Correlation between non-invasive measurements and intracardiac pressures using machine learning techniques

Abstract Background Solutions targeting early recognition of congestion in heart failure (HF) patients have the potential to prevent readmissions and can thus significantly reduce the burden on HF care. A gold standard measure of congestion is the invasively measured pulmonary capillary wedge pressure (PCWP). However, the invasive nature and accessibility of this measurement limits its clinical use. Non-invasive approximation of the PCWP using biosensing wearables could be a promising replacement for HF monitoring. Purpose The primary aim of this retrospective study was to create a model that estimates the PCWP based on non-invasive measurements of vital signs using both traditional statistics and machine learning (ML) techniques. Methods The study cohort consisted of right-sided heart catheterizations between 23/6/2017 and 19/8/2022. The following models were used: linear regression or classification, k-nearest neighbors, random forest, gradient boosting and multilayer perceptron. The outcome measure for the regression models was the continuous PCWP as measured during the catheterization. The two outcome classes for the classification models were low (<12 mmHg) and high (≥12 mmHg) PCWP. Non-invasive mean arterial blood pressure (MAP), saturation, heart rate, weight and temperature measured at most 72 hours before or after the catheterization were collected as the features for the models. Additionally, ECG-signals acquired during the catheterization were used to calculate the heart rate variability (HRV). The data was split in a validation (20%) and training (80%) data set. The models were built based on the training set and then applied on the validation set to determine the coefficient of determination (R2) for the regression models and the AUC of the ROC for the classification models. Results A total of 859 catheterizations were included of which 31.2% had HF as primary diagnosis. Average age of the cohort was 59 ± 14 years and 52.2% were male. In the HF group, the features with the highest correlation with the PCWP were the HRV and gender (Table 1). All the used regression techniques resulted in low R2 values of up to 0.10 and the classification techniques in AUC of the ROC values of up to 0.61 (Table 2). Conclusion In the current study, PCWP could not be approximated with non-invasive measurements using traditional statistics and ML techniques. These findings support the notion that traditional measures for monitoring HF are poorly correlated with hemodynamic parameters. Perhaps repeated measurements over time could overcome the shortcomings of a single vital signs measurement for congestion evaluation. Future prospective research is needed to evaluate the potential of wearable devices measuring trends over a longer period and their potential in predicting hemodynamics.