Towards an Automatic Recognition of Artifacts and Features in Plethysmographic Traces

Abstract

A plethysmograph is a device that quantitatively assesses volumetric variations in an organ or the entire body, typically resulting from fluctuations in blood flow. In this study, a strain-gauge sensor that measures changes in the volume of the neck was used to detect the the cerebral venous outflow in the internal jugular veins. The resulting electronic signal was susceptible to several external factors, complicating the identification of relevant features. A reliable analysis of the waveform, without the need for a manual intervention to analyze the data, is of paramount importance to provide real-time analysis of the vital parameters of the patient. In this work, we demonstrate that specifically designed neural networks can detect artifacts in plethysmographic traces and identify the most important features in the signal with reasonable accuracy, eliminating the need to perform these tasks manually for each patient.

1. Introduction

The determination of the venous outflow from the brain serves as an important metric in the diagnosis of cardiovascular diseases [1]. For instance, brain damage in a patient with diagnosed heart failure is investigated through the jugular venous pulse (JVP) waveform [2], a quantity related to the venous outflow within the internal jugular veins (IJVs) [2].

Currently, the most common method for conducting this assessment entails the utilization of ultrasound technology performed by a qualified operator [2]. However, this approach presents a notable challenge in terms of automation and does not provide the possibility to perform continuous monitoring. An alternative method for the continuous monitoring of the JVP is central venous line catheterization, but this approach requires medical procedures, sometimes including surgery.

Plethysmography is a non-invasive technique that measures the variation in the volume of a certain body district, typically in the wrist, the ankle, or the neck [3,4,5]. A plethysmograph provides complementary information with respect to other non-invasive techniques (e.g., ECG); however, its signal is more complex to analyze because of the dependence on the position and the biological variables of each patient [6,7,8]. Since the IJVs are the main vessels for blood return from the brain to the heart and travel through the neck right under the skin, they have been proven to be one of the most advantageous sites for the use of a plethysmograph [4].

The IJVs are collapsible vessels whose cross-sectional area (CSA) depends on the transmural pressure on the vessel wall [9]. The pressure is also influenced by the position of the patient: In an upright position, the IJVs collapse, significantly reducing their CSA. When the patient is supine, the IJVs are the vessels with the largest blood outflow from the brain [6]. The variation in the venous outflow during a cardiac cycle yields a measurable variation in the IJVs’ CSA, which can be detected through a small variation in the volume of the neck [10].

The JVP pattern consists of three ascents or positive waves (a, c, and v) and three descents or negative waves (x, $x^{\prime }$, and y) [1]:

a 

: originates from active atrial contraction leading to retrograde blood flow into neck veins.

x 

: caused by continued atrial relaxation.

c 

: due to the impact of the carotid artery adjacent to the jugular vein and retrograde transmission of a positive wave in the right atrium, produced by the right ventricular systole and the bulging of the tricuspid valve into the right atrium.

$x^{\prime }$ 

: caused by the descent of the right atrium floor (tricuspid valve) during right ventricular systole and continued atrial relaxation.

v 

: corresponds to the maximal atrial filling; it is less prominent than the a ascent wave.

y 

: follows the v wave and corresponds to the emptying of the atrium.

These features are schematically shown in Figure 1.

A plethysmographic trace exhibits significant variability in its features, largely influenced by the individual characteristics of each patient and their position during measurement (e.g., supine or sitting) [10]. As a result, the data typically require visual inspection and analysis. Artifacts caused by patient movements must be manually identified and removed from the records, while any visible waves require manual annotation by an expert.

In this work, we demonstrate that the application of supervised machine learning techniques, and specifically neural networks, can significantly simplify this task by automatically recognizing artifacts and by identifying the most easily recognizable waves with a reasonable accuracy and specificity.

2. Related Works

The present study aims to explore a non-invasive methodology for monitoring the jugular venous pulse signal. In recent years, numerous research groups have investigated the feasibility of using wearable devices to detect the characteristic peaks of the JVP waveform.

Inertial sensors, such as accelerometers and gyroscopes, have been employed for this purpose due to their correlation with neck movements induced by venous pulse activity. Nguyen et al. [11] used an accelerometer-based sensing system for the preliminary identification of the JVP signal by connecting the peaks of interest from the measured accelerometric trace. Similarly, George et al. [12] achieved a reliable JVP waveform through the double integration of the accelerometric signal. In contrast, Karhinoja et al. [13] extracted the JVP signal by integrating the vector sum of angular velocity measurements obtained from a three-axis gyroscope. Additionally, a contact piezoelectric sensor was employed to measure the JVP signal, and the peak locations were determined using pulse contour markers [14].

Another contact-based system consists of single-crystal ultrasound coupled to a force-sensing load cell to measure the JVP based on the force required to collapse the IJV walls [15]. Furthermore, Hill et al. [16] used a non-invasive optical sensing system to estimate the venous oxygen saturation of both the external and internal jugular veins, while García-López and Rodriguez-Villegas [17] proposed an anterior neck contact photoplethysmography system to extract information on the JVP.

Regarding non-contact methodologies for JVP waveform analysis, researchers are validating the effectiveness of using microwave radar-based systems in both near- and far-field regions [18,19,20]. Custom photoplethysmography imaging systems, consisting of a near-infrared sensitive camera paired with an illumination source, have been used to quantify JVP waveforms based on skin displacements, revealing a strong negative correlation with arterial waveforms [21,22]. Moreover, camera-based systems that acquire video recordings of neck skin movement are being developed to facilitate remote, non-contact JVP measurements for telehealth applications [23,24].

To date, research has predominantly focused on applying artificial intelligence techniques to arterial signals, which are easily detected by photoplethysmography (PPG) [25,26]. Arterial blood pressure can be measured by commercially available wearable devices, where machine learning frameworks are widely used to accurately identify characteristic points of the PPG waveform, not only for cardiovascular applications [27] but also for the classification and prediction of depressive behavior [28], among others [29].

In contrast, our research is dedicated to the automated analysis of the jugular venous pulse instead of the arterial pulse with machine learning techniques, which represents a research topic that is currently largely unexplored. Conversely to the studies presented in refs. [11,13,17,19], the method presented in this article removes artifacts and identifies JVP peaks of interest in an automatic manner, while most studies on JVP signal analysis so far manually annotate the peaks of the main waves for each cardiac cycle [30].

In this work, the JVP signal is acquired via an easy-to-use, wearable device that enables patients to autonomously monitor their health in an uncontrolled environment, such as their own home. Therefore, it is crucial to implement an algorithm capable of automatically eliminating signal artifacts and providing the quasi-real-time identification of JVP peaks, such as c and y, which correspond to the bulging of the tricuspid valve into the right atrium and the rapid filling phase of the right ventricle after the tricuspid valve opens, respectively [31]. These are useful for the assessment of blood pulsation on the brain–heart axis. Future developments will aim to extend the number and the accuracy of features recognized.

3. Data Collection and Preparation

3.1. A Summary of the Capacitive Strain-Gauge Sensor and Experimental Activity

The device considered in this study consists of a wearable, non-invasive plethysmographic sensor that is externally wrapped around the neck of the patient [3,4]. The device contains a highly stretchable central zone with a capacitive strain-gauge sensor [32]. The capacitive strain-gauge sensor is formed by a single-layer dielectric electro-active polymer (DEAP), PolyPowerfilm^®, sandwiched between two deformable electrodes [33]. It results in a sensing element in the form of a stretchable capacitor (ElastiSense Sensor Technology, formerly LEAP technology, Aabenraa, Denmark). The capacitance of the sensor increases when the device is stretched, allowing it to detect minimal deformations. The sensor adheres to the skin and records the perimeter variation in the examined anatomic part. The device, applied to a patient, is depicted in Figure 2.

Integrated in the wearable device, an electronic unit collects data from the strain-gauge sensor and from three electrodes, providing a complementary and independent ECG trace. The plethysmographic and ECG signals are then transferred via a Bluetooth^® connection to a PC for analysis and storage.

A detailed description of the device used in this study can be found in ref. [4,32]. The characterization of the capacitive strain-gauge and ECG sensors are also discussed in ref. [4]. The device is derived from the one developed in the context of the Drain Brain project [10,32,34].

3.2. Data Collection

The data sample consisted of the combined plethysmographic and ECG signals of twenty healthy subjects, seven of which were females, with an average age between 25 and 35 years. The calculated body mass index (BMI) for all the participants was in the range between 19 and 23 [3]. This dataset, consisting of physiological measurements, was collected by researchers at the Center for Exercise Science and Sports, University of Ferrara, as part of the study published in [3]. The study received approval from the Ethics Committee, with the corresponding approval number and date documented therein.

After positioning the sensor around the neck of the patient and placing the ECG electrodes in their designated positions, the patient was allowed to assume the required posture (supine, sitting, and upright). Data acquisition was started only half a minute later to minimize artifacts and allow the patient to relax. For each subject, three traces of the JVP waveform were recorded, one for each of the three postures. Each sample recorded the timestamp, the plethysmographic signal, and the ECG signals, captured at a sampling frequency of 66 Hz and lasting approximately 30–35 s. The raw plethysmographic signals were processed with a Wavelet filter [35] to remove low-frequency artifacts caused by respiratory movements and to enhance the cardiac contribution. Instead, the ECG signal was sufficiently clear and did not require any filtering.

Figure 3 shows an example of signals collected from a supine patient during a 5 s interval. The QRS complex from the ECG trace can be visually correlated with the peaks of the plethysmographic trace. As in Figure 1, the c peak is generally more prominent than the a peak, with the former being easier to identify. This difference is attributed to the pressure exerted by the carotid artery.

The ECG signal serves only as a visual reference to validate the plethysmographic trace. As this study aimed to conduct a comprehensive and independent analysis of the plethysmographic signal, the ECG trace was not utilized in the subsequent steps.

3.3. Annotations

The data collected were complemented with the metadata required for the supervised training of machine learning algorithms. These metadata included the timestamp indicating the location of each peak classified in Section 1. Because of the signal variability, not all peaks are consistently visible or recognizable: while JVP waveforms are clearly identifiable when the patient is in a supine position, the carotid artery’s contribution dominates in the IJV in sitting and upright positions, obscuring some features of the JVP waveform [4]. For these reasons, the current analysis focused on identifying and labeling the c and y peaks, which typically correspond to the global maximum and minimum of the recorded signal in each cycle, respectively. The labeled peaks were then visually inspected and validated by team members.

4. Detection of Artifacts

The identification and localization of artifacts in the plethysmgraphic trace represent the primary challenge in JVP waveform analysis. Both involuntary and voluntary muscle movements can alter the volume of the neck, introducing a potential nuisance in the waveform signal. These artifacts may range from being barely detectable to large enough to obscure the JVP signal.

4.1. Model Description

This study considered a model based on Long Short-Term Memory [36] networks, a class of recurrent neural networks (RNNs) [37] designed to effectively capture temporal patterns in sequential data, such as plethysmographic traces. The LSTM architecture preserves long-term dependencies while filtering out irrelevant variations, making it particularly well suited for detecting transient anomalies within the signal. The network model was implemented using Keras [38], with TensorFlow [39] as the backend framework.

In order to capture temporal dependencies, the network model incorporates two bidirectional LSTM layers, which are designed to process sequential data with a length of 50 timesteps, learning both past and future dependencies within the sequence. The first bidirectional LSTM layer consists of 64 units in each direction, producing an output of size 128 that encodes information from both the forward and backward temporal contexts. The second bidirectional LSTM layer processes this sequence further, reducing the dimensionality to 64 units and providing a compact temporal representation.

The output of the second LSTM layer is fed into a series of dense layers, forming a Multi-Layer Perceptron [40]. An intermediate dense layer reshapes the sequence representation back to its original length of 50 timesteps, with one unit per timestep. Subsequently, three additional dense layers, each with one output unit, generate the predicted values. The entire architecture consists of 78,923 trainable parameters among the two LSTM layers and the dense layers.

Table 1. A summary of the LSTM network architecture, showing the layer type, output shape, and number of parameters.

Layer Type	Output Shape	Number of Parameters
Input	50, 2	-
Bidirectional (1)	50, 128	34,304
Bidirectional (2)	64	41,216
Dense	50	3250
Dense	1	51
Dense	1	51
Output	1	51
Total parameters		78,923

summarizes the number and type of layers, the output size, and the number of trainable parameters for each layer in the model.

For each sequence of 50 timesteps, the model provides three values derived from a standard quantile regression. These values represent the prediction for the next point in the sequence (corresponding to the 50th quantile), as well as the 10th and the 90th quantiles of the distribution modeling the output.

The model was trained from a sample of seven patients selected for their minimal noise and absence of artifacts in the JVP waveform. The Adam optimizer [41] was used to minimize the loss function. Once trained, the model was then applied to other patients who shared the same body position (supine or sitting) as those in the training dataset. In the evaluation phase, the probability distribution function of the output was expected to exhibit a narrow peak centered close to the predicted value, while in presence of an artifact the function tended to be wider.

4.2. Results

The separation between the 10th and 90th quantiles (${\Delta }_{quant}$) provides an indication of the accuracy of the prediction. Additionally, the residual difference between the predicted (50th quantile) and the measured value ${\Delta }_{pred}$ indicates how large the deviation in a specific timestep is from the expectation, as shown in Figure 4. A threshold, set to half of the maximum of ${\Delta }_{comb}={\Delta }_{pred}+{\Delta }_{quant}$, was applied to classify data points exceeding this value as artifacts. Figure 5 illustrates an example where the model identified artifacts in a compromised plethysmographic trace. By using three distinct indicators—the quantile separation, the prediction residual, and their combined effect—the exclusion criteria could be adjusted for various scenarios. Identified artifact intervals can be excluded from subsequent analyses to ensure data integrity.

Future improvements in the network and in the training sample could enable its application to quasi-real-time data, providing immediate feedback to the operator acquiring the plethysmographic trace. The system could notify the operator about the validity of the data if one or more artifacts are detected during acquisition. Since the alert can be issued either during or immediately after data collection, the operator has the opportunity to repeat the measurement if necessary. This approach contrasts with the current standard, where analyses are often conducted hours or days after data acquisition. Such delays preclude the possibility of repeating measurements in cases where the artifact count is excessively high, as the patient may no longer be available.

5. Feature Identification

The extraction of clinically relevant information from the plethysmographic trace required identifying the peaks in the JVP waveform. The number of recognizable peaks was largely dependent on both the position of the patient and the level of noise in the sample, as these factors could obscure some of the peaks described in Section 1.

5.1. Model Description

The model chosen to address the challenge of locating and identifying the peaks was a convolutional neural network (CNN) [42]. A CNN consists of a set of convolutional filters designed to detect characteristic waveform patterns, including peak shapes, slopes, and amplitude variations. This architecture is particularly well-suited for peak identification, as it remains robust with variations in the timing and amplitude of peaks within the signal.

The CNN convolutional filters consisted of one-dimensional (1D) segments containing 15 timesteps of the plethysmographic trace. The network model consisted of three 1D convolutional layers, followed by a flattening layer that converted the output of the convolutional layers into a 1D vector. This vector was fed to a dense fully connected layer, which produced another 1D vector for each input sample. The final dense layer output a vector three times the size of the input segment, assigning the probability of each timestep belonging to one of these three categories: a c peak, a y peak, or neither.

The kernel size [43] for the convolutional layers was set to eight timesteps, which was optimal to capturing local patterns in the signal. No additional padding [44] was applied to the input, resulting in reduced spatial dimensions after each convolution operation. A unitary stride [44] ensured that the filter moved one element at a time, thoroughly exploring the input space without omitting any region. The number of filters in the convolutional layers, the kernel sizes, the dropout rates, and the number of units in the fully connected layer were selected based on model’s performance on the validation set, ensuring that the chosen hyperparameters provided the best balance between efficiency and purity. The model was implemented in Keras [38], based on TensorFlow [39], and is summarized in

Table 2. A summary of the convolutional neural network architecture, showing the layer type, output shape, and number of parameters.

Layer Type	Output Shape	Number of Parameters
Input	15, 1	-
Convolutional 1D (1)	8, 32	288
Convolutional 1D (2)	5, 16	2064
Convolutional 1D (3)	4, 8	264
Flatten	32	-
Dense	256	8448
Output	45	11,565
Total parameters		22,629

.

The model was trained, validated, and tested on a homogeneous sample of supine patients that did not produce experimental noise or artifacts during recording the data. In total, seven patients fulfilled these requirements. The seven samples were split into the training group (six patients, corresponding to approximately 85% of the available dataset), while the one remaining patient (accounting for 15% of the dataset) was considered a validation sample to evaluate the performance of the model. During the training of the model, it was verified that shuffling the training and validation datasets did not change in a significant way the performance of the model.

Because of the limited amount of data available for training, the performance of the model was enhanced using the following technique. Since almost every data point was fed to the model multiple times following the creation of 1D segments, the predictions for each time step were repeated N times. The final classification was obtained by averaging these multiple predictions and applying a threshold of $1/N$, which classified as peaks the timesteps with at least one c or y peak classification.

This procedure followed the “map-reduce” principle [45,46], which consists of processing datasets by dividing tasks into parallel “map” functions, which process and transform data, and a “reduce” function, which aggregates the results to produce the final output. This computing paradigm inherits various desirable properties, such as scalability, which is crucial for real-life applications where the model has to deal with long time series signals. Additionally, it provides a tunable window, allowing adaptation to different frequencies in the data acquisition process, and facilitates a live implementation, enabling real-time analysis of the JVP trace during patient monitoring.

5.2. Results

In the test sample, which was statistically uncorrelated with the training and validation sets, the model predictions were compared against the annotations described in Section 3.3, which were considered the reference to evaluate the model. The numbers of c and y peaks identified in the testing sample are reported in

Table 3. The number of peaks in the testing patient, reported separately for each peak type (c or y) of the JVP waveform. The bottom rows report the efficiency and the purity with the corresponding statistical uncertainties.

	c Peak	y Peak
True positives	51	47
True predicted positives	59	54
Total positives	54	55
Total predicted positives	64	61
Efficiency	$0.{94}_{-0.18}^{+0.06}$	$0.{85}_{-0.17}^{+0.15}$
Purity	$0.{92}_{-0.17}^{+0.08}$	$0.{89}_{-0.16}^{+0.11}$

. Figure 6 represents an example of a plethysmographic trace where both the annotations and the model predictions are shown and compared to each other.

The performance of the model was evaluated in terms of efficiency, also known as recall, and purity, sometimes referred to as precision. Efficiency is defined as the ratio of true positives, i.e., the number of times where a peak was correctly identified, to the total number of positives in the testing sample, reflecting the completeness of the prediction. Purity is the ratio of true predicted positives to the total predicted positives, indicating the accuracy of the predictions. These figures of merit were chosen since there was no cross-contamination between categories, e.g., no c peak was labeled as a y peak and vice versa.

The model achieved an efficiency of $0.{94}_{-0.18}^{+0.06}$ and $0.{85}_{-0.17}^{+0.15}$ for the the c and y peaks, respectively. The purity was $0.{92}_{-0.17}^{+0.08}$ for the c peaks and $0.{89}_{-0.16}^{+0.11}$ for the y peaks. These results, along with their uncertainties, are summarized in Table 3. The uncertainties are purely statistical and reflect the limited number of peaks in the testing sample.

The difference in the performance between the two peaks was primarily due to the characteristics of the JVP waveform. The y peaks were often closely spaced and may have overlapped with other features, making them more challenging to detect. In contrast, c peaks were typically more isolated, simplifying their identification. Additionally, the performance of the model was constrained by the limited size and variability of the training dataset. Expanding the dataset to include a larger number of patients in diverse positions would likely reduce the disparity between the accuracy of c and y peak classifications, improve overall accuracy, and decrease statistical uncertainty. Future work will focus on addressing these limitations to enhance the robustness and generalizability of the model.

6. Conclusions

In this study, we demonstrated the feasibility of automatically recognizing artifacts and key features in the jugular venous pulse (JVP) waveform, significantly reducing the manual workload required for inspection and labeling. A model, based on LSTM, was developed to isolate trace segments containing artifacts, i.e., variations in the shape that do not come from the variation in the cross-sectional area of the internal jugular veins. This model enables the automated validation of data quality and exclusion of regions affected by artifacts, eliminating the need for manual analysis.

Additionally, a dedicated convolutional neural network (CNN) was designed to identify the most clinically relevant features in the waveform, specifically the c and y peaks. An efficiency of $0.{94}_{-0.18}^{+0.06}$ and $0.{85}_{-0.17}^{+0.15}$ and a purity of $0.{92}_{-0.17}^{+0.08}$ and $0.{89}_{-0.16}^{+0.11}$ could be achieved for the c and y peaks, respectively. The analyses were conducted using only the plethysmographic signal, without incorporating concurrent signals like the ECG signal.

These results highlight the potential for expanding the approach. With a larger, more diverse dataset and higher sampling frequencies, the models could be extended to automatically recognize additional features and determine the posture of the patient (sitting or supine). Incorporating a more extensive and heterogeneous dataset, including patients with specific pathologies and individuals over 50 years old, would enable the model to be trained for the identification of pathological indicators across a wide spectrum of ages, sexes, and body positions, thereby enhancing its applicability and clinical relevance.