This retrospective analysis used previously collected BCG signals and reference HR data from multiple research studies, including those in which HR evaluation was either a primary or secondary objective.

For studies in which HR evaluation was not a primary or secondary focus, as well as for the overall retrospective analysis, ethical approval (ECG024/2024) was obtained from the Genebandhu Independent Ethics Committee, New Delhi, India, to ensure adherence to established ethical guidelines. Patient confidentiality was rigorously maintained through the anonymization of all data in compliance with data protection regulations. Furthermore, informed consent was obtained from all participants during the original data collection, upholding ethical standards and safeguarding participants’ rights. No compensation was provided to participants.

The dataset includes BCG signals and reference HR measurements gathered from a range of clinical and nonclinical environments. In these studies, BCG and reference HR measurements were captured simultaneously, with HR data obtained from clinical-grade ECG devices. Reference HR values from clinical-grade ECG devices served as an accurate benchmark for evaluating the algorithm’s performance. BCG data were gathered using the Dozee contactless monitoring system. Data collection covered a range of conditions, including rest, movement, and sleep. All reference ECG-derived HR values underwent quality screening to remove physiologically implausible readings or artifact-induced errors. Spurious values (eg, HR<20 or >250 bpm, or sudden changes >50% within 1 measurement interval) were flagged and excluded. By filtering out these outliers and any HR readings noted as affected by clinical interventions, we ensured that the reference standard was reliable for algorithm training and validation.

For the algorithm’s development and validation, the dataset was segmented into 3 subsets—training, tuning, and testing—reflecting different stages of the algorithm’s development. These subsets encompass a diverse participant pool, ensuring representation across various age groups, sex, and health conditions.

Notably, the training or tuning and testing sets were defined a priori by separate clinical researchers, and the data scientists developing the model were blinded to the allocation. The training or tuning and testing partitions were selected to ensure broad representation across HR ranges, comorbidities, demographics, and participant characteristics. This strict separation and blinding process followed recommended practices, eliminating any risk of data leakage between training and validation cohorts and ensuring an unbiased performance evaluation.

The training and tuning datasets were drawn from 10 studies conducted at 6 locations across India (Table 1), covering both clinical and home settings to support the comprehensive development of the algorithm across diverse environments and patient demographics. The training dataset included data from 373 participants, totaling 129,976 data points (Table 2). This dataset was used to train a CNN to recognize HR patterns within BCG signals, covering an HR range of 36 to 230 bpmTable 2 and Figure 2A) and a varied range of comorbidities (Figure 3A). This extensive range supports the algorithm’s robustness across varied cardiac activities and demographics.

The tuning dataset included 75,970 data points from 192 participants Table 2 and was used to refine the algorithm’s parameters for optimal performance. This subset facilitated adjustments to the CNN’s architecture and hyperparameters, enhancing the algorithm’s ability to generalize beyond the training data. The HR range within this tuning dataset, 46 to 169 bpm (Table 2 and Figure 2B), reflects typical cardiac activity in the general population, and the inclusion of a varied range of comorbidities (Figure 3B) further strengthens the algorithm’s applicability across diverse user groups.

The testing dataset was sourced from 5 independent clinical studies conducted at various locations (Table 3), entirely separate from the training and tuning data to ensure unbiased evaluation and minimize the risk of overfitting. This dataset included data from 205 participants and 70,211 data points, covering an HR range of 40 to 169 bpm (Table 4 and Figure 4) and a range of comorbidities, each with at least 3000 data points (Figure 5). The diversity in participant demographics, including different age groups, sex, and health conditions, provided a robust foundation for evaluating the algorithm’s performance across a wide range of patient populations and clinical environments.

The HR measurement algorithm was designed to extract fundamental and harmonic frequency components from the BCG signal, combining signal processing techniques with a CNN for robust HR detection under diverse conditions. Figure 6 illustrates the algorithm’s process on a sample segment: the raw BCG waveform and its filtered version, alongside the short-term Fourier transform (STFT) spectrogram. In these examples, the algorithm correctly identifies an HR of approximately 39 bpm and approximately 182 bpm from BCG signals, closely matching the reference ECG. The HR measurement algorithm includes the following processes:

The performance of the HR measurement algorithm was assessed on the testing dataset by comparing measured HR values to reference values using the following metrics to evaluate accuracy and agreement:

These metrics were chosen to facilitate comparison with prior studies. MAE, Bland-Altman limits of agreement, correlation coefficients, and detection rates are standard benchmarks for validating unobtrusive HR estimation algorithms [28-34]. Using this common set of metrics allows us to contextualize our results against existing BCG-based and other noncontact HR monitoring approaches.

These evaluation metrics provide a comprehensive analysis of the algorithm’s accuracy, reliability, and applicability in real-world clinical settings, reinforcing its potential as a contactless solution for HR monitoring.

The HR measurement algorithm demonstrated high accuracy and reliability across datasets from studies conducted in both the United States and India. In total, the algorithm processed 70,211 reference points and produced 58,908 measured points, resulting in an MAE of 2.50 bpm and a DR of 83.90%.

In the US-based study 1, the algorithm achieved an MAE of 2.27 bpm and a DR of 92.34%. The algorithm’s performance remained robust in the India-based studies (studies 2‐5), with a combined MAE of 2.52 bpm and a DR of 83.10%. Specifically, study 2 reported an MAE of 2.73 bpm and a DR of 81.40%, while study 3 achieved an MAE of 2.49 bpm and a DR of 81.12%. Study 4 achieved an MAE of 2.39 bpm and a DR of 84.52%. Notably, study 5 demonstrated the best performance, with an MAE of 1.78 bpm and a DR of 97.26%.

These findings highlight the algorithm’s consistency across diverse geographic locations and clinical environments, maintaining an MAE below 3 bpm and a DR exceeding 80% across all datasets. Table 5 provides a detailed summary of the performance metrics from the five studies, offering a comprehensive overview of the algorithm’s accuracy and detection capabilities.

Figure 7 shows the HR trend for an ICU patient, comparing HR values measured by the CNN-based algorithm with clinical-grade ECG reference HR measurements over several hours. The figure illustrates that the HR measurement algorithm closely aligns with the ECG reference values, even as the patient’s HR fluctuates within a broad range of 50 to 150 bpm. The visual alignment between the Dozee HR measurements and the ECG HR reference across varying HR levels highlights the algorithm’s precision and responsiveness in a real-world clinical setting. This result demonstrates the algorithm’s capability to maintain accurate HR tracking in the high-noise ICU environment, underscoring its potential as a viable, contactless solution for real-time HR monitoring in health care settings.

A demographic analysis of the algorithm’s performance showed stable accuracy across a variety of subgroups, including different HR ranges, age groups, sex, BMI categories, racial backgrounds, and comorbidities. Table 6 provides a summary of the algorithm’s performance across these demographic variables.

The demographic subgroups are as follows:

These findings confirm the algorithm’s adaptability to diverse population profiles, underscoring its potential as a reliable HR monitoring tool across broad demographic groups in clinical settings.

The Bland-Altman analysis, shown in Figure 8, demonstrates the algorithm’s high accuracy and reliability. The analysis revealed a minimal bias, with a point estimate of 0.25, indicating close alignment with reference HR values. The limits of agreement (LOA), ranging from 8.59 to −8.08, indicate that most differences between the measured and reference values are within 9 bpm, affirming the algorithm’s precision. Notably, 2551 (4.33%) points fall outside the LOA, which represents a small proportion of the total dataset and underscores the algorithm’s strong overall performance. These narrow LOA boundaries, combined with the low bias value, collectively highlight the algorithm’s strong agreement with reference data, confirming its reliability for HR estimation in clinical environments where accuracy is essential.

The Deming regression analysis, complemented by the Pearson correlation coefficient, provided further evidence of the algorithm’s alignment with reference HR values. Figure 9 illustrates the linear relationship between the algorithm’s HR measurements and the clinical reference values. With a Pearson correlation coefficient of 0.97, the algorithm demonstrated a strong correlation with the reference HR data, reinforcing its accuracy and reliability. Notably, 3150 (5.35%) points fall outside the ±7 bpm region, further highlighting the precision of the algorithm’s measurements. The high correlation, together with minimal observed bias, underscores the algorithm’s potential for consistent and accurate HR monitoring in diverse clinical settings. These findings validate the algorithm’s performance, confirming its suitability for contactless patient monitoring applications that require reliable HR data.

This study introduces a highly accurate, contactless HR measurement algorithm based on BCG signals, specifically designed for continuous, noninvasive monitoring in clinical settings. By combining adaptive signal processing techniques with ML-based peak detection, the algorithm effectively addresses challenges in BCG monitoring, such as noise, motion artifacts, and physiological variability. Unlike traditional methods that rely on fixed filtering and thresholding, this algorithm dynamically suppresses noise while preserving HR-related features, ensuring reliable performance under nonstationary conditions caused by patient movement and physiological fluctuations. The algorithm achieved an MAE within 3 bpm and a DR above 80% across diverse settings, demonstrating reliable HR monitoring despite motion artifacts and physiological variability.

A significant strength of this algorithm is its adaptability across demographic subgroups, including age, sex, BMI, and race. This versatility enhances its applicability in dynamic, real-world environments. The CNN architecture represents a substantial advancement over conventional threshold-based methods, which can obscure true cardiac peaks in noisy settings. By maintaining independent datasets for training, tuning, and testing, we minimized the risk of overfitting, ensuring the algorithm generalizes well across different environments. This careful design enables the algorithm to consistently deliver reliable HR measurements, making it suitable for hospital wards and potential applications in home and remote monitoring.

To our knowledge, this is the first BCG-based HR monitoring study to combine time-domain and frequency-domain processing with a deep learning model and to validate it across such a broad range of conditions. Our CNN-based algorithm demonstrates notable improvements in accuracy, robustness, and adaptability compared to existing BCG-based HR measurement solutions. With an MAE of 3 bpm, our algorithm surpasses the accuracy of some commercial contactless solutions, such as EarlySense, which reports an MAE of 5 bpm, setting a new benchmark in contactless HR monitoring. Few studies report comparable MAE values. For example, Vesterinen et al [28] assessed the Emfit QS, a commercial contact-free BCG device, achieving an MAE of 2.4 bpm but with a higher rate of missing data compared to ECG. Similarly, Clemente et al [29] used a bed-mounted sensor system with an MAE of 2.41 bpm, while Proll et al [30] achieved an average MAE of 2.07 bpm across a smaller sample size of 14 patients. Jung et al [31] tested a multichannel load cell bed, achieving an MAE of 1.76 bpm in the supine posture, with a maximum error of 3.03 bpm in lateral postures. Although these solutions show promising accuracy, their limited validation in diverse clinical environments restricts generalizability.

The Bland-Altman analysis underscores the algorithm’s high accuracy and reliability, validating its potential for clinical application. The minimal bias observed indicates a close alignment between the algorithm’s HR measurements and the reference values, highlighting its precision and robustness under diverse conditions. The LOA show that most differences between measured and reference HR values fall within a clinically acceptable range, with only a small fraction of data points outside these limits. This reinforces the algorithm’s reliability and strong agreement with reference data, even in the face of physiological and environmental variability. Notably, the bias and LOA observed in this analysis are comparable to those reported for HR measurements using BCG under real-world conditions. For instance, studies using BCG-based technologies, such as EarlySense in postoperative settings and Emfit in sleep lab environments [32,33], have demonstrated similar bias (−1.4 and 1.79) and LOA (10.4 to −13.2 and 19.11 to −15.33), further validating the robustness and reliability of the algorithm in practical, uncontrolled scenarios. This alignment with real-world performance benchmarks emphasizes the algorithm’s suitability for clinical applications where accuracy and consistency are paramount.

The Deming regression analysis, supported by the Pearson correlation, highlights the strong alignment between the algorithm’s HR measurements and the reference standard. The analysis demonstrates a robust linear relationship, reflecting the algorithm’s precision and reliability across diverse conditions. This strong correlation is particularly important in clinical settings, where accurate trend detection and reliable HR measurements are essential for timely interventions and effective patient management. By ensuring minimal variability and maintaining close agreement with reference values, the algorithm enables health care providers to monitor physiological changes confidently. The findings align with those of other noninvasive HR monitoring systems [34], further validating the algorithm’s performance. This establishes the algorithm as a viable tool for both hospital and remote monitoring, offering a reliable solution for continuous, contactless HR monitoring to improve patient care and outcomes.

Many existing approaches to HR estimation exhibit limitations in sample diversity and applicable HR ranges, which can impact their generalizability in clinical practice. For instance, Sumali et al [35] used a CNN algorithm for peak detection but limited their analysis to an HR range of 60 to 100 bpm, potentially overlooking extreme HRs encountered in clinical practice. Feng et al [36] used STFT and empirical mode decomposition for peak extraction on a small sample of 10 healthy participants and 2 patients with coronary heart disease, which limited applicability due to narrow demographics. Pino et al [37] tested only 34 normal participants and 24 with atrial fibrillation in controlled settings, which restricts the generalizability of their findings. Moukadem et al [38] based their analysis on data from just 3 participants, further limiting the applicability of their results. Yang et al [22] used peak detection with discrete wavelet transform across a small dataset of 24 participants, with recordings lasting about 3 minutes; this restricted duration and participant diversity limit applicability. Schranz et al [39] introduced a supervised deep learning method for J-peak extraction from BCG measurements. By modeling discrete heartbeat events with a symmetric, continuous kernel function (surrogate signal), this approach enhances target heartbeat detection. Their evaluation is based on a BCG dataset from 11 participants, collected over 17 nights of sleep, totaling 134 hours of ECG reference data that investigated various deep learning architectures, including convolutional, recurrent, and hybrid models, on a held-out testing set of 14 patients [37,39]. Choe et al [40] proposed an advanced algorithm using a moving dispersion calculation method to accurately identify heartbeat locations. This is enhanced by an adaptive peak detection technique that dynamically adjusts the detection window based on predicted heartbeat locations. The method effectively operates across a HR range of 40 to 110 bpm. In contrast to the above-mentioned studies, our algorithm stands out due to its validation across a broad HR range and diverse patient demographics, enhancing robustness and generalizability for real-world applications [40].

Currently, patients in hospital wards are often monitored solely through spot checks [41], highlighting the need for continuous HR monitoring. In this context, an error margin of 3 bpm is deemed acceptable for several reasons. This margin emphasizes trend detection over precise measurements, enabling health care providers to identify significant changes in HR that may indicate patient deterioration or response to treatment. Recognizing the natural fluctuations in HR due to physiological variability, this margin accommodates normal variations while still delivering clinically relevant data. Furthermore, it has been reported in the literature that HR is deemed clinically acceptable if it aligns within ±5 bpm or within ±10% of the reference standard, supporting the adequacy of these thresholds. Moreover, previous studies validate the adequacy of a ±3 bpm for effective monitoring, reinforcing its appropriateness for enhancing patient management and outcomes [3,42].

The key strength of this study lies in its innovative dual-domain algorithm, which integrates time-domain HR peak detection with STFT analysis. While the STFT effectively captures the fundamental heart-rate frequencies, the CNN adds significant value by identifying complex temporal and spectral patterns that may not be apparent in traditional frequency-domain representations and by improving robustness to noise and artifacts through learned pattern recognition. This is particularly advantageous in real-world clinical settings, where physiological variability and movement-related disturbances are common. This dual-domain CNN approach addresses the noise or artifact challenges noted in earlier methods and elevates the performance to a level suitable for demanding health care environments.

By combining these advanced techniques, our algorithm sets a new benchmark in contactless HR monitoring. Rigorously trained on a large and diverse dataset that includes varying levels of motion artifacts and physiological diversity, particularly from ICU settings with inherent noise and HR fluctuation, this algorithm ensures reliable HR detection across a wide range of clinical scenarios, positioning it as a transformative solution for HR monitoring in challenging health care environments. Importantly, we validated the algorithm in both controlled and real-world environments, demonstrating practical robustness. Performance remains consistent across age, sex, BMI, ethnicity, and comorbidities, underscoring its strong generalizability.

While this study validates the algorithm effectively, several limitations must be considered. The BCG data used for training and validation were exclusively collected from piezoelectric sensors, which may limit the algorithm’s applicability to other sensor types and reduce its generalizability. Testing across various BCG sensor technologies is recommended to enhance adaptability. Furthermore, while the algorithm performs well across a broad HR range, its accuracy with highly irregular heart rhythms, such as severe arrhythmias, remains untested and may require tailored adaptations. The CNN-based architecture also introduces computational demands that could hinder real-time processing in resource-limited environments; optimizing the algorithm for lower computational needs could improve scalability for telemedicine and home care applications. Additionally, we acknowledge a limitation in methodological transparency: the full CNN architectural details and certain parameter values are not disclosed due to the proprietary nature of the algorithm. This constraint, while necessary from an intellectual property standpoint, means that external researchers cannot yet fully reproduce the model. We have mitigated this by rigorously validating the algorithm on independent datasets. Finally, the algorithm’s performance is influenced by the characteristics of the training data, potentially affecting generalization to demographics or conditions that are underrepresented in the dataset.

Future studies could further enhance the algorithm by integrating additional vital signs, such as respiratory rate, to create a more comprehensive monitoring system. Expanding the dataset to include more data points for comorbidities will provide a stronger foundation for these refinements. Investigating the impact of comorbidities on BCG signals could further optimize the algorithm for specific patient populations, enabling customization for individual health profiles. As the demand for contactless patient monitoring grows, ongoing updates and algorithm refinements based on real-world data, combined with enriched datasets, will be essential to maintaining the algorithm’s accuracy, adaptability, and relevance in clinical practice.

In summary, this study presents a novel CNN-based, contactless HR measurement algorithm using BCG signals, designed for reliable, continuous HR monitoring in diverse clinical environments. By integrating advanced signal processing with ML, this approach achieves a high degree of accuracy, adaptability, and robustness, establishing a new benchmark for noninvasive monitoring solutions. The algorithm’s capacity to perform reliably across varied demographics and conditions highlights its potential as a transformative tool for real-time HR monitoring, promising improved patient outcomes through timely detection and interventions. Future research will focus on expanding the algorithm’s validation across various clinical environments and integrating additional vital signs to create a more holistic and effective monitoring solution.