Dynamic Prediction of Cardiovascular Death among Old People with Mildly Reduced Kidney Function Using Deep Learning Models Based on a Prospective Cohort Study.

Abstract

Introduction: Cardiovascular disease (CVD) is more likely to occur in old people with mildly reduced kidney function. We aimed to identify target features in this cohort to reduce cardiovascular death using deep learning models.

Methods: A total of 12,650 older people (age ≥60) with mildly reduced kidney function from Tianjin Community Health Promotion Prospective Study were recruited from 2014 to 2020. Cardiovascular death was verified by the death certificates from the provincial vital statistics offices. Mildly reduced kidney function was defined when estimated glomerular filtration rate (eGFR) between 45 mL/min/1.73 m2 ≤ and 90 mL/min/1.73 m2. Data were analyzed using Cox regression, random survival forest (RSF), DeepHit (DH), and Dynamic DH (DDH). Concordance Index (C-index) and Brier Score (B-S) were used to compare the models' performances.

Results: During the follow-up of 7 years, 838 people died of CVD (6.62%). Age, gender, hypertension, diabetes, and eGFR were closely related to cardiovascular death. Both accuracy and precision of models, predictive performance gets better as the number of follow-up visits increases. In predicting cardiovascular death, the C-index and B-S value of COX were only 0.711 and 0.001 at the first follow-up, and values were 0.767 and 0.073 at last time, respectively. This trend is similar in the other three models, with the DDH model standing, which showed the individual survival prediction with more accuracy at different time points (for the 6-year survival prediction, the C-index = 0.797 and B-S = 0.022 for the average of all time points) than the Cox, RSF, and DH.

Conclusion: A novel deep learning algorithm used in our study has shown its superior performance in the prediction of individual dynamics in longitudinal studies, which improves predictive power with increasing data input over time.