High evidence
More
Heart-disease models for 2,166 kidney patients lost accuracy moving between US and China
Study title: Cardiovascular Classification in Middle-Aged and Elderly CKD Patients: A Machine Learning Approach in China and the United States
- Evidence type
- Observational study
- Sample
- n = 2,166
- Sources
- 5 sourcesPeer-reviewed
- Published
- 1 min read
- Truvace record
- TRV-2026-1220cite or check it
Study title: Cardiovascular Classification in Middle-Aged and Elderly CKD Patients: A Machine Learning Approach in China and the United States
Researchers developed interpretable machine learning models to classify cardiovascular disease among middle-aged and elderly CKD patients using US NHANES (n=1057) and China CHARLS (n=1109) cohorts. Seven algorithms were trained and tested internally, with XGBoost achieving test AUC 0.753 in NHANES and logistic regression achieving 0.769 in CHARLS, followed by bidirectional cross-population external validation and SHAP analysis.
Read the study abstract
Background Cardiovascular disease (CVD) drives mortality in chronic kidney disease (CKD) patients. Conventional risk tools underperform in CKD, and few machine learning (ML) models undergo cross-national validation. We developed interpretable ML models for classifying CVD amongst Chinese and US middle-aged and elderly CKD patients. Methods We conducted retrospective cross-sectional analyses using two nationally representative cohorts: NHANES (2011-2016, US) and CHARLS (Wave 1 and 3, China). A total of 1057 CKD p…

The problem
In a study of kidney patients in the US and China, models built to classify heart disease performed worse when tested on the other country's patients, and key risk factors differed: age in the US, hypertension in China.
In brief
The findings matter because CVD drives mortality in CKD yet conventional risk tools underperform, and the observed drop in performance across countries plus differing top features (age in US, hypertension in China) suggests a single unified model may misclassify risk. Uncertainty remains about prospective clinical utility, as the work was retrospective cross-sectional classification rather than longitudinal prediction or implementation.
Main points
- Retrospective cross-sectional analysis included 1057 CKD patients from NHANES 2011-2016 and 1109 from CHARLS Wave 1 and 3.
- Seven ML algorithms were trained and optimized with internal test sets and bidirectional cross-population external validation.
- SHAP analysis showed Age ranked as the top contributing feature in the US cohort, whereas hypertension was the primary contributing features in the Chinese cohort.
The full story
The study used two nationally representative cohorts, NHANES 2011-2016 with 1057 CKD patients and CHARLS Wave 1 and 3 with 1109 CKD patients, to train seven ML algorithms with optimization and SHAP interpretability.
Internal testing identified XGBoost as best in NHANES and logistic regression as optimal in CHARLS, but bidirectional external validation showed performance decline across populations, with distinct SHAP profiles pointing to age in the US and hypertension in China.
Sources
5 on file- Diabetes, Obesity and Metabolismdoi:10.1111/dom.71380
- Scientific Reportsdoi:10.1038/s41598-025-18443-7
- Annals of Internal Medicinedoi:10.7326/annals-24-01926
- Frontiers in Digital Healthdoi:10.3389/fdgth.2023.1193467
- Clinical Journal of the American Society of Nephrologydoi:10.2215/cjn.15691020
Talk it over
The researchers trained models on 1057 kidney patients from NHANES and 1109 from CHARLS, then tested each on the other country's patients. What does a drop in performance across countries tell you about using one model everywhere?
Heart-disease models for 2,166 kidney patients lost accuracy moving between US and Chinahttps://truvace.com/item/cardiovascular-classification-in-middle-aged-and-elderly-ckd-patients-a-machine-
The researchers trained models on 1057 kidney patients from NHANES and 1109 from CHARLS, then tested each on the other country's patients. What does a drop in performance across countries tell you about using one model everywhere?
Answer in the threadAge ranked as the top feature in the US cohort, while hypertension did in the Chinese cohort. Which explanations for that difference would you want the researchers to check first, and why?
Answer in the threadIf you ran a kidney clinic in a country that was in neither cohort, how would you decide whether a model built on NHANES or CHARLS data could be used there?
Answer in the thread
If you have more time
The study classified existing cardiovascular disease in a cross-sectional design rather than predicting it over time. What kind of follow-up study would show whether these models help patients in a clinic?
Answer in the threadLook up how NHANES and CHARLS recruited their participants. Which differences in who was included could explain why the models did worse on the other country's patients?
Answer in the thread
ace
What readers say