BMC ARTIFICIAL INTELLIGENCE REPRODUCIBILITY INFORMATION
Collection: AI in chronic disease management

STUDY
Benchmarking Machine Learning Models for Cardiovascular Disease Classification in Population Health Surveys: Discrimination, Calibration, Interpretability, and Transportability

ANALYTIC COHORT
Source analytic records: 455,341
Countries or territories in source file: 89
Countries with at least one valid harmonized CVD outcome: 48
Countries eligible for binary benchmarking: 46
Excluded single-class surveys: Bahamas and Sudan
Primary analytic participants: 187,791
CVD cases: 16,305
Subregions: 15
Survey years represented in benchmarking cohort: 2013 to 2023

OUTCOME
cvd_history: self-reported history of heart attack, angina/chest pain from heart disease, or stroke.
The outcome is prevalent reported CVD and is not a clinically adjudicated incident endpoint.

PREDICTOR SETS
Full set, 16 predictors:
age, gender, residence, edlev, occupation, mstatus, smokehis, alcharm,
physical_activity, highsalt, lfi, lvi, bmi, hypertension, diabetes, dash_score.

Detection-adjusted set:
The full set with hypertension and diabetes removed.

VALIDATION DESIGN
Primary geographic validation: five-fold country-grouped internal-external cross-validation.
All participants from a country are assigned to one external fold.
Development-calibration partition: 15 percent of development observations, stratified by CVD outcome.
No external-country observation is used for imputation, encoding, model fitting, calibration, or threshold selection.

TEMPORAL BENCHMARKS
Mongolia: 2013 development -> 2019 evaluation
Uganda: 2014 development -> 2023 evaluation
Viet Nam: 2015 development -> 2021 evaluation

RANDOM SEED
20260903

CLASS-IMBALANCE AND SURVEY WEIGHTING
Survey weights are converted to numeric form and normalized to mean 1 within country-year.
Nonpositive or missing survey weights are replaced by the median positive weight in the same country-year.
Country-year balancing uses an inverse sample-size factor scaled to the mean country-year sample size.
Class factors balance the weighted positive and negative outcome mass during model fitting.
Synthetic oversampling is not used in the primary analysis.
PR AUC is reported together with ROC AUC.

MODEL SETTINGS
Logistic regression: L2 regularization; C = 1.0 in the reference implementation.
Elastic net: C = 0.5; L1/L2 mixing ratio = 0.5.
Random forest: 160 trees; maximum depth 14; minimum leaf size 10; square-root feature sampling.
XGBoost: 200 boosting rounds; depth 4; learning rate 0.05; subsample 0.8; column subsample 0.8;
minimum child weight 5; L2 regularization 1.0; binary logistic objective.

PREPROCESSING
Numeric predictors age, BMI, and DASH score: median imputation. Numeric predictors are standardized for regression models.
Categorical predictors: modal imputation, rare development levels with fewer than 10 observations collapsed,
then one-hot encoding. Evaluation-only categories are handled without refitting the encoder.

PROBABILITY CALIBRATION
Platt-type logistic recalibration is fitted only in the development calibration partition.
The logit of the raw model probability is the single recalibration predictor.
Classification thresholds from 0.02 to 0.70 in steps of 0.01 are evaluated in calibration data.
The threshold maximizing survey-weighted F1 is carried to the untouched external fold.

PRIMARY METRICS
ROC AUC
PR AUC
Brier score
Calibration intercept
Calibration slope
Sensitivity
Specificity
Positive predictive value
Negative predictive value
F1 score

REFERENCE SOFTWARE ENVIRONMENT USED TO VERIFY THE NUMERICAL PACKAGE
Python 3.13.5
NumPy 2.3.5
pandas 2.2.3
scikit-learn 1.8.0
XGBoost 3.1.3
SHAP 0.50.0
Linux x86_64, glibc 2.41

R REPRODUCTION IMPLEMENTATION
Additional File 2 provides an R implementation using:
data.table
Matrix
glmnet
ranger
xgboost
ggplot2

The R script writes sessionInfo() to analysis_outputs/R_sessionInfo.txt when executed.
For archival reproducibility, retain that generated sessionInfo file with the executed analysis outputs.

DATA REDISTRIBUTION CAUTION
Additional File 3 is included in this private submission package because it was requested as an analysis-ready supplementary dataset.
Do not upload the participant-level dataset to the journal until WHO STEPS and applicable country-specific data-use terms have been checked and confirmed to allow redistribution.
If redistribution is not permitted, submit the code, data dictionary, harmonization specification, and aggregated tables, and provide the appropriate controlled-access data availability statement.

FILE MANIFEST AND SHA-256 HASHES
da3f62265160e7a23d174bdc177df32ae30433ac734e4b0f04f5ffae9df4b379  /mnt/data/BMC_AI_CVD_Supplementary_Package/Additional_File_1_Supplementary_Tables.xlsx
8aff37109a7972054acecef05f135aba0d483c876e4e968205a1097527f956dd  /mnt/data/BMC_AI_CVD_Supplementary_Package/Additional_File_2_Reproducible_R_analysis.R
672c367a1a6d1452cc1b2914a445a479ffbc1b7c35bddbaf237957b93b44a898  /mnt/data/BMC_AI_CVD_Supplementary_Package/Additional_File_3_CVD_analysis_dataset.csv.gz
e10a070a3676ca8b106d36e6b27404cf6dd5967bc390a3ac4aa9bffd5da0daef  /mnt/data/BMC_AI_CVD_Supplementary_Package/README_SUPPLEMENTARY_FILES.txt