Evaluation of GPT-4 for 10-year cardiovascular risk prediction: Insights from the UK Biobank and KoGES data

Figures

• 

  Figure 1. Example of a GPT prompt and response Tabular data extracted from UK Biobank and KoGES were organized and queried into a sentence format. The 10-year cardiovascular disease risk percentage was extracted using regular expressions from the corresponding answers. GPT: generative pretrained transformer, UK: United Kingdom, KoGES: Korean Genome and Epidemiology Study, HDL: high-density lipoprotein, LDL: low-density lipoprotein, BMI: body mass index.

• 

  Figure 2. Performance evaluation and comparison of the risk scoring methods in the UK Biobank and KoGES cohorts GPT-4’s performance was comparable to conventional risk prediction models. Substantial correlation was found between the GPT-based risk score, ACC/AHA risk score and Framingham risk score. (A) AUROC curves (UK Biobank). (B) AUROC curves (KoGES). (C) Scatterplot (UK Biobank, GPT-4 vs. ACC/AHA risk score). (D) Scatterplot (KoGES, GPT-4 vs. ACC/AHA risk score). UK: United Kingdom, KoGES: Korean Genome and Epidemiology Study, GPT: generative pretrained transformer, ACC/AHA: American College of Cardiology/American Heart Association, AUROC: area under the receiver operating characteristics curve.

• 

  Figure 3. Kaplan-Meier curves stratified by risk categories in the UK Biobank cohort All pairwise comparisons between curves with the log rank test with post-hoc Bonferroni correction were statistically significant. (A) Kaplan-Meier curve stratified by GPT-4 based risk category. (B) Kaplan-Meier curve stratified by Framingham risk score category. (C) Kaplan-Meier curve stratified by ACC/AHA risk score category. GPT: generative pretrained transformer, ACC/AHA: American College of Cardiology/American Heart Association, UK: United Kingdom.