RESUMEN
Background: Previous results on the association between the estimated glomerular filtration rate (eGFR) and stroke are mixed. Most studies derived the eGFR from serum creatinine, which is affected by non-kidney determinants and thus has possibly biased the association with stroke risk. Methods: In this cohort study, we included 429 566 UK Biobank participants (94.5% white, 54% women, age 56 ± 8 years) free of stroke at enrollment. The eGFRcys and eGFRcr were calculated with serum cystatin C and creatinine, respectively. Outcomes of interest were risk of total stroke and subtypes. We investigated the linear and nonlinear associations using Cox proportional hazards models and restricted cubic splines, corrected for regression dilution bias. Results: During an average follow-up of 10.11 years, 4427 incident strokes occurred, among which 3447 were ischemic and 1163 were hemorrhagic. After adjustment for confounders, the regression dilution-corrected hazard ratios (95% confidence intervals) for every 10 mL/min/1.73 m2 decrement in eGFRcys were 1.10 (1.05-1.14) for total stroke and 1.11 (1.08-1.15) for ischemic stroke. A similar pattern was observed with eGFRcr, although the association was weaker. When either type of eGFR was below 75 mL/min/1.73 m2, the risks of total and ischemic stroke increased exponentially as eGFR decreased. A U-shaped relationship was witnessed if eGFRcr was used instead. There was a null association between eGFR and hemorrhagic stroke. Conclusions: The risks of total stroke and ischemic stroke increased exponentially when the eGFRcys fell below 75 mL/min/1.73 m2.
RESUMEN
Background: Coronary heart disease (CHD) and cerebral ischemic stroke (CIS) are two major types of cardiovascular disease (CVD) that are increasingly exerting pressure on the healthcare system worldwide. Machine learning holds great promise for improving the accuracy of disease prediction and risk stratification in CVD. However, there is currently no clinically applicable risk stratification model for the Asian population. This study developed a machine learning-based CHD and CIS model to address this issue. Methods: A case-control study was conducted based on 8,624 electronic medical records from 2008 to 2019 at the Tongji Hospital in Wuhan, China. Two machine learning methods (the random down-sampling method and the random forest method) were integrated into 2 ensemble models (the CHD model and the CIS model). The trained models were then interpreted using Shapley Additive exPlanations (SHAP). Results: The CHD and CIS models achieved good performance with the areas under the receiver operating characteristic curve (AUC) of 0.895 and 0.884 in random testing, and 0.905 and 0.889 in sequential testing, respectively. We identified 4 common factors between CHD and CIS: age, brachial-ankle pulse wave velocity, hypertension, and low-density lipoprotein cholesterol (LDL-C). Moreover, carcinoembryonic antigen (CEA) was identified as an independent indicator for CHD. Conclusions: Our ensemble models can provide risk stratification for CHD and CIS with clinically applicable performance. By interpreting the trained models, we provided insights into the common and unique indicators in CHD and CIS. These findings may contribute to a better understanding and management of risk factors associated with CVD.