Incorporating intrathoracic pressure fluctuation can improve noninvasive prediction of left ventricular end-diastolic pressure

Abstract Introduction Heart failure (HF) is a leading cause of mortality and morbidity, with a prevalence above 10% in elderly population (1). Left ventricular end-diastolic pressure (LVEDP) is a known indicator for all types of HF. An invasive pressure catheter is the gold standard for measuring LVEDP (2). However, it involves an in-hospital procedure and is costly, which limits the accessibility. A noninvasive strategy would improve early diagnosis and reduce the burden on secondary care. Several noninvasive classifiers are able to detect elevated LVEDP (3-5), with some receiving FDA clearance (6). While beneficial for triaging patients, these classifiers cannot quantify the severity and track disease progression. Developing a noninvasive LVEDP regression model requires accounting for respiration dynamics. Even invasively, the dynamics make measuring LVEDP challenging. This study uses an error metric from ANSI/AAMI/ISO 81060-2:2019 (AAMI) (7) to evaluate the effect of respiration on noninvasive LVEDP regression results. Methods The dataset (147 patients, 21+ yrs old, mean 65 ± 10 yrs; 68% male) consisted of patients referred for nonemergent left heart catheterisation inclusive of direct assessment of LVEDP. Noninvasive data was collected by a modified brachial cuff with a single-lead electrocardiogram (ECG). Invasive data was acquired using a Millar pressure catheter. LVEDP was derived based on the time of the R-wave of the ECG (8). For each patient, 40 secs of noninvasive data with synchronised catheter signals were analysed to build an LVEDP regression model. Inputs were generated from the ECG and brachial pressure waves based on fiducials identified in both signals. Leave-one-subject-out cross-validation (LOSO CV) was used to train and test the algorithm. The impact of respirophasic variation on the model was analysed by comparing the prediction with the mean LVEDP and the range of the mean ± standard deviation (std) (i.e. AAMI error metric). Results The results, evaluated using Bland-Altman and correlation analysis, showed a decrease in the mean absolute error (MAE) from 3.6 mmHg to 1.4 mmHg, the std from 4.9 mmHg to 3.0 mmHg, and the correlation (r) from 0.57 to 0.77 when incorporating intrathoracic pressure swings (Fig.1 and Fig.2). A Levene test on the residuals of the two error determinations showed a significant difference (p-value < 0.01). Conclusion Irrespective of the error metric, the model performance meets the AAMI standard for blood pressure (BP) measurement. Moreover, when respiratory motion is included, agreement with invasive LVEDP improves considerably. From a clinical perspective, this emphasises the importance of targeting end-expiratory LVEDP. With respect to noninvasive BP measurement, it suggests that both brachial BP evaluation and LVEDP prediction should attempt to quantify fluctuations. While the challenge remains, these results indicate that noninvasive LVEDP prediction could be realised in the near future.Fig.1 Prediction vs mean LVEDP.Fig. 2 Prediction vs mean +/- std.