Deriving machine learning and visualization frameworks for decoding inter-related factors in preventing cardiac ageing

Tóm tắt

Background

Ageing produces changes in cardiac function that increases risks of cardiovascular disease (CVD). However, age-related changes in body composition such as visceral adiposity or sarcopenia are rarely studied in association with physical activity (PA) factors in older adults. Given that PA is widely advocated as a preventive strategy against CVD, we explored methods in machine learning particularly with visualization techniques to delineate these factors separately.

Methods

Participants without cardiovascular disease (CVD) from a community-based cohort had their PA factors collected prospectively. Doppler and tissue Doppler echo-derived mitral E and e' ratio (a marker of left ventricular filling pressure) (E/E¹), and anthropometric measurements were obtained simultaneously. Using the random forest model, multiple decision trees were created. Each tree represents a set of clinical parameters, such as the PA factors, the adiposity parameters, and clinical parameters (age, gender, blood pressure). Each tree was studied individually to provide a prediction, then merged together, to produce a majority prediction. (Figure 1). The magnitude of impact of each parameter on individual left ventricular filling pressure is depicted by intelligible visualization using SHapley Additive exPlanations (SHAP).

Results

Using repeated K-cross-validation on the train set (n = 473), we found the Random Forest Regressor with the most optimal hyperparameters, which achieved the lowest mean squared error. With the trained model, we evaluated its performance by reporting its mean absolute error and plotting the correlation on the test set (n = 119).  Based on Figure 2, the intensity of PA is the most important feature in determining LV filling due to its greatest average impact on the model output as indicated by the mean absolute SHAP values. Figure 3 describes the relationship between the features and their global impact based on the computed SHAP values for each instance. Increased intensity, duration and frequency of exercise contributes to a lower prediction of elevated left ventricular filling pressure whereas high BMI and waist circumference contributes to increased prediction of elevated left ventricular filling pressure.

Conclusion

Use of machine learning techniques with intelligible visualization is a promising method for discovering, targeting, monitoring, and reporting the outcomes of preventive strategies, physical activity or otherwise, for optimizing the cardiovascular health of older adults.