The Application of Deep Learning Algorithms for PPG Signal Processing and Classification

Vital Sign and Biochemical Data Collection Using Non-contact Photoplethysmography and the Comestai Mobile Health App: Protocol for an Observational Study

BACKGROUND

Early detection of vital sign changes is key to recognizing patient deterioration promptly, enabling timely interventions and potentially preventing adverse outcomes.

OBJECTIVE

In this study, vital parameters (heart rate, respiratory rate, oxygen saturation, and blood pressure) will be measured using the Comestai app to confirm the accuracy of photoplethysmography methods compared to standard clinical practice devices, analyzing a large and diverse population. In addition, the app will facilitate big data collection to enhance the algorithm's performance in measuring hemoglobin, glycated hemoglobin, and total cholesterol.

METHODS

A total of 3000 participants will be consecutively enrolled to achieve the objectives of this study. In all patients, personal data, medical condition, and treatment overview will be recorded. The "by face" method for remote photoplethysmography vital sign data collection involves recording participants' faces using the front camera of a mobile device (iOS or Android) for approximately 1.5 minutes. Simultaneously, vital signs will be continuously collected for about 1.5 minutes using the reference devices alongside data collected via the Comestai app; biochemical results will also be recorded. The accuracy of the app measurements compared to the reference devices and standard tests will be assessed for all parameters. CIs will be calculated using the bootstrap method. The proposed approach's effectiveness will be evaluated using various quality criteria, including the mean error, SD, mean absolute error, root mean square error, and mean absolute percentage error. The correlation between measurements obtained using the app and reference devices and standard tests will be evaluated using the Pearson correlation coefficient. Agreement between pairs of measurements (app vs reference devices and standard tests) will be represented using Bland-Altman plots. Sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and likelihood ratios will be calculated to determine the ability of the new app to accurately measure vital signs.

RESULTS

Data collection began in June 2024. As of March 25, 2025, we have recruited 1200 participants. The outcomes of the study are expected at the end of 2025. The analysis plan involves verifying and validating the parameters collected from mobile devices via the app, reference devices, and prescheduled blood tests, along with patient demographic data.

CONCLUSIONS

Our study will enhance and support the accuracy of data on vital sign detection through PPG, also introducing measurements of biochemical risk indicators. The evaluation of a large population will allow for continuous improvement in the performance and accuracy of artificial intelligence algorithms, reducing errors. Expanding research on mobile health solutions like Comestai can support preventive care by validating their effectiveness as screening tools and guiding future health care technology developments.

TRIAL REGISTRATION

ClinicalTrials.gov NCT06427564; https://clinicaltrials.gov/study/NCT06427564.

Synthetic photoplethysmogram (PPG) signal generation using a genetic programming-based generative model

Nowadays, photoplethysmograph (PPG) technology is being used more often in smart devices and mobile phones due to advancements in information and communication technology in the health field, particularly in monitoring cardiac activities. Developing generative models to generate synthetic PPG signals requires overcoming challenges like data diversity and limited data available for training deep learning models. This paper proposes a generative model by adopting a genetic programming (GP) approach to generate increasingly diversified and accurate data using an initial PPG signal sample. Unlike conventional regression, the GP approach automatically determines the structure and combinations of a mathematical model. Given that mean square error (MSE) of 0.0001, root mean square error (RMSE) of 0.01, and correlation coefficient of 0.999, the proposed approach outperformed other approaches and proved effective in terms of efficiency and applicability in resource-constrained environments.

A Guide to Measuring Heart and Respiratory Rates Based on Off-the-Shelf Photoplethysmographic Hardware and Open-Source Software

The remote monitoring of vital signs via wearable devices holds significant potential for alleviating the strain on hospital resources and elder-care facilities. Among the various techniques available, photoplethysmography stands out as particularly promising for assessing vital signs such as heart rate, respiratory rate, oxygen saturation, and blood pressure. Despite the efficacy of this method, many commercially available wearables, bearing Conformité Européenne marks and the approval of the Food and Drug Administration, are often integrated within proprietary, closed data ecosystems and are very expensive. In an effort to democratize access to affordable wearable devices, our research endeavored to develop an open-source photoplethysmographic sensor utilizing off-the-shelf hardware and open-source software components. The primary aim of this investigation was to ascertain whether the combination of off-the-shelf hardware components and open-source software yielded vital-sign measurements (specifically heart rate and respiratory rate) comparable to those obtained from more expensive, commercially endorsed medical devices. Conducted as a prospective, single-center study, the research involved the assessment of fifteen participants for three minutes in four distinct positions, supine, seated, standing, and walking in place. The sensor consisted of four PulseSensors measuring photoplethysmographic signals with green light in reflection mode. Subsequent signal processing utilized various open-source Python packages. The heart rate assessment involved the comparison of three distinct methodologies, while the respiratory rate analysis entailed the evaluation of fifteen different algorithmic combinations. For one-minute average heart rates' determination, the Neurokit process pipeline achieved the best results in a seated position with a Spearman's coefficient of 0.9 and a mean difference of 0.59 BPM. For the respiratory rate, the combined utilization of Neurokit and Charlton algorithms yielded the most favorable outcomes with a Spearman's coefficient of 0.82 and a mean difference of 1.90 BrPM. This research found that off-the-shelf components are able to produce comparable results for heart and respiratory rates to those of commercial and approved medical wearables.

DeepVAQ : an adaptive deep learning for prediction of vascular access quality in hemodialysis patients

BACKGROUND

Chronic kidney disease is a prevalent global health issue, particularly in advanced stages requiring dialysis. Vascular access (VA) quality is crucial for the well-being of hemodialysis (HD) patients, ensuring optimal blood transfer through a dialyzer machine. The ultrasound dilution technique (UDT) is used as the gold standard for assessing VA quality; however, its limited availability due to high costs impedes its widespread adoption. We aimed to develop a novel deep learning model specifically designed to predict VA quality from Photoplethysmography (PPG) sensors.

METHODS

Clinical data were retrospectively gathered from 398 HD patients, spanning from February 2021 to February 2022. The DeepVAQ model leverages a convolutional neural network (CNN) to process PPG sensor data, pinpointing specific frequencies and patterns that are indicative of VA quality. Meticulous training and fine-tuning were applied to ensure the model's accuracy and reliability. Validation of the DeepVAQ model was carried out against established diagnostic standards using key performance metrics, including accuracy, specificity, precision, F-score, and area under the receiver operating characteristic curve (AUC).

RESULT

DeepVAQ demonstrated superior performance, achieving an accuracy of 0.9213 and a specificity of 0.9614. Its precision and F-score stood at 0.8762 and 0.8364, respectively, with an AUC of 0.8605. In contrast, traditional models like Decision Tree, Naive Bayes, and kNN demonstrated significantly lower performance across these metrics. This comparison underscores DeepVAQ's enhanced capability in accurately predicting VA quality compared to existing methodologies.

CONCLUSION

Exemplifying the potential of artificial intelligence in healthcare, particularly in the realm of deep learning, DeepVAQ represents a significant advancement in non-invasive diagnostics. Its precise multi-class classification ability for VA quality in hemodialysis patients holds substantial promise for improving patient outcomes, potentially leading to a reduction in mortality rates.

Physiological sensor data cleaning with autoencoders

Objective.Physiological sensor data (e.g. photoplethysmograph) is important for remotely monitoring patients' vital signals, but is often affected by measurement noise. Existing feature-based models for signal cleaning can be limited as they might not capture the full signal characteristics.Approach.In this work we present a deep learning framework for sensor signal cleaning based on dilated convolutions which capture the coarse- and fine-grained structure in order to classify whether a signal is noisy or clean. However, since obtaining annotated physiological data is costly and time-consuming we propose an autoencoder-based semi-supervised model which is able to learn a representation of the sensor signal characteristics, also adding an element of interpretability.Main results.Our proposed models are over 8% more accurate than existing feature-based approaches with half the false positive/negative rates. Finally, we show that with careful tuning (that can be improved further), the semi-supervised model outperforms supervised approaches suggesting that incorporating the large amounts of available unlabeled data can be advantageous for achieving high accuracy (over 90%) and minimizing the false positive/negative rates.Significance.Our approach enables us to reliably separate clean from noisy physiological sensor signal that can pave the development of reliable features and eventually support decisions regarding drug efficacy in clinical trials.

Blood pressure estimation and classification using a reference signal-less photoplethysmography signal: a deep learning framework

The markers that help to predict th function of a cardiovascular system are hemodynamic parameters like blood pressure (BP), stroke volume, heart rate, and cardiac output. Continuous analysis of hemodynamic parameters such as BP can detect abnormalities earlier, preventing cardiovascular diseases (CVDs). However, sometimes due to motion artifacts, it becomes difficult to monitor the BP accurately and classify it. This work presents an optimized deep learning model having the capability to estimate the systolic blood pressure (SBP) and diastolic blood pressure (DBP) and classify the BP stages simultaneously from the same network using only a single channel photoplethysmography (PPG) signal. The proposed model is designed by exploiting the deep learning framework of a convolutional neural network (CNN), exhibiting the inherent ability to extract features automatically. Moreover, the proposed framework utilizes the superlet transform method to transform a 1-D PPG signal into a 2-D super-resolution time-frequency (TF) spectrogram. A superlet transform separates the peaks related to true PPG signal components and motion artifacts components. Thus, the superlet provides a robust realtime approach to accurately estimating and classifying BP using a single PPG sensor signal and does not require additional ECG and PPG sensor signals for reference. Using a super-resolution spectrogram and CNN model makes the method profitable in motion artifact removal, feature selection, and extraction. Hence the proposed framework becomes less complex for deployment on wearable devices having limited battery resources. The performance of the proposed framework is demonstrated on the publicly available larger dataset MIMIC-III. This work obtained a mean absolute error (MAE) of 2.71 mmHg and 2.42 mmHg for SBP and DBP, respectively. The classification accuracy for the SBP prediction is about 96.79%, whereas it is 98.94% for DBP. From a motion artifact-affected PPG signal, SBP and DBP are estimated. Then the estimated BP is classified into three categories: normotension, prehypertension, and hypertension, and is compared with the state of art methods to show the effectiveness of the proposed optimized framework.

Photoplethysmography-Based Distance Estimation for True Wireless Stereo

Recently, supplying healthcare services with wearable devices has been investigated. To realize this for true wireless stereo (TWS), which has limited resources (e.g. space, power consumption, and area), implementing multiple functions with one sensor simultaneously is required. The Photoplethysmography (PPG) sensor is a representative healthcare sensor that measures repeated data according to the heart rate. However, since the PPG data are biological, they are influenced by motion artifact and subject characteristics. Hence, noise reduction is needed for PPG data. In this paper, we propose the distance estimation algorithm for PPG signals of TWS. For distance estimation, we designed a waveform adjustment (WA) filter that minimizes noise while maintaining the relationship between before and after data, a lightweight deep learning model called MobileNet, and a PPG monitoring testbed. The number of criteria for distance estimation was set to three. In order to verify the proposed algorithm, we compared several metrics with other filters and AI models. The highest accuracy, precision, recall, and f1 score of the proposed algorithm were 92.5%, 92.6%, 92.8%, and 0.927, respectively, when the signal length was 15. Experimental results of other algorithms showed higher metrics than the proposed algorithm in some cases, but the proposed model showed the fastest inference time.

A learning-based image processing approach for pulse wave velocity estimation using spectrogram from peripheral pulse wave signals: An in silico study

Carotid-to-femoral pulse wave velocity (cf-PWV) is considered a critical index to evaluate arterial stiffness. For this reason, estimating Carotid-to-femoral pulse wave velocity (cf-PWV) is essential for diagnosing and analyzing different cardiovascular diseases. Despite its broader adoption in the clinical routine, the measurement process of carotid-to-femoral pulse wave velocity is considered a demanding task for clinicians and patients making it prone to inaccuracies and errors in the estimation. A smart non-invasive, and peripheral measurement of carotid-to-femoral pulse wave velocity could overcome the challenges of the classical assessment process and improve the quality of patient care. This paper proposes a novel methodology for the carotid-to-femoral pulse wave velocity estimation based on the use of the spectrogram representation from single non-invasive peripheral pulse wave signals [photoplethysmography (PPG) or blood pressure (BP)]. This methodology was tested using three feature extraction methods based on the semi-classical signal analysis (SCSA) method, the Law's mask for texture energy extraction, and the central statistical moments. Finally, each feature method was fed into different machine learning models for the carotid-to-femoral pulse wave velocity estimation. The proposed methodology obtained an $R²\geq0.90$ for all the peripheral signals for the noise-free case using the MLP model, and for the different noise levels added to the original signal, the SCSA-based features with the MLP model presented an $R²\geq0.91$ for all the peripheral signals at the level of noise. These results provide evidence of the capacity of spectrogram representation for efficiently assessing the carotid-to-femoral pulse wave velocity estimation using different feature methods. Future work will be done toward testing the proposed methodology for in-vivo signals.

On the Feasibility of Real-Time HRV Estimation Using Overly Noisy PPG Signals

Heart Rate Variability (HRV) is a biomarker that can be obtained non-invasively from the electrocardiogram (ECG) or the photoplethysmogram (PPG) fiducial points. However, the accuracy of HRV can be compromised by the presence of artifacts. In the herein presented work, a Simulink® model with a deep learning component was studied for overly noisy PPG signals. A subset with these noisy signals was selected for this study, with the purpose of testing a real-time machine learning based HRV estimation system in substandard artifact-ridden signals. Home-based and wearable HRV systems are prone to dealing with higher contaminated signals, given the less controlled environment where the acquisitions take place, namely daily activity movements. This was the motivation behind this work. The results for overly noisy signals show that the real-time PPG-based HRV estimation system produced RMSE and Pearson correlation coefficient mean and standard deviation of 0.178 ± 0.138 s and 0.401 ± 0.255, respectively. This RMSE value is roughly one order of magnitude above the closest comparative results for which the real-time system was also used.

Peak Detection and HRV Feature Evaluation on ECG and PPG Signals

Heart Rate Variability (HRV) evaluates the autonomic nervous system regulation and can be used as a monitoring tool in conditions such as cardiovascular diseases, neuropathies and sleep staging. It can be extracted from the electrocardiogram (ECG) and the photoplethysmogram (PPG) signals. Typically, the HRV is obtained from the ECG processing. Being the PPG sensor widely used in clinical setups for physiological parameters monitoring such as blood oxygenation and ventilatory rate, the question arises regarding the PPG adequacy for HRV extraction. There is not a consensus regarding the PPG being able to replace the ECG in the HRV estimation. This work aims to be a contribution to this research area by comparing the HRV estimation obtained from simultaneously acquired ECG and PPG signals from forty subjects. A peak detection method is herein introduced based on the Hilbert transform: Hilbert Double Envelope Method (HDEM). Two other peak detector methods were also evaluated: Pan-Tompkins and Wavelet-based. HRV parameters for time, frequency and the non-linear domain were calculated for each algorithm and the Pearson correlation, T-test and RMSE were evaluated. The HDEM algorithm showed the best overall results with a sensitivity of 99.07% and 99.45% for the ECG and the PPG signals, respectively. For this algorithm, a high correlation and no significant differences were found between HRV features and the gold standard, for the ECG and PPG signals. The results show that the PPG is a suitable alternative to the ECG for HRV feature extraction.