Detection of Low Cardiac Index using a Polyvinylidene Fluoride-Based Wearable Ring and Convolutional Neural Networks

Abstract

This study investigated the use of a wearable ring made of polyvinylidene fluoride film to identify a low cardiac index (≤2 L/min). The waveform generated by the ring contains patterns that may be indicative of low blood pressure and/or high vascular resistance, both of which are markers of a low cardiac index. In particular, the waveform contains reflection waves whose timing and amplitude are correlated with pulse travel time and vascular resistance, respectively. Hence, the pattern of the waveform is expected to vary in response to changes in blood pressure and vascular resistance. By analyzing the morphology of the waveform, our aim was to create a tool to identify patients with low cardiac index. This was done using a convolutional neural network which was trained on data from animal models. The model was then tested on waveforms that were collected from patients undergoing pulmonary artery catheterization. The results indicate high accuracy in classifying patients with a low cardiac index, achieving an area under the receiver operating characteristics and precision-recall curves of 0.88 and 0.71, respectively.

Graphical Abstract

I. Introduction

Cardiac output (CO) is one of the most important indicators of cardiac health and hemodynamic balance. It measures the amount of blood that the heart supplies to the organs and cells, indicating how much oxygen and nutrients are delivered and how much waste is removed. Assessments of CO are frequently used to guide diagnosis and treatment in critical care settings. For example, cardiogenic and septic shock may both present with low blood pressure and similar presenting symptoms. The pathophysiology of these diseases, however, is markedly different with one having a low CO and the other (often) a high CO, respectively. The treatment will thus be different for the two diseases. Hence, identification of a low CO could enable expedited diagnoses and targeted therapy.

The Fick method is the gold standard for assessing CO and is defined as the ratio of oxygen uptake (VO2) to the systemic arterio-venous oxygen difference. This, however, can only be calculated intermittently and through invasive measurements of blood oxygen content [1]. While VO2 is most accurately measured directly using either mass spectometry analysis of exhaled air collected using a Douglas bag or a metabolic cart, it is more often estimated using derived formulas. Additional techniques for measuring CO include pulmonary artery thermodilution and Doppler ultrasound. The former is invasive and requires a medical provider to place a pulmonary artery catheter [2]. The latter is non-continuous and requires a trained ultrasound technician to obtain measurements [3].

Other alternatives for measuring CO have been proposed; however, these methods have not been studied rigorously. In addition, the majority of existing studies indicate reduced accuracy for these methods [4]. One such method is thoracic electrical bioimpedance (also known as impedance cardiography) which uses variations in electrical impedance to estimate CO [5]. However, several studies and clinical trials have found poor correlation between bioimpedance-derived CO and thermodilution techniques in both adult and pediatric populations [6], [7], [8], [9], [10]. Moreover, bioimpedance is sensitive to electrode positioning, electrical noise, movement, temperature and humidity [4], [3]. This makes it challenging to use thoracic electrical bioimpedance as a reliable tool to measure CO.

Another technique used for estimation of CO is pulse contour analysis [11]. This method is based on invasive arterial blood pressure monitoring and uses an extension of the Frank Windkessel model to derive CO from arterial blood pressure readings. A non-invasive version of this method uses the volume clamp technique to estimate arterial waveforms. The CO is then derived using pulse contour analysis. While some studies have shown promise for this method in controlled settings, clinical studies have shown weak correlations with gold-standard measurements of CO [12]. Moreover, the required devices are non-portable, require frequent calibration, and are not suitable for in-home monitoring. Cardiac output is a function of body size and is thus most often indexed to body surface area, termed cardiac index (CI).

Polyvinylidene fluoride (PVDF) based materials have been actively studied for sensing [13], energy harvesting [14] and transducers and actuators [15], among others, due to their flexibility, easy fabrication, inexpensive cost and shape adaptability. They have also have gained popularity in health monitoring applications such as pulse rate monitoring [16], [17], respiration monitoring [18], [19], sleep monitoring [20], [21], and body motion monitoring [22], since they are comfortable to wear and easy to use [23]. PVDF sensors are commonly composed of two plate-shaped electrodes covering a piezoelectric layer of the polymer. The polymer acts as dielectric material, forming a capacitor upon which charge is altered when a mechanical strain is induced.

In this study, we investigated the use of a wearable ring made of PVDF material to identify patients with a low CI (≤2 L/min). The waveform was analyzed using a convolutional neural network (CNN) that was trained on data from animal experiments and validated on data from patients.

II. Polyvinylidene Fluoride Ring

A. Sensor description

The piezoelectric sensor consists of a custom-built flexible PVDF sensing layer as shown in Figure 1a. The PVDF layer is silver plated with 52 μm thickness (Precision Acoustics) covered by layers of polyimide tape (Kapton Tape, Uline Inc.). A copper strip that is placed on the inside surface of the PVDF sensor acts as a grounding electrode and creates contact between skin and the electronics and reduces 60 Hz noise and other interference. The ring is composed of a Velcro or elastic band that covers the PVDF sensor, as shown in Figure 1b, with a screw that allows for tension adjustment. The ring tension was set to a relatively consistent level in all experiments. A more detailed description of the sensor and its mechanical and electrical dynamics can be found in [24]

Fig. 1:

(a) Depicts the sensor, including the PVDF sensing layer and a photoplethysmography sensor. The photoplethysmography sensor was not utilized in this study. The grounding electrode, shown on the top left corner, reduces the effect of the 60Hz power line noise and other interference. (b) The ring that encompasses the sensor, worn along with a signal collection wristband. The schematic diagram for the ring and the sensing circuit that is used to model electrical dynamics of the interface between the sensor and the signal acquisition system. The sensor capacitance and initial input resistance are denoted by C and R, respectively.

The ring may be worn at multiple different sites on the body. Patients in this study wore the ring on their finger, most often their fourth finger. Occasionally, it was placed on the second or third finger if the sensor signal from the ring was not felt to be optimal. In the swine model, the sensor was placed distally over the radial artery on the animal’s foreleg adjacent to the hoof. This setup uses animal’s smallest accessible radial artery which is analogous to the human wrist.

The analyses done in this study depend on the morphology of the ring waveform which is determined by systemic factors. Hence, morphology of the measured waveform is independent of the measuring site as long as the sites are on the same arterial branch and approximately at the same distance from the heart. As a result, the proposed models do not depend on the amplitude of the generated signal and no calibration is needed for the operation of the sensor. Furthermore, because the frequency range of cardiac cycles does not vary substantially, any dynamic effects should be relatively uniform in signals used for training the models. This does leave the possibility that features could be altered by differing non-linear dynamics from patient to patient; however, local modeling suggests that nonlinear effects are small due to the small displacement amplitude at the ring.

B. Signal conditioning

The PVDF sensor can be modeled as a charge source in parallel with a capacitor, as shown in Figure 1c. The output electrical charge can be converted to voltage using a commercial data acquisition system. In this study, all waveforms were measured using a general purpose analog amplifier with high input impedance (DA100C, BIOPAC Systems Inc.) and a gain of 200. The PVDF sensor has a zero-mean baseline output due to the high-pass filtering effect of the piezoelectric material in connection with a finite impedance output. In addition, the DA100C module applies a 10Hz low-pass analog filter that removes most of the high-frequency and power line noise. The waveforms were initially sampled at 200Hz and were later down-sampled to 128Hz for analysis.

A custom-built dedicated sensing circuit, shown in Figure 1b on the wrist, is also available for measuring the generated signal. The frequency response of this circuit is similar to that of the DA100C module. The sensing circuit, shown in Figure 1c, consists of a first stage amplifier and a low-pass filter to minimize high-frequency interference, and compensates for the low current from the PVDF sensor (1–5nA).

The elastic band that covers the sensor creates a firm contact between the sensor and the skin. After adjusting the tension with the screw, the signal can be measured with high fidelity and low noise levels when the finger is stationary. Movement of the finger can introduce large intermittent disturbances which dissipate quickly after termination of the movement. The effects of noise and motion artifacts on the signal are similar to that in photoplethysmography sensors that are widely used in clinical practice for measurement of oxygen saturation and in consumer products such as smartwatches for health monitoring. Hence, the authors do not envision noise and motion artifacts to be barriers for adoption of this technology.

C. Waveform Description

A sample ring waveform from a swine is displayed in Figure 2. The sensor acts as a high-pass filter due to the input impedance and internal capacitance of the PVDF sensor and data acquisition system [24]. The amplification circuit produces a transfer function between strain and voltage with a zero at the origin and two real poles. The dominant pole-zero dynamics produce a high-pass filter with cutoff frequency higher than the signal frequency, such that it behaves as a differentiator in the frequency range of the cardiac cycles. Hence, the generated signal is approximately the derivative of the pressure waveform with respect to time at the measurement site.

Fig. 2:

A sample waveform generated by the PVDF sensor with clear reflection waves.

The hydraulic version of Ohm’s law can be used to calculate CO,

$$Q=\Delta P/R\approx P_a/R,$$ (1)

where Q denotes the CO, ΔP the gradient between the mean arterial and mean right atrial pressures, and R the vascular resistance, mainly caused by resistance in the small arteries. Since the right atrial pressure is often close to zero, the CO can be approximated as the ratio of mean arterial pressure, P_a, and R. Hence, assessment of CO requires quantification of arterial blood pressure and resistance. It is well-known that the elastic modulus of large arteries is a function of the intra-arterial pressure,

$$E(P)=E_0{e}^{\alpha P},$$ (2)

where E₀ is the elastic modulus at zero arterial pressure and α is an artery dependent constant [25]. Assuming the artery is an elastic tube, the Moens-Kortweg equation [26], [27] relates pulse wave velocity (PWV) to the artery elastic modulus,

$$\text{PWV}=\sqrt{\frac{hE}{2r\rho }},$$ (3)

where h is the artery thickness, r the radius, and ρ the blood density. Using Equations 2 and 3, the classical relationship between arterial pressure and PWV can be derived as

$$P={\alpha }^{-1}\text{log}\left(\frac{2r\rho }{h{E}_0}\right)+2{\alpha }^{-1}\text{log}(\text{PWV}).$$ (4)

The changes in the first term and α are negligible; hence, the arterial pressure is mainly dependent on the PWV. Conventionally, the PWV is derived from pulse transit time measured using two sensors at two different anatomic sites, such as electrocardiography and finger photoplethysmography. However, the pulse transit time can also be estimated by measuring the time delay between the incident pulse wave and its reflection. Hence, we hypothesize that the inverse of the delay between the incident beat and the first reflection wave in the PVDF waveform correlates with arterial blood pressure.

The other component in the right side of Eq. 1 is the vascular resistance, which is mainly influenced by artery radius, r, according to the Hagen—Poiseuille equation,

$$R=\frac{8L\eta }{\pi r^4}$$ (5)

where L is the length of the artery and η is blood viscosity. Arterioles and capillaries are the predominant drivers of systemic vascular resistance (SVR) given their small radius. The arterioles have a layer of smooth muscle which modulates their radius. Since L is constant and the variations in η are negligible compared to variations in r⁴, changes in R are mainly determined by changes in the arteriole radius. The body maintains homeostasis through vasoconstriction (a decrease in r which leads to an increase in R) and vasodilation (an increase in r which leads to a decreases R). By controlling R in Eq. 1, vasoconstriction and vasodilation are responsible for maintaining blood pressure and CO within their normal ranges.

As the pulse wave travels through the distal arteries, arterioles and capillaries, the changes in the radius of the vessels leads to pressure wave reflections. This is due to the change in the characteristic impedance of the vessel at these points, resulting in a backward propagation of the wave. The amplitude of the backward wave is determined by the reflection coefficient, Γ,

$$\Gamma =\frac{Z_T-Z_0}{Z_T+Z_0}$$ (6)

where Z₀ and Z_T are the characteristic impedance of the proximal and terminal (distal) arteries, respectively [28]. Hence, an increase in Z_T caused by vasoconstriction leads to an increase in the amplitude of the reflected wave. Therefore, we hypothesize that the amplitude of the reflection wave, as measured by the PVDF ring, correlates with vascular resistance. As a result, the PVDF waveform contains information regarding both P_a and R that can be used to estimate Q through Eq. 1.

III. Data

The CNN was trained using animal data and validated using patient data. The two datasets are described below.

A. Animal Data for Model Development

The training data for the CNN was collected from 52 animals. The experiments were conducted on swine undergoing various individual protocols to study hemorrhage, cardiac arrest, traumatic brain injury, and sepsis. All procedures were approved by the University of Michigan’s Institutional Animal Care and Use Committee and the principles stated in the eighth edition of the Guide for the Care and Use of Laboratory Animals [29] were followed.

The experiments were conducted on male Yorkshire cross swine, approximately 14 weeks of age with a mean (SD) weight of 40kg (3kg). Animals remained under full general anesthesia using inhaled isoflurane or total intravenous anesthesia (Fentanyl, midazolam, and Propofol) for the duration of the procedure. Following induction of anesthesia, all animals underwent invasive continuous monitoring of arterial pressure and central venous pressure as well as electrocardiography, photoplethysmography, pulse oximetry, end-tidal carbon dioxide, and the experimental PVDF ring. In addition, CO was measured continuously using a pulmonary artery thermodilution catheter and Vigilance II Monitoring System (CCOmbo V, 8F, Edwards Lifesciences, Irvine, CA). Continuous waveform data was collected using a Biopac MP150 data acquisition system and Acqknowledge software (Biopac Inc. Goleta, CA).

As part of the individual protocols, animals were taken systematically through a range of hemodynamic perturbations, modeling states commonly seen in critical illness and injury. They included a combination of arterial hemorrhage (controlled and uncontrolled) producing hemorrhagic shock, fluid and whole blood resuscitation, sudden cardiac arrest and cardiopulmonary resuscitation, traumatic brain injury using a simulated epidural hematoma or direct cortical impact, norepinephrine challenge, and sepsis resulting in septic shock, all of which produce dynamic changes in CO.

B. Patient Data for Model Validation

Between October 2017 and June 2019, the PVDF ring was validated on a convenience sample of 60 human subjects undergoing 65 pulmonary artery catheterizations (PACs) in the heart failure lab at the University of Michigan, excluding subjects who had a left ventricular assist device (LVAD) in place. Data from four subjects was unusable; in two cases this was because of a noisy signal due to a tremor and in the other two cases due to faulty PVDF sensors. Thus, data was available on 56 subjects undergoing 61 PACs. Most were male (78.6%) and white (87.5%) with a mean age of 57.7±14.5 years (Table I). Twenty-four (42.9%) subjects had heart failure (HF) with a reduced ejection fraction. Twenty-five (44.6%) subjects had a history of heart transplantation (HT) of whom 14 (56.0%) had a prior LVAD.

TABLE I:

Clinical characteristics for the 56 unique patients included in the analysis.

Variable	n(%)	Mean (SD)
Demographics
Patient age, years		57.7 (14.5)
Female gender	12 (21.4)
Race
White	59 (87.5)
Black	6 (10.7)
Other	1 (1.8)
Comorbid conditions
History of a left ventricular assist device	14 (25.0)
History of heart transplantation	25 (44.6)
Heart failure, reduced ejection fraction	24 (42.9)
Heart failure, preserved ejection fraction	7 (12.5)
Obesity	29 (51.8)
Diabetes mellitus	23 (41.1)
Hypertension	39 (69.6)
Chronic kidney disease	34 (60.7)

Key: SD = standard deviation.

The majority of subjects (50.8%) underwent a PAC in the setting of HF, both as part of routine clinical care and for pre-operative assessments (Table II). A large minority (42.6%) of subjects underwent a PAC in the setting of routine surveillance following HT. Sixteen (26.2%) subjects had a CI less 2 L/min/m². For 7 patients, resting oxygen consumption was measured using a metabolic cart while in the remaining 52 patients oxygen consumption was estimated.

TABLE II:

Hemodynamics from 61 PACs.

Variable	n(%)	Mean (SD)	Range
Indication for pulmonary artery catheterization
HF	31 (50.8)
Follow-up HT	26 (42.6)
Pulmonary hypertension	2 (3.3)
Shortness of breath	2 (3.3)
Hemodynamics
MAP, mmHg		92.4 (17.2)	56.0–134.0
PCWP, mmHg		18.5 (7.3)	6.0–36.0
Mean PAP, mmHg		28.2 (9.8)	9.0–48.0
RAP, mmHg		11.0 (5.1)	2.0–28.0
Fick CI, L/min/m2		2.4 (0.8)	1.26–6.7
Fick CI < 2 L/min/m2	16 (26.2)
TD CI, L/min/m2		2.6 (0.8)	1.35–6.3
PVR, dyn.s/cm5		167.4 (99.5)	17.0–553.0
SVR, dyn.s/cm5		1408.2 (436.8)	254.3–2515.2

Key: HT=heart transplant; MAP=mean arterial pressure; PCWP=pulmonary capillary wedge pressure; PAP=pulmonary artery pressure; RAP=right atrial pressure; TD=thermodilution; PVR=pulmonary vascular resistance; SVR=systemic vascular resistance.

Similar to the animal data, a Biopac MP150 data acquisition system and Acqknowledge software were used to collect all the waveforms. Informed consent was obtained from all the participants in this study and data collection was conducted under IRB protocol HUM00144242 approved by the University of Michigan.

C. Data Cleansing

The collected waveforms (both animal and patient data) were visually inspected to identify sections that were contaminated by noise and annotated in the Acqknowledge Software prior to data analysis and model validation.

IV. Methods

Both the animal and patient data were divided into 3 second non-overlapping windows. For every window of the animal data, the mean CI was calculated from measurements made by the Edwards continuous CO monitor, which is the gold standard for continuous measurement of CO. For the patients, the CI was measured using the Fick method, which is the gold standard used in clinical practice. The Fick formula is calculated as $Q=\frac{{\text{VO}}_2}{C_a-C_v}$ where VO₂ denotes oxygen consumption and is either directly measured or estimated as 125×body surface area. C_a and C_v correspond to the oxygen content of arterial and mixed venous blood, respectively.

The animal data was used to train CNNs to identify instances of low CI and then validated on patient data. The CNN architecture and training are detailed below.

A. Convolutional Neural Network

The model used in this study is shown in Figure 3. It consisted of a 384×1 (3 seconds at 128Hz sampling rate) input layer, followed by 3 blocks each containing 3 convolutional layers and a maximum pooling layer. The output of the third block was flattened and fed into a fully connected layer with 128 neurons and ReLU activation, followed by another fully connected layer with 1 neuron and sigmoid activation. Each convolutional layer contained 64 filters of size 11×1 with strides of size 1, zero padding and ReLU activation.

Fig. 3:

The architecture of the convolutional neural network. The input is 3 seconds of PVDF signal after down-sampling to 128Hz. The convolutional layers and the filters are shown in orange and red, respectively. The red arrows indicate max pooling. The last two layers are fully connected with ReLU and sigmoid activations.

The network was trained using data from 52 animal experiments, where 36 (70%) randomly selected animals were used for training and the other 16 (30%) for validation. The model was trained using a binary cross-entropy loss function and RMSProp optimizer. A batch size of 128 was used and the training continued for a maximum of 200 epochs. The learning rate was reduced by a factor of 0.1 if the validation accuracy did not improve for 3 consecutive epochs. Training was terminated early if the validation accuracy did not improve after 15 consecutive epochs. The training was repeated 50 times. The models were implemented using Keras with Tensorflow backend in Python 3.6.

B. Model Aggregation

For validation, each of the 50 trained models was applied to ten 3-second non-overlapping noise-free windows closest to the time when the Fick CO was measured. An aggregate model (referred to as median CNN model) was built by computing the median of the class probabilities (output of the CNN model) for the ten PVDF windows and then computing the median across the 50 models. This resulted in a single probability for each patient, indicating whether the patient had a low CI.

Another model was built (referred to as a pruned CNN model) by selecting the models with the least uncertainty in their output, measured by the range of scores obtained from each patient’s ten windows (i.e., a model generating output probabilities with less variability across the ten PVDF windows was preferred over a model with more variable outputs). Twenty models with the lowest range of probabilities were selected for each patient and their outputs were used to calculate the median probability for that patient.

C. Heart Rate and Mean Arterial Pressure Models

Heart rate (HR) and mean arterial pressure (MAP) are two hemodynamic parameters that can be measured non-invasively and directly impact CI. Hence, we compared the results of our two CNN models against two models that use HR and MAP to identify patients with a low CI. As these two models each contain a single variable, the variables were directly used as model scores, after being multiplied by −1 due to the inverse relationship between these variables and low CI; hence, no training was conducted. In addition, we used the combination of HR and MAP (HR+MAP) to classify the patients. To do so, we used the patient data to train a logistic regression model with a binary outcome variable (normal versus low CI).

V. Results

The HR, MAP and HR+MAP models were compared to the median and pruned CNN models in Table III and Figure 4. Note that the logistic regression for the HR+MAP model was directly trained on the patient data and no external validation was conducted for this model. This decision was made to avoid a drop in the performance of the HR+MAP model arising from the differences in the physiology of swines and humans, and to give the HR+MAP model an advantage over the median and pruned CNN models. Nevertheless, all three models based on HR and MAP failed to accurately identify patients with low CI.

TABLE III:

The accuracies for the two CNN models against the HR, MAP and HR+MAP models. The threshold for each model was selected to achieve a sensitivity ≥ 0.8, corresponding to the solid dots on the ROC and PR curves in Figure 4. The numbers in parentheses correspond to the 95% confidence intervals calculated using bootstraps with 10000 replicas.

Model	Sensitivity	Specificity	PPV	NPV	AUROC	AUPRC
Median CNN	0.81 (0.54,0.95)	0.82 (0.68,0.92)	0.62 (0.40,0.82)	0.93 (0.80,0.98)	0.85 (0.71,0.94)	0.65 (0.44,0.85)
Pruned CNN	0.81 (0.53,0.95)	0.89 (0.77,0.96)	0.72 (0.47,0.90)	0.93 (0.81,0.98)	0.88 (0.73,0.95)	0.71 (0.51,0.89)
HR	0.81 (0.54,0.95)	0.24 (0.13,0.39)	0.28 (0.16,0.43)	0.79 (0.50,0.94)	0.57 (0.39,0.73)	0.30 (0.16,0.49)
MAP	0.81 (0.54,0.95)	0.18 (0.09,0.32)	0.26 (0.15,0.40)	0.73 (0.38,1.00)	0.53 (0.35,0.70)	0.27 (0.15,0.44)
HR+MAP	0.81 (0.55,0.95)	0.24 (0.14,0.39)	0.28 (0.16,0.42)	0.79 (0.50,0.98)	0.57 (0.39,0.73)	0.30 (0.16,0.49)

The bold font indicates the best value for each metric.

Fig. 4:

The receiver operating characteristic (ROC) and precision-recall (PR) curves are shown. The dashed lines show the curves for the 50 individual CNN models (mostly overlapping). For the five models shown in solid lines, the points corresponding to a sensitivity of ≥ 0.8 are indicated by solid dots. Note that the dots for the HR and HR+MAP models are overlapping.

For comparisons, we chose a threshold for each model that led to a sensitivity of ≥ 0.8. The results using these thresholds are reported in Table III. The pruned CNN model achieved the highest sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC). These results were followed by the Median CNN model that achieved the second best accuracies.

VI. Discussion

The PVDF signal contains features that are dependent on blood pressure and vascular resistance, as detailed in Section II-C. As described earlier, increased blood pressure results in higher PWV and shortens the time delay between the incidence and reflection waves. This is illustrated in the examples shown in Figure 5. In this example, the blue line corresponds to the highest pressures, leading to the shortest delay for the reflection wave, while the green line corresponds to the lowest pressures and has the longest wave reflection delay.

Fig. 5:

Examples of the PVDF waveform at different levels of blood pressure (BP), CI and SVR. The output probabilities (P) calculated by the pruned CNN model are also indicated. The red dashed lines indicate the location of the first reflection waves. The BPs are expressed as systolic/diastolic/MAP. The measurement units for BP, CI and SVR are mmHg, L/min/m² and dyn·s/cm⁵, respectively.

The amplitude of the reflection wave is determined by the impedance changes in the terminal arteries, which is correlated with vascular resistance. The three examples shown in Figure 5 correspond to vastly different SVRs. The green line comes from a patient with a low SVR and has a diminished reflection wave while the yellow line comes from a patient with a normal SVR and has a more moderate amplitude. The line in blue comes from a patient with a high SVR. In this case, the reflection wave, although less visible due to the overlap with the incidence beat, has the highest amplitude.

Since MAP and SVR are the two factors that affect CI, certain patterns for the reflection waves correspond to different values of CO. The CNN models trained on the animal data learn these patterns and use them to classify patients based on their CI. The results, shown in Figure 4 and Table III, indicate that a PVDF ring and CNN models can be leveraged to achieve this objective.

VII. Conclusions and Future Works

The main objective of this study was to demonstrate that variable signal output from a PVDF ring can be used to distinguish patients based on their CI. To do so, we used a CNN model to classify the PVDF waveforms. The model was trained on waveforms from an animal model and validated using waveforms from patients undergoing PAC procedures. The results demonstrate that the combination of the ring and the CNN model are effective for identifying patients with a low CI. This technology can serve as a tool by which to obtain more rapid and accurate diagnoses for patients presenting with undifferentiated shock, allowing for early delivery of targeted therapies.

The proposed approach needs to be further validated using a larger sample of patients with a range of CIs. Moreover, the correlation between blood pressure and wave reflection delays, as well as the correlation between vascular resistance and the amplitude of the reflection waves need to be further studied and confirmed. Finally, the noisy signals were identified and discarded manually in this study. In the future, this steps needs to be automated. Once approach toward identification and reduction of noise can be based on convolutional autoencoders that are trained on noise-free sections of the animal waveforms. Such a model can be used both for detection of noise using the error between the input and output of the autoencoder model, as well as reduction of noise by utilizing the output of the model.

Biographies

Sardar Ansari received his PhD and MS degrees in Computer Science and MS Degree in Statistics from Virginia Commonwealth University. Dr. Ansari is a data scientist with a special focus on signal processing applications. His research interests are in biomedical Big Data analytics, machine learning, and development of wearable diagnostic devices.

Jessica R. Golbus received her MD and MS from the University of Michigan. She is a general cardiology fellow at the University of Michigan on the Cardiovascular Medicine T32 training grant. Her research focuses on leveraging digital health technology to improve cardiovascular care with a particular interest in patients with advanced heart failure.

Mohamad H. Tiba received his MD from the University of Damascus in 1991. Dr. Tiba is a career researcher in the field of critical care and shock, with research interests spanning non-invasive monitoring techniques and devices to development of innovative resuscitative fluids and hemostatic strategies.

Brendan McCracken is the Assistant Director for MCIRCC’s Preclinical Critical Care Laboratory and specifically focused on the development, management, and execution of experimental protocols for models of critical illness, injury, and care. Areas of interest include hemostasis, coagulation, and noninvasive or advanced hemodynamic and physiologic monitoring.

Lu Wang received his PhD and MS degrees in Mechanical Engineering from the University of Michigan, Ann Arbor. His research focuses on the intersection of dynamics and estimations with small-scale sensing. He is particularly interested in sensing, modeling and dynamics, with current research focused on applications to cardiovascular system.

Keith D. Aaronson received his MD from Baylor College of Medicine, Houston, Texas, and MS in Clinical Research Design and Statistical Analysis from the University of Michigan. He is the Co-Director of both the Section of Heart Failure and Transplantation and the Heart Failure and Transplantation Inpatient Service at the University of Michigan.

Kevin R. Ward received his MD from Tulane University. Dr. Ward serves as the Executive Director for the Michigan Center for Integrative Research in Critical Care at the University of Michigan. His research interests are in developing platform technologies that improve the care of critically ill and injured patients.

Kayvan Najarian received his BSc in Electrical Engineering from Sharif University, Tehran, MSc in Biomedical Engineering from Amirkabir University, Tehran, and PhD in electrical and computer engineering from the University of British Columbia, Vancouver. Dr. Najarian’s research focuses on the design of signal/image processing and machine learning methods to improve patient care.

Kenn R. Oldham received the PhD and BS in Mechanical Engineering from UC Berkeley and Carnegie Mellon University, respectively. His research focuses on the intersection of control systems and micro-scale sensing and actuation, with interests in design for controllability, optimal and robust control, microsystem estimation and identification, and novel sensor and actuator design.