STUDY OBJECTIVES:Sleep staging is usually performed by manual scoring of polysomnography (PSG), which is expensive, laborious, and poorly scalable. We propose an alternative to PSG for ambulatory sleep staging using wearable photoplethysmography (PPG) recorded by a smartwatch and automated scoring. METHODS:We previously trained a deep learning model on public datasets, with the specific purpose of performance generalizability to unseen datasets. In the present work, the model was assessed on two datasets of reflective PPG collected from wrist-worn devices: (1) 68 overnight recordings and (2) for the first time, 493 long-term recordings each lasting for 24 hours (170 subjects). Findings were compared either to (1) expert scored sleep stages from PSG for the night recordings or (2) actigraphy for the long-term recordings. RESULTS:For the overnight recordings, the PPG-based model achieved 78.7% accuracy and a Cohen's κ of 0.68 on reflective PPG collected using wrist-worn devices compared to PSG using a 4-class setup (wake, N1, and N2 combined, N3 and REM), and a sleep/wake accuracy of 94.1%, with a Cohen's κ of 0.71. For the long-term recordings, a sleep/wake accuracy of 92.5% with a Cohen's κ of 0.80 was achieved when compared to a state-of-the-art actigraphy-based deep learning model. CONCLUSIONS:This state-of-the-art accuracy achieved on wrist-worn devices represents a significant advancement for home sleep monitoring and a valuable alternative to PSG-based sleep staging. Additionally, our model demonstrated promising results on long-term ambulatory recordings, paving the way towards continuous ambulatory monitoring of sleep stages and sleep-wake cycles. Statement of Significance Sleep staging is crucial to diagnose sleep disorders, but traditional methods are laborious and costly. We developed a sleep staging model that demonstrates high performance and exceptional generalization to unseen datasets, including those from wrist-worn devices, thereby possibly enabling accurate sleep staging from wearable technology. Furthermore, we evaluated the model's performance on 24-hour recordings of subjects of various health conditions, offering valuable insights for clinical applications and future research. These advancements significantly enhance the feasibility of continuous sleep monitoring at home, a low-cost, scalable, and comfortable alternative to current methods.

Sleep apnea is a common chronic sleep-related disorder which is known to be a comorbidity for cerebro- and cardio-vascular disease. Diagnosis of sleep apnea usually requires an overnight polysomnography at the sleep laboratory. In this paper, we used a wearable device which measures reflectance photoplethysmography (PPG) at the wrist and upper arm to estimate continuous SpO2 levels during sleep and subsequently derive an oxygen desaturation index (ODI) for each patient. On a cohort of 170 patients undergoing sleep apnea screening, we evaluated whether this ODI value could represent a surrogate marker for the apnea-hypopnea index (AHI) for the diagnosis and severity assessment of sleep apnea. As the ODI was simultaneously obtained at the fingertip, upper arm and wrist, we compared ODI diagnostic performance depending on the measurement location. We then further evaluated the accuracy of ODI as a direct predictor for moderate and severe sleep apnea as defined by established AHI thresholds. We found that ODI values obtained at the upper arm were good predictors for moderate or severe sleep apnea, with 86% accuracy, 96% sensitivity and 70% specificity, whereas ODI values obtained at the wrist were less reliable as a diagnostic tool.

Introduction Obstructive sleep apnea syndrome (OSAS) is a prevalent sleep disorder associated with significant morbidity and mortality, particularly due to its links with cardiovascular diseases like hypertension (HT). Continuous positive airway pressure (CPAP) remains the standard treatment for OSAS, yet individualized therapy and monitoring are crucial for optimizing patient outcomes. This study explores the feasibility of utilizing connected devices to remotely monitor OSAS patients undergoing CPAP treatment. Methods Ten patients diagnosed with OSAS were enrolled in a prospective observational feasibility study. Participants wore two wearables continuously: CenterPoint Insight Watch ™ for sleep and physical activity monitoring, and Aktiia™ bracelet for blood pressure measurement. CPAP usage data were collected using the DreamStation™ device. Data synchronization and processing were conducted using a dedicated Python script. Primary outcomes included acceptability, compliance, autonomy in device usage, and data quality. Secondary outcomes focused on the feasibility of integrating a centralized platform for analysis. Results Acceptability among patients was reasonable, with 58% consenting to participate. However, two patients discontinued the study due to skin allergies and device interference with professional activities. Most participants demonstrated autonomy in using the devices, although two required assistance with synchronization. Data quality varied, particularly with nocturnal blood pressure measurements, affected by technical issues and individual factors. Integration of data from all devices onto a centralized platform was feasible, enabling comprehensive analysis. Discussion The study highlighted successes in continuous remote monitoring of OSAS patients undergoing CPAP treatment. Challenges included device-related issues and manual data processing. A centralized platform for data integration and analysis proved promising for longitudinal monitoring and personalized healthcare delivery. Conclusion This feasibility study demonstrates the potential of remote monitoring in CPAP-treated OSAS patients. Future efforts should focus on addressing technical challenges and optimizing data integration on a common platform to realize the full benefits of continuous monitoring in personalized healthcare management.

Traditionally used for measuring heart rate and oxygen saturation, photoplethysmography (PPG) has emerged as a promising non-invasive alternative for diagnosing sleep related disorders. Unlike the gold-standard polysomnography (PSG) performed in-lab at the hospital, PPG offers a more scalable and cost-effective solution. Recent advancements in deep learning have significantly enhanced the precision of these methods for sleep stage inference. This study extends the evaluation of a PPG-based deep learning model to a clinical cohort of 134 patients with suspected sleep apnea (SA). These participants, enrolled in an ongoing clinical trial, underwent overnight PSG alongside simultaneous recording of PPG and accelerometer signals using CSEM’s wearable devices, positioned at both the wrist and upper arm. When compared to PSG, the PPG-based deep learning model achieved a median accuracy of 80.8% with a Cohen's Kappa of 0.7 in identifying wakefulness, light sleep (S1 + S2), deep sleep (S3), and rapid eye movement (REM) sleep stages using wrist-worn sensors. A reduction in performance was observed when the device was worn at the upper arm, with accuracy decreasing by approximately 6.2% and Cohen’s Kappa by 10%. Additionally, a lightweight alternative of the model leveraging inter-beat-intervals (IBIs) yielded comparable results at the wrist, with no performance degradation at the upper arm, highlighting its potential for deployment in resource-constrained settings. Overall, these findings demonstrate the feasibility of the approach as an accessible complement to PSG for home-based sleep monitoring.

The combination of electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) at ultra-high field (7 Tesla) offers appealing new possibilities to probe human brain function non-invasively with high coverage, millisecond temporal precision and sub-millimeter spatial precision, unraveling cortical layers and small subcortical structures. Unfortunately, this technique has remained largely inaccessible at 7T, due to prohibitive cross-modal interference effects and physical constraints. Here, we developed a first-of-its-kind EEG-fMRI acquisition framework on a clinical 7T system combining key improvements from previous research: compact EEG transmission to reduce artifact incidence, reference sensors for artifact correction, and adapted leads for compatibility with a dense radiofrequency receive array allowing state-of-the-art fMRI sensitivity and acceleration. Two implementations were tested: one using an EEG cap adapted in-house, and another using a recently designed prototype from an industrial manufacturer, intended to be further developed into a commercial device accessible to the broader community. A comprehensive evaluation in humans showed that simultaneous acquisitions, including sub-millimeter fMRI resolution, could be conducted without detectable safety issues or major practical constraints. The EEG exerted relatively mild perturbations on fMRI quality (6-11% loss in temporal SNR), without measurably affecting the detection of resting-state networks and visual responses. The artifacts induced on EEG could be corrected to a degree where the spatial, spectral, and temporal characteristics were comparable with outside recordings, and hallmark features such as resting-state alpha and eyes-closing alpha modulation could be clearly detected. Altogether, these findings indicate excellent prospects for neuroimaging applications, which can leverage the unique possibilities achievable at 7T.

Wearable electroencephalography (EEG) sensors are becoming increasingly accessible, enabling long-term, at-home automatic sleep profiling, with outstanding potential for personalized medicine and wellness applications. However, current devices remain less reliable for sleep stage classification than gold-standard, lab-based polysomnography (PSG). One of the main bottlenecks is the need to collect extensive datasets for each new wearable device together with a gold standard reference, to train robust algorithms. To tackle this challenge, here, we investigated machine learning and feature engineering techniques on a recent PSG dataset containing multi-channel EEG and reference sleep stage labels, to train classifiers tailored for sleep staging on a separate, wearable EEG headband (ULTEEMNite). We found that the classification performance improved when: (i) training on EEG configurations spatially closer to that of the wearable device, (ii) filtering the training data to approximate the spectral profile of the wearable, and (iii) including information from neighboring signal periods. We also identified the most relevant signal features for classification, across different stages and sensor configurations. These findings enable developing more effective algorithms for sleep staging with wearable EEG, to pursue its vast clinical potential.Clinical Relevance— Sleep stage disruptions are a meaningful indication in diverse pathologies, and increasingly recognized as an early sign of neurodegenerative diseases such as dementia.

Sleep apnea (SA) is a chronic sleep-related disorder consisting of repetitive pauses or restrictions in airflow during sleep and is known to be a risk factor for cerebro- and cardiovascular disease. It is generally diagnosed using polysomnography (PSG) recorded overnight in an in-lab setting at the hospital. This includes the measurement of blood oxygen saturation (SpO2), which exhibits fluctuations caused by SA events. In this paper, we investigate the accuracy and utility of reflectance pulse oximetry from a wearable device as a means to continuously monitor SpO2 during sleep. To this end, we analyzed data from a cohort of 134 patients with suspected SA undergoing overnight PSG and wearing the watch-like device at two measurement locations (upper arm and wrist). Our data show that standard requirements for pulse oximetry measurements are met at both measurement locations, with an accuracy (root mean squared error) of 1.9% at the upper arm and 3.2% at the wrist. With a rejection rate of 3.1%, the upper arm yielded better results in terms of data quality when compared to the wrist location which had 30.4% of data rejected.Clinical Relevance- Our findings confirm the feasibility of wearables for unobtrusive continuous monitoring of SpO2, especially in patients with suspected SA.

The diagnosis of sleep disorders is still often based on polysomnography, an in-lab exam allowing experts to perform accurate sleep staging, although this is labor-intensive, expensive, and exposing patients to unusual sleep conditions. A state-of-the-art deep learning model - called SleepPPGNet - was recently proposed. It achieves an accuracy of 82% and a Cohen's kappa of 0.74 on a completely new dataset through transfer learning, using only raw fingertip photoplethysmography (PPG) as input, paving the way toward more efficient sleep staging methods.We applied this model to PPG data collected with our own wrist-worn devices in adults and reached 78% accuracy and a Cohen's kappa of 0.68. This is encouraging in the prospect of patients collecting their own data at home. In addition, we built upon the model's architecture to include activity counts as additional input. This increased global accuracy from 78.5% to 80.0% and Cohen's kappa from 0.67 to 0.69 on our main dataset. Finally, although this model has demonstrated remarkable potential on subjects with normal cardiac rhythms, it has shown limitations when applied to patients with cardiac arrhythmia, with an accuracy drop of 10% compared to a control group.

Artificial intelligence (AI) is gaining increasing interest in the field of medicine because of its capacity to process big data and pattern recognition. Cardiotocography (CTG) is widely used for the assessment of foetal well-being and uterine contractions during pregnancy and labour. It is characterised by inter- and intraobserver variability in interpretation, which depends on the observers’ experience. Artificial intelligence (AI)-assisted interpretation could improve its quality and, thus, intrapartal care. Cardiotocography (CTG) raw signals from labouring women were extracted from the database at the University Hospital of Bern between 2006 and 2019. Later, they were matched with the corresponding foetal outcomes, namely arterial umbilical cord pH and 5-min APGAR score. Excluded were deliveries where data were incomplete, as well as multiple births. Clinical data were grouped regarding foetal pH and APGAR score at 5 min after delivery. Physiological foetal pH was defined as 7.15 and above, and a 5-min APGAR score was considered physiologic when reaching ≥7. With these groups, the algorithm was trained to predict foetal hypoxia. Raw data from 19,399 CTG recordings could be exported. This was accomplished by manually searching the patient’s identification numbers (PIDs) and extracting the corresponding raw data from each episode. For some patients, only one episode per pregnancy could be found, whereas for others, up to ten episodes were available. Initially, 3400 corresponding clinical outcomes were found for the 19,399 CTGs (17.52%). Due to the small size, this dataset was rejected, and a new search strategy was elaborated. After further matching and curation, 6141 (31.65%) paired data samples could be extracted (cardiotocography raw data and corresponding maternal and foetal outcomes). Of these, half will be used to train artificial intelligence (AI) algorithms, whereas the other half will be used for analysis of efficacy. Complete data could only be found for one-third of the available population. Yet, to our knowledge, this is the most exhaustive and second-largest cardiotocography database worldwide, which can be used for computer analysis and programming. A further enrichment of the database is planned.

Biomedical data generation and collection have become faster and more ubiquitous. Consequently, datasets are increasingly spread across hospitals, research institutions, or other entities. Exploiting such distributed datasets simultaneously can be beneficial; in particular, classification using machine learning models such as decision trees is becoming increasingly common and important. However, given that biomedical data is highly sensitive, sharing data records across entities or centralizing them in one location are often prohibited due to privacy concerns or regulations. We design PrivaTree, an efficient and privacy-preserving protocol for collaborative training of decision tree models on distributed, horizontally partitioned, biomedical datasets. Although decision tree models may not always be as accurate as neural networks, they have better interpretability and are helpful in decision-making processes, which are crucial for biomedical applications. PrivaTree follows a federated learning approach, where raw data is not shared, and where every data provider computes updates to a global decision tree model being trained, on their private dataset. This is followed by privacy-preserving aggregation of these updates using additive secret-sharing, in order to collaboratively update the model. We implement PrivaTree, and evaluate its computational and communication efficiency on three different biomedical datasets, as well as the accuracy of the resulting models. Compared to the model centrally trained on all data records, the obtained collaborative model presents a modest loss of accuracy, while consistently outperforming the accuracy of the local models, trained separately by each data provider. Moreover, PrivaTree is more efficient than existing solutions, which makes it usable for training decision trees with numerous nodes, on large complex datasets, with both continuous and categorical attributes, as often found in the biomedical field.

Objective.Cardiac arrhythmias are a leading cause of mortality worldwide. Wearable devices based on photoplethysmography give the opportunity to screen large populations, hence allowing for an earlier detection of pathological rhythms that might reduce the risks of complications and medical costs. While most of beat detection algorithms have been evaluated on normal sinus rhythm or atrial fibrillation recordings, the performance of these algorithms in patients with other cardiac arrhythmias, such as ventricular tachycardia or bigeminy, remain unknown to date.Approach. ThePPG-beatsopen-source framework, developed by Charlton and colleagues, evaluates the performance of the beat detectors namedQPPG,MSPTDandABDamong others. We applied thePPG-beatsframework on two newly acquired datasets, one containing seven different types of cardiac arrhythmia in hospital settings, and another dataset including two cardiac arrhythmias in ambulatory settings.Main Results. In a clinical setting, theQPPGbeat detector performed best on atrial fibrillation (with a medianF1score of 94.4%), atrial flutter (95.2%), atrial tachycardia (87.0%), sinus rhythm (97.7%), ventricular tachycardia (83.9%) and was ranked 2nd for bigeminy (75.7%) behindABDdetector (76.1%). In an ambulatory setting, theMSPTDbeat detector performed best on normal sinus rhythm (94.6%), and theQPPGdetector on atrial fibrillation (91.6%) and bigeminy (80.0%).Significance. Overall, the PPG beat detectorsQPPG,MSPTDandABDconsistently achieved higher performances than other detectors. However, the detection of beats from wrist-PPG signals is compromised in presence of bigeminy or ventricular tachycardia.

The combination of electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) at ultra-high field (7 Tesla) offers unprecedented possibilities to probe human brain function non-invasively with high coverage, millisecond temporal precision and sub-millimeter spatial precision, unraveling cortical layers and small subcortical structures. Unfortunately, this technique has remained largely inaccessible at 7T, due to prohibitive cross-modal interference effects and physical constraints. Here, we developed a first-of-its-kind EEG-fMRI acquisition framework on a clinical 7T system combining key improvements from previous works: compact EEG transmission chain to reduce artifact incidence, reference sensors for artifact correction, and adapted leads for compatibility with a dense radiofrequency receive-array allowing state-of-the-art fMRI sensitivity and acceleration. Two implementations were tested: one using an EEG cap adapted in-house, and another using a recently-designed prototype from an industrial manufacturer, intended to be further developed into a commercial product accessible to the broader community. A comprehensive evaluation in humans showed that simultaneous acquisitions, including with sub-millimeter fMRI resolution, could be conducted without detectable safety issues or major practical constraints. The EEG exerted relatively mild perturbations on fMRI quality (6-11% loss in temporal SNR), without measurably affecting the detection of resting-state networks and visual responses. The artifacts induced on EEG could be corrected to a degree where the spatial, spectral and temporal characteristics were comparable to outside recordings, and hallmark features such as resting-state alpha and eyes-closing alpha modulation could be clearly detected. Altogether, these findings indicate excellent prospects for neuroimaging applications, that can leverage the unique possibilities achievable at 7T. ### Competing Interest Statement Authors Cilia Jager and Tracy Warbrick are employed by Brain Products GmbH. The other authors have no conflicts of interest to disclose.

Wearable devices based on photoplethysmography (PPG) allow for the screening of large populations at risk of cardiovascular disease. While PPG has shown the ability to discriminate atrial fibrillation (AF) - the most common cardiac arrhythmia (CA) - versus normal sinus rhythm, it is not clear whether such AF detectors are efficient in presence of CAs other than AF. We propose to apply a simple recurrent neural network (RNN) on a newly acquired dataset containing eight different types of CAs. The classifier takes sequences of inter-beat intervals (IBIs) as input and discriminates between normal and abnormal rhythm. The RNN achieved 84% accuracy in detecting abnormal rhythms. Some CAs were well detected (AF: 99.6%; atrial tachycardia: 100%), whereas other CAs were more difficult to detect (atrial flutter: 65.4%; bigeminy: 72.4%; ventricular tachycardia 80%). This study shows the potential of PPG technology to detect not only AF but also other types of CA. It highlights the strengths and weaknesses of IBI-based detection of abnormal rhythms and paves the way towards continuous monitoring of CAs in everyday life.

Cardiac arrhythmias affect millions of individuals worldwide and can lead to severe complications such as stroke or heart failure. They can be difficult to diagnose with ambulatory electrocardiogram monitors due to their transient nature. We propose a system for long-term ar-rhythmia monitoring that takes single-lead electrocardio-gram and tri-axis acceleration signals as inputs. It is composed of a beat detector to extract interbeat intervals and a classifier to detect arrhythmias. This system is evalu-ated on two datasets including 42 patients and achieves an accuracy of 0.988 for the abnormal class, 0.967 for the normal class, and 0.979 for the tachycardia class.

The use of 24-h ambulatory blood pressure monitoring (ABPM) has been continuously increasing over the last decades. However, cuff-based devices may cause discomfort, particularly at night, leading to potentially non-representative blood pressure (BP) values. We investigated the feasibility of a cuff-less BP monitoring solution in 67 subjects undergoing conventional 24-h ABPM. A watch-like optical sensor was attached at the upper arm or wrist at the contralateral side of the cuff. Systolic (SBP) and diastolic BP (DBP) values were estimated from the measured optical signals by pulse wave analysis. Average 24-h, daytime and nighttime BP values were compared between the conventional monitor and the cuff-less sensor. The differences between both methods—expressed as mean ± standard deviation (95% limits of agreement)—were of − 1.8 ± 6.2 mmHg (− 13.9, 10.3) on SBP and − 2.3 ± 5.4 mmHg (− 13.0, 8.3) on DBP for 24-h averages, of − 1.5 ± 6.6 mmHg (− 14.4, 11.4) on SBP and − 1.8 ± 5.9 mmHg (− 13.4, 9.9) on DBP for daytime averages, and of 0.4 ± 7.5 mmHg (− 14.4, 15.1) on SBP and − 1.3 ± 6.8 mmHg (− 14.7, 12.0) on DBP for nighttime averages. These results encouragingly suggest that cuff-less 24-h ABPM may soon become a clinical possibility.

Cardiac arrhythmias present a significant global health concern. The advent of wearable devices utilizing photoplethysmography gives the opportunity to screen large populations, hence offering the potential for early detection of pathological rhythms and reducing risks of complications and associated medical costs. While most beat detection algorithms have been evaluated on normal sinus rhythm or atrial fibrillation recordings, their performance in patients with other cardiac arrhythmias remains unexplored to date. To address this gap, we leveraged the open-source framework PPG-beats, developed by Charlton and colleagues, to analyse a newly acquired dataset comprising seven distinct types of cardiac arrhythmia in hospital settings. Among the thirteen beat detectors evaluated, the QPPG detector performed best on atrial fibrillation (with a median F1 score of 94.4%), atrial flutter (95.2%), atrial tachycardia (87.0%), sinus rhythm (97.7%), ventricular tachycardia (83.9%) and was ranked second for bigeminy (75.7%) behind the ABD detector (76.1%). Overall, the QPPG beat detector achieved high performances and consistently outperformed other detectors. However, the detection of beats from wrist-PPG signals is compromised in the presence of bigeminy or ventricular tachycardia.

Blood pressure (BP) is a crucial biomarker giving valuable information regarding cardiovascular diseases but requires accurate continuous monitoring to maximize its value. In the effort of developing non-invasive, non-occlusive and continuous BP monitoring devices, photoplethysmography (PPG) has recently gained interest. Researchers have attempted to estimate BP based on the analysis of PPG waveform morphology, with promising results, yet often validated on a small number of subjects with moderate BP variations. This work presents an accurate BP estimator based on PPG morphology features. The method first uses a clinically-validated algorithm (oBPM®) to perform signal preprocessing and extraction of physiological features. A subset of features that best reflects BP changes is automatically identified by Lasso regression, and a feature relevance analysis is conducted. Three machine learning (ML) methods are then investigated to translate this subset of features into systolic BP (SBP) and diastolic BP (DBP) estimates; namely Lasso regression, support vector regression and Gaussian process regression. The accuracy of absolute BP estimates and trending ability are evaluated. Such an approach considerably improves the performance for SBP estimation over previous oBPM® technology, with a reduction in the standard deviation of the error of over 20%. Furthermore, rapid BP changes assessed by the PPG-based approach demonstrates concordance rate over 99% with the invasive reference. Altogether, the results confirm that PPG morphology features can be combined with ML methods to accurately track BP variations generated during anesthesia induction. They also reinforce the importance of adding a calibration measure to obtain an absolute BP estimate.

Smartphones may provide a highly available access to simplified hypertension screening in environments with limited health care resources. Most studies involving smartphone blood pressure (BP) apps have focused on validation in static conditions without taking into account intraindividual BP variations. We report here the first experimental evidence of smartphone-derived BP estimation compared to an arterial catheter in a highly dynamic context such as induction of general anesthesia. We tested a smartphone app (OptiBP) on 121 patients requiring general anesthesia and invasive BP monitoring. For each patient, ten 1-min segments aligned in time with ten smartphone recordings were extracted from the continuous invasive BP. A total of 1152 recordings from 119 patients were analyzed. After exclusion of 2 subjects and rejection of 565 recordings due to BP estimation not generated by the app, we retained 565 recordings from 109 patients (acceptance rate 51.1%). Concordance rate (CR) and angular CR demonstrated values of more than 90% for systolic (SBP), diastolic (DBP) and mean (MBP) BP. Error grid analysis showed that 98% of measurement pairs were in no- or low-risk zones for SBP and MBP, of which more than 89% in the no-risk zone. Evaluation of accuracy and precision [bias ± standard deviation (95% limits of agreement)] between the app and the invasive BP was 0.0 ± 7.5 mmHg [− 14.9, 14.8], 0.1 ± 2.9 mmHg [− 5.5, 5.7], and 0.1 ± 4.2 mmHg [− 8.3, 8.4] for SBP, DBP and MBP respectively. To the best of our knowledge, this is the first time a smartphone app was compared to an invasive BP reference. Its trending ability was investigated in highly dynamic conditions, demonstrating high concordance and accuracy. Our study could lead the way for mobile devices to leverage the measurement of BP and management of hypertension.

Photoplethysmography $(PPG)$ is one of the most promising alternatives for non-invasive and cuffless blood pressure $(BP)$ monitoring. In recent years, several machine learning approaches have been considered for this task: either feature engineering-based to map PPG-derived features into $BP$ values or feature learning-based with an automated feature extraction process. The generalization capability, namely the ability of a model to adapt to unseen data, is an important aspect of such data-driven models due to the heterogeneity of PPG waveforms. However, in published studies, this point is generally omitted. Therefore, we propose to assess the generalization capability of a feature learning model built to estimate $BP$ from $PPG$ signals by comparing the model accuracy (bias) and precision (standard deviation of the error) on two datasets with different recording protocols. On unseen subjects from the training dataset, the proposed model achieved mean and standard deviation errors of - $0.88\pm 10.29$ mmHg for systolic $BP(SBP)and-0.60\pm 5.76$ mmHg for diastolic $BP(DBP)$. Whereas, on the other dataset, the same metrics were $1.13\pm 12.86$ mmHg for $SBP$ and $-0.44\pm 6.99$ mmHg for $DBP$ Taken together, these results show that a feature learning model can extract feature representation that are generalizable over different populations and different $PPG$ sensors.