Synapse
⌘+K
Synapse
PulseExploreJournal ClubResearchersJournals
Instagram
HomeJournal ClubExplore
June 12, 2026European Journal of Preventive CardiologyOpen Access

Enhancing cardiovascular risk prediction with neighbourhood determinants of health: a machine learning analysis in a nationwide population-based cohort of 1.8 million patients

View Full Paper
Ask AI
Bookmark
Share

Why the study?

Do machine learning survival models improve 10-year cardiovascular risk prediction compared to a Cox proportional hazards model when incorporating social and environmental determinants in a primary care population?

Population

1,776,865 adults aged 25–84 years registered with general practices across Wales, mean age 47.4 years, 50.6%…

Comparison

Machine learning survival models incorporating… vs Cox proportional hazards model incorporating the…

Design

Cohort

Follow-up

median 15.2 years

Key result

Machine learning survival models (Kernal SVM C-index 0.7896) performed marginally worse than a Cox proportional hazards model (C-index 0.8082) for predicting 10-year CVD risk.

Authors

JBJ R G BrownPBP J BaptisteHHH Hajmohammadi

Discussion

Loading...

Member takes

Overview

Large cohort data confirm high first-time CVD incidence; leaves open whether neighbourhood determinants meaningfully improve clinical risk models.

Key Points

  • This study aims to improve cardiovascular disease risk prediction by integrating neighbourhood health determinants and comparing traditional and machine learning models.
  • Utilized longitudinal data from the Secure Anonymised Information Linkage (SAIL) Databank of 1.8 million patients.
  • Developed a Cox proportional hazards model to estimate 10-year CVD risk, incorporating social determinants like area deprivation and air pollution.
  • Compared performance with machine learning models including Random Forest, Support Vector Machine, and Neural Network.
  • Cox proportional hazards model achieved a C-index of 0.8082, outperforming machine learning models.
  • First-time CVD events recorded at 254,040 (14.3%) during the follow-up period.
  • Best AUC at 10 years for the Cox model was 0.8528, indicating superior predictive performance.

Study Design

Type

Cohort (n=1,776,865)

Multicenter

Yes

Structured PICO

Do machine learning survival models improve 10-year cardiovascular risk prediction compared to a Cox proportional hazards model when incorporating social and environmental determinants in a primary care population?

P
Population
1,776,865 adults aged 25-84 years registered with general practices in Wales, followed for a median of 15.2 years to predict incident CVD events.
E
Exposure
Machine learning survival models (Random Forest, Support Vector Machine [SVM], Gradient Boosting, Penalised Cox Regression, Neural Network) incorporating standard QRISK predictors plus area-level deprivation (IMD) and residential air pollution (NO2).
C
Comparator
Cox proportional hazards model incorporating the same clinical, social, and environmental predictors.
O
Outcome
Model performance for predicting 10-year incident cardiovascular disease (CVD) events, evaluated using concordance index (C-index), Brier score, and area under the curve (AUC) at 10-years.

Traditional Cox proportional hazards models performed marginally better than machine learning approaches for predicting 10-year cardiovascular risk when incorporating social and environmental determinants in a large primary care cohort.

Limitations

  • Having initially used a 1% sample, our results may be biased.
  • Having initially used a 1% sample, results may be biased

Cite This Study

Brown et al. (2026) conducted a cohort in Cardiovascular disease (n=1,776,865). Machine learning survival models vs. Cox proportional hazards model was evaluated on 10-year CVD risk prediction performance (C-index). Machine learning survival models (Kernal SVM C-index 0.7896) performed marginally worse than a Cox proportional hazards model (C-index 0.8082) for predicting 10-year CVD risk.

synapsesocial.com/papers/6a2bd1386550ea4541ffe9b1https://doi.org/10.1093/eurjpc/zwag249.192
View Full Paper
Ask AI
Bookmark
Share