Non-invasive monitoring of key hemodynamical and cardiac parameters using physics-based modelling and artificial intelligence

In a progressively aging population, it is of utmost importance to develop reliable, non-invasive, and cost-effective tools to estimate biomarkers that can be indicative of cardiovascular risk. Clinical parameters directly measured in the heart or the aorta are crucial for the diagnosis and management of disease. However, their clinical use is severely hampered by their invasive nature, cost, or need for special equipment. Aortic systolic blood pressure (aSBP), cardiac output (CO), end-systolic elastance (Ees), and arterial stiffness indices provide valuable information about the cardiovascular state in humans and are strongly associated with clinical outcomes. This thesis presents original predictive algorithms suitable for estimating such cardiovascular biomarkers from readily available, non-invasive clinical data.The first aim of this thesis is to develop and validate methods to estimate central hemodynamics. Firstly, an inverse problem-solving method is presented to estimate aSBP and CO from non-invasive measurements of cuff pressure and pulse wave velocity (PWV). The method relies on the adjustment of a previously validated computational arterial tree model. Assessment of the method indicates a high performance on a large human cohort. The second approach involves the machine learning-based estimation of aSBP and CO using again cuff pressure and PWV. Validation of the method on in silico data shows that machine learning offers a greatly accurate alternative for monitoring aSBP and CO. Additionally, we develop and test a gamut of machine learning frameworks for predicting Ees. First, a machine learning model is trained/tested using as inputs cuff pressure and PWV. The importance of incorporating ejection fraction (EF) as additional input is also assessed. Results indicate that Ees cannot be predicted from pressure-based data alone, while the addition of the EF is indispensable. Alternatively, we propose a novel artificial intelligence-based approach to estimate Ees using the information embedded in clinically relevant systolic timing intervals. Lastly, by means of a deep learning algorithm, we demonstrate that Ees can be accurately predicted using the morphology of the brachial blood pressure waveform. Furthermore, we aim to improve in vivo assessment of aortic characteristic impedance (Zao) and total arterial compliance (CT). Given that regional PWV measurements are non-invasive and clinically available, we present a non-invasive method for estimating Zao and CT using cuff pressure, cfPWV, and carotid-radial PWV via regression analysis. In silico validation verifies that the method may offer a valuable tool for assessing arterial stiffness while reducing the cost and the complexity of the existing techniques. As a step forward, we introduce a non-invasive method for estimating CT from a single carotid waveform using artificial neural networks. The proposed methodology is appraised using the large Asklepios human cohort. Precise estimates of CT are yielded, indicating that such an approach could offer promising applications, ranging from fast and cost-efficient hemodynamical monitoring by the physician to integration in wearable technologies. Finally, in view of the conflicting clinical and experimental evidence regarding the influence of HR on arterial stiffness, we highlight the importance of accounting for the effect of HR-induced blood pressure effects on the PWV measurement.

Noninvasive estimation of aortic hemodynamics and cardiac contractility using machine learning

Vasiliki Bikia, Stamatia Zoi Pagoulatou, Georgios Rovas, Nikolaos Stergiopulos

Cardiac and aortic characteristics are crucial for cardiovascular disease detection. However, noninvasive estimation of aortic hemodynamics and cardiac contractility is still challenging. This paper investigated the potential of estimating aortic systolic pressure (aSBP), cardiac output (CO), and end-systolic elastance (E-es) from cuff-pressure and pulse wave velocity (PWV) using regression analysis. The importance of incorporating ejection fraction (EF) as additional input for estimating E-es was also assessed. The models, including Random Forest, Support Vector Regressor, Ridge, Gradient Boosting, were trained/validated using synthetic data (n = 4,018) from an in-silico model. When cuff-pressure and PWV were used as inputs, the normalized-RMSEs/correlations for aSBP, CO, and E-es (best-performing models) were 3.36 +/- 0.74%/0.99, 7.60 +/- 0.68%/0.96, and 16.96 +/- 0.64%/0.37, respectively. Using EF as additional input for estimating E-es significantly improved the predictions (7.00 +/- 0.78%/0.92). Results showed that the use of noninvasive pressure measurements allows estimating aSBP and CO with acceptable accuracy. In contrast, E-es cannot be predicted from pressure signals alone. Addition of the EF information greatly improves the estimated E-es. Accuracy of the model-derived aSBP compared to in-vivo aSBP (n = 783) was very satisfactory (5.26 +/- 2.30%/0.97). Future in-vivo evaluation of CO and E-es estimations remains to be conducted. This novel methodology has potential to improve the noninvasive monitoring of aortic hemodynamics and cardiac contractility.

2020

Novel non-invasive techniques for hemodynamic monitoring

Orestis Vardoulis

The monitoring of hemodynamic parameters is crucial for optimal patient management and treatment. However, established methods are often invasive and possibly inapplicable to sensitive age groups. Novel, non-invasive techniques have the potential to transform cardiovascular (CV) assessment, offering reduced hospitalization periods, lower infection risk and decreased staff-to-patient ratio. Cardiac output (CO) monitoring provides valuable insight on global tissue perfusion. To overcome limitations of traditional methods, a novel technique was proposed. Taking into account the basic principles of CV physiology we derived an equation where the required variables can be extracted non-invasively. The novel technique was validated in a controlled, in silico, environment against previously established wave analysis methods. Subsequent data analysis showed that the new method yielded superior accuracy and precision. The new method was further evaluated in vivo using a 2D transthoracic echocardiography protocol for CO estimation. The corresponding comparison showed that the SVB method provides highly reproducible and accurate estimates of CO when compared to ultrasound. The estimation of pulse transit time (PTT) affects critically the accuracy and precision of pulse wave velocity (PWV) measurements. We proposed the "Diastole-Patching" method (DPm), which estimates PTT based on a signal region in the diastolic foot of the wave and is more robust against noisy arterial waveforms. DPm and previously established techniques were tested with a computational model of the arterial tree where the "real" values of PWV are known. The novel method yielded the highest agreement, accuracy and precision even when the waves were distorted with noise. The technique was further evaluated against the corresponding â"reference" algorithm of a commercially available device with in vivo acquired waves. DPm yielded more precise and reproducible measurements of PWV. Total Systemic Compliance (CT) expresses the ability of the arterial system to dilate and store blood during cardiac ejection. Our aim was to develop a relationship between aPWV and CT. The proposed, generalized Bramwell-Hill model of the systemic arterial tree estimates CT by a single non-invasive measurement of PWV. The proposed scheme was found to accurately estimate CT from aPWV. The mechanical properties of woven Dacron grafts differ significantly from the human aorta. As a result, aortic graft implantation changes the aortic geometry and mechanical properties, thus affecting blood pressure and wave reflections. A dedicated cardiovascular model was developed to include graft geometry and in vitro measured Dacron properties. It was found that aortic reconstruction affected aortic hemodynamics. For the proximal graft, the primary reason for pulse pressure rise is augmentation of the forward wave. For the distal graft, the pulse pressure rise is associated with augmented wave reflections resulting from the compliance mismatch between the aorta and the distal graft. To conclude, this body of work has shown that given the current computational tools and biosensors, it is possible to design and validate a set of non-invasive techniques that will provide valuable clinical insight for a wide gamut of cardiovascular diseases.

EPFL2014

In vivo application and validation of a novel noninvasive method to estimate the end-systolic elastance

Vasiliki Bikia, Stamatia Zoi Pagoulatou, Georgios Rovas, Nikolaos Stergiopulos

Accurate assessment of the left ventricular (LV) systolic function is indispensable in the clinic. However, estimation of a precise index of cardiac contractility, i.e., the end-systolic elastance (E-es), is invasive and cannot be established as clinical routine. The aim of this work was to present and validate a methodology that allows for the estimation of E-es from simple and readily available noninvasive measurements. The method is based on a validated model of the cardiovascular system and noninvasive data from arm-cuff pressure and routine echocardiography to render the model patient-specific. Briefly, the algorithm first uses the measured aortic flow as model input and optimizes the properties of the arterial system model to achieve correct prediction of the patient's peripheral pressure. In a second step, the personalized arterial system is coupled with the cardiac model (time-varying elastance model) and the LV systolic properties, including E-es, are tuned to predict accurately the aortic flow waveform. The algorithm was validated against invasive measurements of E-es (multiple pressure-volume loop analysis) taken from n = 10 patients with heart failure with preserved ejection fraction and n = 9 patients without heart failure. Invasive measurements of E-es (median = 2.4 mmHg/mL, range = [1.0, 5.0] mmHg/mL) agreed well with method predictions (normalized root mean square error = 9%, rho = 0.89, bias = -0.1 mmHg/mL, and limits of agreement = [-0.9, 0.6] mmHg/mL). This is a promising first step toward the development of a valuable tool that can be used by clinicians to assess systolic performance of the LV in the critically ill. NEW & NOTEWORTHY In this study, we present a novel model-based method to estimate the left ventricular (LV) end-systolic elastance (E-es) according to measurement of the patient's arm-cuff pressure and a routine echocardiography examination. The proposed method was validated in vivo against invasive multiple-loop measurements of E-es, achieving high correlation and low bias. This tool could be most valuable for clinicians to assess the cardiovascular health of critically ill patients.

2021

Novel non-invasive techniques for hemodynamic monitoring

Orestis Vardoulis

EPFL2014

In vivo application and validation of a novel noninvasive method to estimate the end-systolic elastance

Vasiliki Bikia, Stamatia Zoi Pagoulatou, Georgios Rovas, Nikolaos Stergiopulos

2021