biases, ultimately supporting surgeons and promoting more equitable care. We conducted an observational, longitudinal study using data from the French National Health Insurance Information System. All adult patients who underwent lower‑limb revascularisation between January 2016 and December 2024 were included. Socio‑demographic, medical, and intra‑operative variables were used to predict major adverse limb events (MALE) or one‑year all‑cause mortality. Logistic regression, random forest, and XGBoost models were evaluated using performance metrics (AUC, accuracy) and fairness measures, focusing on True Positive Rate Balance (TPRB) across FDEP‑defined socio‑economic groups. A total of 702,016 patients were analysed. Among them, 14.7% belonged to the first and least deprived FDEP quintile, while 22.8% were classified in FDEP quintile 5 (most deprived). The best-performing model was a XGBoost algorithm incorporating medical and intra‑operative variables (AUC=0.779). However, across FDEP quintiles 1 to 5, the AUC decreased from 0.787 to 0.774, and accuracy decreased from 0.714 to 0.694, indicating a socio‑economic gradient in model performance. The TPRB was −0.02, reflecting lower sensitivity in the most deprived group. These findings highlight the need for equity‑oriented strategies. In this workshop, we propose to present our results and illustrate the importance of sampling approaches to reduce performance gaps while preserving global accuracy. Overall, we will demonstrate that integrating social and territorial dimensions into ML models is essential to improve risk prediction while avoiding reinforcement of existing healthcare inequalities.
Related
- Digital health literacy, proactive health behaviours, the digital healthcare environment and health-related quality of life among older adults with type 2 diabetes in Chinese primary care: a cross-sectional study
- Professional practice environment and advanced nursing interventions associated with nurse-perceived quality of healthcare delivery in selected rural facilities in Saudi Arabia: a multisite cross-sectional correlational study
- Introduction to using symbolic regression for interpretable machine learning in healthcare
- Self-reported sleep quality, physical activity and anxiety associated with heart rate variability in youth tiers derived via unsupervised machine learning
- Biomechanical and movement-quality predictors of lower-limb load-related injury in collegiate athletes: a prospective cohort study using functional movement screening, wearable inertial sensing, and explainable machine learning
BibTeX
@article{cabon2026when,
title = {{When living environment affects healthcare access and quality: a machine learning approach}},
author = {T Cabon and V Garès and M Giacofci and Y Goueffic and N Le Meur},
journal = {The European Journal of Public Health},
year = {2026},
month = oct,
doi = {10.1093/eurpub/ckag169.712},
url = {https://doi.org/10.1093/eurpub/ckag169.712},
note = {CC BY-NC 4.0},
}
RIS
TY - JOUR
TI - When living environment affects healthcare access and quality: a machine learning approach
AU - T Cabon
AU - V Garès
AU - M Giacofci
AU - Y Goueffic
AU - N Le Meur
PY - 2026
DA - 2026/10/01
JO - The European Journal of Public Health
DO - 10.1093/eurpub/ckag169.712
UR - https://doi.org/10.1093/eurpub/ckag169.712
AB - Rural and socially deprived areas often face limited access to healthcare services and professionals, leading to suboptimal care and poorer outcomes. The AIPAD project develops a scalable, data‑driven framework to quantify inequalities in peripheral artery disease (PAD) and build ML models that account for and reduce these biases, ultimately supporting surgeons and promoting more equitable care. We conducted an observational, longitudinal study using data from the French National Health Insurance Information System. All adult patients who underwent lower‑limb revascularisation between January 2016 and December 2024 were included. Socio‑demographic, medical, and intra‑operative variables were used to predict major adverse limb events (MALE) or one‑year all‑cause mortality. Logistic regression, random forest, and XGBoost models were evaluated using performance metrics (AUC, accuracy) and fairness measures, focusing on True Positive Rate Balance (TPRB) across FDEP‑defined socio‑economic groups. A total of 702,016 patients were analysed. Among them, 14.7% belonged to the first and least deprived FDEP quintile, while 22.8% were classified in FDEP quintile 5 (most deprived). The best-performing model was a XGBoost algorithm incorporating medical and intra‑operative variables (AUC=0.779). However, across FDEP quintiles 1 to 5, the AUC decreased from 0.787 to 0.774, and accuracy decreased from 0.714 to 0.694, indicating a socio‑economic gradient in model performance. The TPRB was −0.02, reflecting lower sensitivity in the most deprived group. These findings highlight the need for equity‑oriented strategies. In this workshop, we propose to present our results and illustrate the importance of sampling approaches to reduce performance gaps while preserving global accuracy. Overall, we will demonstrate that integrating social and territorial dimensions into ML models is essential to improve risk prediction while avoiding reinforcement of existing healthcare inequalities.
SN - 1101-1262
SN - 1464-360X
ER -