Predicting the Risk of Cardiovascular Diseases in the Elderly Based on Clinical Data and Heart Rate Variability Using...

Cardiovascular diseases (CVDs) remain the leading cause of morbidity and mortality in the elderly worldwide. Over the past two decades, there has been a wealth of evidence of a close relationship between autonomic nervous system activity and cardiovascular mortality, including sudden cardiac death. Heart rate variability (HRV), derived from photoplethysmographic (PPG) signals, is increasingly recognized as a promising non-invasive digital marker for evaluating autonomic nervous system function and stratifying CVD risk. The application of machine learning algorithms to PPG-derived HRV analysis offers a promising approach for improving CVD risk stratification and facilitating the development of personalized medicine strategies. Background/Objectives: To evaluate the potential of heart rate variability indicators in predicting the risk of developing CVD in individuals aged 65 years and older. Methods: The study involved individuals aged 65 years and older, divided into two groups: those with a risk of developing CVD (n = 54) and those without risk (n = 46). The first stage included a questionnaire as well as anthropometric and hemodynamic measurements. At the second stage, a PPG was performed using the Eldar computer photoplethysmograph and Eldar-Vario software, followed by an analysis of time-domain and spectral HRV parameters. Statistical data analysis was conducted using the SPSS Statistics 22.0 software package, focusing on the evaluation of associations between HRV indicators and the presence of CVD. Interpretable machine learning models were developed using logistic regression and a random forest algorithm within a nested cross-validation framework. In addition to the discriminatory characteristics, Brier score, LogLoss, calibration analysis, error matrices, permutation importance, and SHAP interpretation were analyzed in the study. Results: In patients with cardiovascular diseases, a statistically significant decrease in heart rate variability was revealed: SDNN by 2 times (26 [Q1-Q3: 15, 35] ms), pNN50 by 3.5 times (4 [3, 5]%), TINN by 5 times (31 [20, 51] ms), and HRV by 2.5 times (6 [4, 8.7]). In addition, a decrease was seen in the spectral components of VLF by one-fold (2450 [Q1-Q3: 2450, 4500] ms²), LF by four-fold (750 [750, 1500] ms²) and HF by five-fold (450 [450, 750] ms²) (p < 0.05). At the same time, there was a significant increase in the VLF/HF and LF/HF ratios, which indicates a predominance of sympathetic activity. According to the results of the correlation analysis, statistically significant associations of HRV indicators with age, physical activity level, body mass index and systolic blood pressure were revealed. The results of machine learning also revealed the association of HRV with arterial hypertension, physical activity and BMI. The best final results were demonstrated by a random forest model with a combined set of clinical and HRV signs of HF and RMSSD (ROC-AUC was 0.9988). The signs of heart rate variability obtained by photoplethysmography demonstrated additional prognostic value in relation to clinical signs. PPG-derived HRV features demonstrated additional discriminatory value for cardiovascular risk stratification. Conclusions: The obtained data demonstrate a close association between the risk of developing cardiovascular disease and autonomic nervous system dysfunction. The decrease in heart rate variability is most pronounced in elderly individuals with existing cardiovascular disease and can be considered a potential tool for developing diagnostic, prognostic, and risk stratification strategies. The use of machine learning demonstrated that heart rate variability features obtained using photoplethysmography improve diagnostic prognostication and classification of cardiovascular diseases compared to models based solely on clinical data.