Evaluating a ballistocardiography derived respiratory rate algorithm through comprehensive clinical validation across multiple settings

Abstract

Continuous respiratory rate (RR) monitoring is key for early deterioration detection, but conventional methods require contact. Ballistocardiography (BCG) offers an unobtrusive alternative using an under-mattress sensor. This study evaluated a BCG-based RR algorithm across diverse clinical settings and populations. BCG data from 11 studies involving 400 subjects across wards, ICUs, sleep labs, and healthy volunteers were analyzed. Reference RR was measured using capnography or polysomnography. An algorithm processed BCG signals via filtering,, dynamic thresholding, and median filtering to estimate RR. Performance was assessed by mean absolute error (MAE), detection rate (DR), and agreement with reference using Bland-Altman, Deming regression, and Pearson’s coefficient. Across 68,342 reference datapoints the algorithm achieved an MAE of 1.29 BrPM and a 92.68% detection rate. Performance was consistent across studies, with MAE ranging from 0.96 to 1.8 BrPM and detection rates from 85 to 97%. Accuracy was higher in controlled settings (e.g., sleep lab: MAE − 0.96 BrPM, 95.8% detection) and slightly lower in wards (MAE ~ 1.6 BrPM, 86–93% detection). Subgroup analyses by geography, demographics, and comorbidities showed consistently low error and good detection, except in COPD patients, where detection dropped to ~ 78% (MAE-1.3 BrPM). Bland–Altman and Deming regression showed minimal bias (–0.39 BrPM) and strong correlation (Pearson’s r = 0.86). A continuous, unobtrusive BCG-based RR measure offers high accuracy and detection, supporting its use in early respiratory event detection and improved patient care.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-33325-8.

Introduction

Respiratory rate (RR) is an essential vital sign for clinical monitoring. An elevated or depressed RR often foreshadows events like cardiac arrest, severe infection, or respiratory failure^1–4​. Because changes in breathing can be an early-warning signal of deterioration^3–5, timely recognition of RR abnormalities is vital for enabling prompt interventions¹​. Studies have shown that continuous RR surveillance in hospitals could reduce intensive care transfers, cardiac arrests, and mortality^6,7​. Despite this importance, routine RR monitoring is remarkably limited. In general wards, nurses typically measure RR by manually counting breaths for 30–60 s every few hours (e.g., once per 4–12 h) ​⁸. Such intermittent spot-checks not only miss transient but critical changes occurring between⁹, but are also prone to human error and inconsistencies ​^10,11. Indeed, RR is often termed the “neglected vital sign” because it is documented infrequently and sometimes inaccurately^12,13. The standard method of visual breath counting is labor-intensive, subjective, and can yield inconsistent results. One systematic review found that manual RR counts are time-consuming and frequently inaccurate, even when performed by trained staff¹⁴​. This gap in monitoring can delay the detection of patient deterioration; for example, opioid-induced respiratory depression episodes have been found to occur within minutes after a nursing round, well before the next scheduled check¹⁵​.

Current technologies for continuous RR tracking have significant limitations. In intensive care units (ICU), impedance pneumography via ECG leads is commonly used to estimate RR by measuring chest electrical impedance changes. This method often triggers false alarms due to motion or electrode issues, contributing to alarm fatigue​¹⁶. Outside the ICU, the requirement of wired leads attached to the patient limits mobility and comfort, making it impractical for general ward use¹​. Capnography, which measures exhaled CO₂ to determine RR, is considered a gold standard for ventilated patients, but for awake, non-intubated patients it requires a nasal cannula or mask. Patients often find these sensors uncomfortable and may dislodge them, and prolonged use on the ward is not feasible¹⁷​. Other approaches have emerged: for example, adhesive RR patches, acoustic monitors placed on the neck, or under-mattress pressure sensors. While these are less obtrusive than face masks or multiple wires, studies report that some commercially available systems have suboptimal accuracy for RR. In fact, the accuracy of wearable or bedside devices was deemed not clinically acceptable for respiration monitoring in at least one analysis¹⁸​. Novel non-contact methods such as camera-based tracking of chest movements, radar sensors, or infrared optical sensors have shown promise in research settings¹⁷.

Ballistocardiography (BCG) is another emerging technology for unobtrusive vital sign monitoring. BCG involves measuring the tiny vibrational forces produced by the heartbeat and breathing,typically via a vibration piezoelectric sensor embedded in the bed or chair on which a person lies. Prior work has demonstrated that BCG signals contain a clear respiratory modulation, allowing extraction of RR from an unrestrained patient ​^19,20. Because BCG sensors are placed under mattresses or chairs, monitoring is completely contact-free to the patient, causing virtually no discomfort^19,20. This approach eliminates the tangle of wires and reduces the burden on healthcare staff once deployed, since it can automatically and continuously log respiratory data. Given these advantages, unobtrusive BCG monitoring has been proposed to fill the respiratory surveillance gap on hospital wards²¹. Nevertheless, this promising methodology requires validation in varied clinical populations.

Dozee is a commercially available, unobtrusive vital signs monitoring system that leverages BCG technology through a thin piezoelectric sensor sheet placed under the mattress or bed foam, typically beneath the patient’s chest (Fig. 1). This under-mattress sensor captures minute vibrations generated by cardiac ejection forces and respiratory movements. The resulting analog voltage signal is processed by advanced algorithms to continuously track heart rate (HR), RR, HR variability (HRV), cardiac function, and sleep patterns^20,22–27. The system requires no physical contact with the patient; monitoring is achieved simply by the patient lying in bed. The HR and RR algorithms can measure HR within ± 3 beats per minute and RR within ± 2 breaths per minute (BrPM)^20,23.

Fig. 1.

A sensor sheet positioned beneath the bed, captures subtle micro vibrations from the patient in an unobtrusive manner.

Initial clinical evaluations of the RR algorithm were conducted on limited datasets. This study presents a comprehensive clinical performance evaluation of the RR measurement algorithm, analyzed retrospectively across multiple clinical studies and diverse patient populations. The primary objective is to assess the algorithm’s accuracy in continuous RR measurement. It is hypothesized that this BCG-based RR measurement approach can achieve accuracy comparable to reference RR measures and maintain reliable performance across varied conditions, including differences in geography, age, gender, body composition, clinical environments, patient and respiratory profiles.

Materials and methods

Study design and ethical considerations

A retrospective analysis was performed on data from 11 clinical studies that evaluated the BCG-based Dozee system (Table 1). These clinical studies encompass both RR reference values collected with the primary objective of evaluating the Dozee algorithm’s RR measurement, as well as studies where RR was recorded as part of broader data collection. These studies were carried out at multiple hospital sites and sleep laboratories, encompassing both Indian and US cohorts to ensure geographic diversity. Each study obtained appropriate ethics approval in accordance with its original protocol (details not included here for brevity). Each site followed its ethics-approved protocol, which specified how participants were approached and consented. In all cases, eligible participants or their legal representative were informed about the study’s scope, data collection procedures, and their rights, and written informed consent was obtained in accordance with ethical guidelines and local regulations.

Table 1.

Overview of study locations.

Study #	Study title	Location	Study subjects	RR monitoring
1	Pilot clinical evaluation of Dozee VS in hospital patients	Eastside Research Centre, Seattle, USA	Healthy Volunteers	Capnography
2	Exploratory study to evaluate relationship between interoception and sleep quality	National Institute of Mental Health and Neurosciences, Bengaluru, India	Healthy Volunteers	Non- capnography
3	Contactless monitoring of body vitals using Ballistocardiogram	SPARSH Hospital, Bangalore, India	Healthy Volunteers	Capnography
4	Validation of the Dozee Pro V1.1 system	SPARSH Hospital, Bangalore, India	Healthy Volunteers	Capnography
5	Validation of Dozee system	SPARSH Hospital, Bangalore, India	Healthy Volunteers	Capnography
6	Validation of the dozee Pro NX	Apollo Hospital, Jubilee Hills, Hyderabad, India	Ward Patients	Capnography
7	Pilot clinical evaluation of Dozee VS in Hospital Patients	Apollo Hospital, Jubilee Hills, Hyderabad, India	Ward Patients	Capnography
8	Contactless monitoring of body vitals using Ballistocardiogram	SPARSH Hospital, Bangalore, India	Ward Patients	Capnography
9	Dozee as a screening tool for sleep Apnea detection and classification	St John’s Medical College Hospital, Bengaluru, India	Apnea Subjects	Non- capnography
10	Dozee as a screening tool for sleep Apnea detection and classification	Nithra Institute of Sleep Sciences, Chennai, India	Apnea Subjects	Non- capnography
11	Evaluation of continuous non-invasive blood pressure monitoring through dozee	BGS Global Gleneagles Hospital, Bengaluru, India	ICU Subjects	Capnography

Non- capnography- nasal cannula/thermistor/ RIP belts are used

Given the multi-center and varied nature of these clinical studies, inclusion and exclusion criteria varied slightly across sites to accommodate study objectives, specific participant populations and local regulatory requirements. The inclusion and exclusion criteria used for each individual study is provided in Supplementary Material 1. These criteria were pre-defined in the respective study protocols and approved by the appropriate Institutional Review Boards or Ethics Committees prior to subject enrollment.

For the present retrospective analysis for this study, ethics approval was obtained from the Genebandhu Independent Ethics Committee, New Delhi, India (Ref No. ECG024/2024), in compliance with established regulatory guidelines.

Study setting and study subjects

The study settings ranged from research clinic, general hospital wards and ICUs to sleep study centers, both in India and the USA thereby covering a broad spectrum of use-cases (Table 1).

Healthy volunteers: Five studies (Studies 1–5) involved healthy individuals, both in sleep lab settings, hospital environments and research facilities, representing baseline conditions without clinical pathology. General Ward patients: Three studies (Studies 6–8) involved non-ICU hospitalized patients, clinically stable and resting in standard hospital ward beds, monitored in typical ward environments characterized by moderate patient movement, variable postures, and ambient clinical noise. Sleep apnea subjects: Two studies (Studies 9–10) enrolled patients with obstructive/central/mixed sleep apnea undergoing overnight sleep studies presenting conditions marked by irregular breathing patterns, frequent respiratory disruptions, and apneic events. ICU patients: One study (Study 11) enrolled critically ill patients observed in an intensive care environment characterized by high clinical acuity, frequent medical interventions, ventilatory support, and physiological instability.

Across 11 studies, a total of 400 subjects were monitored (Table 2). This combined cohort had a mean age of ~ 35 ± 14 years (range 18–94 years), including 235 males and 165 females. The population was diverse in body morphometry, with body mass index (BMI) ranging from underweight to obese (mean ~ 25.2 kg/m², range ~ 14–45). The combined dataset included 68,342 reference RR data points, covering a wide range from 5 to 52 BrPM. Most RR measurements fall within the typical clinical range (15–25 BrPM), with fewer points at the extremes, indicating broad and clinically relevant coverage of RR values (Fig. 2). Common comorbidities (Fig. 3) present (in various subsets) included hypertension, diabetes, cardiovascular disease, chronic obstructive pulmonary disease (COPD), and hypothyroid disorders, reflecting a real-world inpatient mix. The diversity in demographic characteristics and clinical conditions makes this cohort well-suited for comprehensive validation of respiratory monitoring across a wide range of scenarios.

Table 2.

Summary of demographics and RR data points across the datasets.

Study	Subjects	Sex (M: F)	Age (years, mean SD) (range)	Weight (Kg, mean SD) (range)	Height (M, mean SD) (range)	BMI (mean SD) (range)	RR Data points	RR range
Combined	400	235:165	35.37 ± 14.34 (18–94)*	70.20 ± 15.35 (40–135)*	1.67 ± 0.10 (1.18–1.93)*	25.18 ± 5.05 (14.17–44.92)*	68,342	5–52
1	50	12:38	33.00 ± 13.74 (18–58)	72.48 ± 15.84 (46–116)	1.68 ± 0.11 (1.50–1.93)	25.52 ± 4.58 (18-35.81)	6142	5–39
2	132	72:60	32.54 ± 11.52 (20–76)	66.48 ± 13.14 (47–110)	1.66 ± 0.10 (1.18–1.86)	24.18 ± 4.73 (16.22–39.44)	28,195	9–33
3	50	32:18	26.32 ± 4.71 (20–44)	69.35 ± 17.19 (40–135)	1.69 ± 0.08 (1.54–1.83)	24.17 ± 4.83 (16.87–40.31)	5277	5–33
4	10	4:6	26.50 ± 3.50 (26–33)	65.40 ± 10.46 (49–84)	1.64 ± 0.07 (1.67–1.73)	24.29 ± 3.59 (17.57–28.93)	968	8–36
5	20	10:10	29.85 ± 8.83 (23–60)	70.45 ± 15.74 (45–110)	1.68 ± 0.11 (1.34–1.87)	25.04 ± 4.92 (18.03–40.90)	1831	8–30
6	11	8:3	48.45 ± 10.42 (35–68)	73.59 ± 19.39 (44–115)	1.64 ± 0.10 (1.52–1.82)	27.14 ± 5.40 (19.04–35.16)	1249	9–52
7	20	13:7	43.95 ± 15.81 (22–69)	66.60 ± 11.31 (45–85)	1.66 ± 0.14 (1.32–1.84)	24.18 ± 3.87 (17.90-32.14)	11,587	5–36
8	51	39:12	37.27 ± 14.59 (20–77)	70.69 ± 14.19 (40–112)	1.68 ± 0.09 (1.49–1.84)	25.15 ± 5.07 (14.17–37.66)	3045	7–35
9	25	18:7	46.58 ± 14.11* (19–84)	79.77 ± 12.79* (55-103.5)	1.65 ± 0.08* (1.52–1.8)	29.58 ± 5.36* (19.86–40.82)	2218	5–31
10	8	8:0	43.63 ± 11.41 (27–59)	89.48 ± 27.01 (60–131)	1.71 ± 0.08 (1.59–1.79)	30.25 ± 6.65 (22.62–41.80)	759	5–28
11	23	17:6	52.65 ± 21.09 (20–94)	74.04 ± 15.50 (48–115)	1.67 ± 0.08 (1.40–1.75)	26.55 ± 5.55 (21.97–44.92)	7071	5–50

*Data is unavailable for one subject.

Fig. 2.

Distribution of RR data points across ranges. The majority fall within the clinical norm (15–25 BrPM).

Fig. 3.

Distribution of RR data across patient comorbidities, reflecting a diverse real-world inpatient population.

Reference respiratory rate measurement

In each study, RR reference data were recorded in parallel with the BCG to serve as the ground truth for the performance evaluation. The choice of reference method varied based on the study setting:

Capnography (EtCO₂-based RR)

In all non-sleep lab settings, capnography with a nasal cannula was employed to continuously measure RR by detecting exhaled carbon dioxide from each breath. Capnography is considered highly accurate and is recognized by the U.S. FDA as a gold-standard method for RR measurement. Capnography provided RR values in 1 min frequency, reflecting standard clinical practice for end-tidal CO₂ monitors. To maintain data integrity, segments affected by factors such as changes in patient posture or temporary absence from the bed were identified through annotations or automated outlier detection and were excluded from analysis using predefined quality control filters.

Polysomnography (PSG)

In sleep lab studies, RR reference was computed every minute from standard overnight PSG systems using three respiratory channels, nasal airflow thermistor/cannula, chest respiratory inductance plethysmography (RIP) belt, and abdominal RIP belt. These three respiratory signals are processed separately for RR extraction over matching 1 min segments and are not fused. Because the nasal cannula/thermistor and RIP belts operate based on distinct physiological signals i.e. airflow versus thoracoabdominal movement, variability among these channels is expected. To ensure the reliability of reference data, only time segments where RR measurements from all three channels closely matched, defined as a maximum inter-channel difference of ≤ 1 BrPM were included in the analysis. For these concordant segments, the nasal cannula/thermistor signal was designated as the representative PSG reference.

The use of both capnography and PSG as reference standards is supported by prior literature^28,29. Studies have shown that these modalities yield virtually interchangeable RR measurements under good signal conditions. For example, capnographic RR typically agrees with simultaneous RIP belt–derived RR within about ± 2 breaths per minute²⁸. Therefore, it is methodologically sound to use either modality as a ground truth, provided that signal quality and internal consistency checks are enforced. Indeed, recent validation studies have employed both capnography and PSG-based measurements as reference benchmarks²⁹, supporting the notion that either modality can serve as a valid RR.

RR measurement algorithm

The core of the study is the algorithm that processes the raw BCG signals to produce an estimated RR every 60 s. The algorithm was designed to be robust against noise and varying signal quality. It comprises several sequential stages as illustrated in Fig. 4 using a 120-second BCG signal.

Fig. 4.

RR extraction pipeline from BCG signals.

The raw BCG signal acquisition sampled at 250 Hz over a 120 s window(Step 1) is first subjected to median filtering and detrending (Step 2) to mitigate macro-movement artifacts and baseline drift. Next, stable segments are identified by computing a rolling standard deviation (STD) over a 5–10 s window (Step 3). Segments with STD values below a dynamic, empirically determined threshold are retained for further analysis.

These stable signal windows are then bandpass filtered in the 0.1–0.7 Hz range (Step 4) to isolate the respiratory frequency components corresponding to typical human breathing rates. Within the filtered signal, peak detection is performed (Step 5), and the intervals between successive peaks are used to compute instantaneous RR. To assess the stability and variability of breathing, a histogram of RR intervals is constructed (Step 6) .

The RR is calculated using the following formula: RR = (Number of Peaks / Median Interval) × 60. In the example shown, this yields an RR of approximately 16 BrPM.

Finally, a quality scoring mechanism which utilizes median interval and number of peaks evaluates each RR estimate based on the regularity of detected peak intervals. Estimates falling below a predefined confidence threshold are discarded, while those exceeding the threshold are retained as valid outputs. This quality filter ensures that only high-confidence RR estimates are generated.

The algorithm is designed to detect RR within the range of 5 to 55 BrPM encompassing the full spectrum of physiologically plausible values encountered in sleep, general ward, and intensive care settings.

Performance evaluation metrics

To comprehensively evaluate algorithm performance, combined data from all 11 clinical studies were used for an aggregated analysis. Additionally, several predefined subgroups were analyzed to assess performance across varied clinical conditions and population characteristics.

Subgroup analyses

By individual study

Each of the 11 clinical studies was evaluated independently to identify any study-specific biases or performance outliers.

By study location (India vs. USA)

To examine potential site-related influences, such as differences in bed types or patient demographics.

By RR range

Reference RR values were grouped into bins (5–15, 15–25, 25–35, 35–45, and > 45 BrPM) to assess performance across bradypnea, normal, and tachypnea ranges.

By Age Group

Participants were categorized by approximate decades (18–24, 25–30, 31–35, 36–42, 43–48, and ≥ 49 years) to explore age-related variations in accuracy.

By gender

Male and female subjects were compared to detect any sex-specific differences in algorithm performance.

By BMI category

Subjects were grouped based on standard BMI cut-offs,underweight (< 18.5), normal (18.5–24.9), and overweight/obese (25–45),to evaluate the effect of body morphometry. Higher BMI ranges were combined due to limited sample sizes.

By clinical category

Participants were classified as healthy volunteers, ICU patients, sleep apnea patients, or ward patients to assess performance across different clinical contexts.

By comorbidity status

Based on medical history, subjects were stratified by presence of comorbidities such as hypertension, diabetes, cardiovascular disease, COPD, and hypothyroidism. These categories were not mutually exclusive, as subjects could present with multiple comorbidities. This stratified approach was intended to identify any specific scenarios,such as COPD patients with irregular breathing patterns,where algorithm performance might be challenged. These subgroup evaluations enhance understanding of algorithm robustness across real-world use cases.

Primary evaluation metrics

Mean absolute error (MAE)

The primary accuracy metric, expressed in BrPM, calculated as the average absolute difference between algorithm measured and reference RR values at each timepoint. MAE was reported for the overall dataset as well as each subgroup.

Detection rate (DR)

The proportion of reference timepoints for which the algorithm successfully generated a valid RR estimate represented in percentage (%). Non-detections typically occurred when signal quality was insufficient, based on the internal confidence scoring mechanism. A higher DR indicates greater reliability and continuity in RR tracking. DR was reported for the overall dataset as well as each subgroup.

Bland-Altman analysis

Used to assess agreement between the algorithm and reference RR values, identifying mean bias and limits of agreement. This analysis was performed only on the combined dataset.

Deming regression and Pearson’s coefficient (r)

Applied to evaluate correlation and proportional bias between the methods. This analysis was performed only on the combined dataset.

All analysis was done using custom written Python scripts developed in-house.

Statistical analysis

Agreement between the algorithm estimated RR and the reference measurement was evaluated using descriptive and inferential statistics. The mean paired difference (bias) and standard deviation (SD) of the paired errors were calculated to characterize systematic deviation and variability. The magnitude of the observed difference was quantified using Cohen’s d for paired measurements, computed as the mean paired difference divided by the SD of the paired differences.

A post-hoc power analysis was performed based on the observed effect size, significance level (α = 0.05), and sample size to determine whether the study was sufficiently powered. All analyses were performed using Python statistical libraries.

Results

Overall algorithm performance

Across all subjects and settings, the RR algorithm performed with a high degree of accuracy and consistency. The overall MAE between the BCG-derived RR and the reference was 1.29 BrPM (Table 3). This indicates that on average, the algorithm’s reading was within ~ 1.3 breaths of the reference value – a small error relative to usual adult RR. The overall DR was 92.7%, meaning the algorithm successfully validates RR reading for about 93% of the reference measurement​. In the remaining ~ 7% of intervals, the system did not produce an RR, because of transient motion artifacts or other factors which affect the BCG quality.

Table 3.

Summary of RR measurement algorithm performance across the studies.

		Reference data points	Measured data points	MAE	DR (%)
Overall		68,342	63,341	1.29	92.68
Studywise	Study 1	6142	5729	1.8	93.28
	Study 2	28,195	27,003	0.96	95.77
	Study 3	5277	4858	1.67	92.06
	Study 4	968	938	1.4	96.9
	Study 5	1831	1686	1.81	92.08
	Study 6	1249	1082	1.59	86.63
	Study 7	11,587	9848	1.65	84.99
	Study 8	3045	2833	1.71	93.04
	Study 9	2218	2135	1.15	96.26
	Study 10	759	711	1.15	93.68
	Study 11	7071	6518	1.09	92.18
Study location	USA	6142	5729	1.8	93.28
	India	62,200	57,612	1.24	92.62

Agreement and correlation analyses

To further illustrate the agreement between the BCG-based RR and the reference RR, two key plots are presented: a Bland-Altman plot (Fig. 5) and a Deming regression plot (Fig. 6).

Fig. 5.

Bland-Altman plot comparing BCG-derived RR and reference RR across all data. Dotted lines indicate mean bias (-0.39 BrPM) and 95% limits of agreement (± 4.6 BrPM). The tight clustering of points shows high agreement with no significant systematic error.

Fig. 6.

Deming regression plot (BCG-derived RR vs. reference RR) with Pearson’s coefficient. The regression line and the Pearson’s r ~ 0.86 illustrate the good correlation and accuracy of the unobtrusive RR measurements.

Bland-Altman Analysis shows that the mean bias was essentially zero (mean difference − 0.39 BrPM, indicated by the central line). The 95% limits of agreement (LoA) were roughly ± 4.6 BrPM around the mean. Specifically, no systematic bias was observed across the range of RRs, the points are symmetrically scattered around zero difference, and there was no trend indicating skew at high or low RR. A slight increase in variability at higher RRs was seen, which is common as there is inherently more variability when breathing is rapid.

Nevertheless, virtually all points lay within the clinically acceptable error band of ± 5 BrPM, and the vast majority within ± 3 BrPM, reinforcing that large errors were very rare. This level of agreement is on par with or better than many contact-based RR monitoring technologies³⁰​.

Deming regression (Fig. 6) was fitted to account for measurement error in both methods. The resulting line had a slope very close to 1 and an intercept near 0, indicating that the BCG method neither consistently overestimates nor underestimates RR across the measured range. The Pearson correlation coefficient was calculated at r = 0.86 for the entire dataset, highlighting a strong linear relationship. This high correlation is visualized in (Fig. 6) by the tight clustering of points around the identity line. In practical terms, this means if a patient’s breathing speeds up or slows down, the BCG system almost always detects that change proportionately. The strong correlation and near-unity regression fit, combined with the low bias from Bland-Altman, confirms that the algorithm achieves accurate tracking of RR in a manner equivalent to the clinical standard measurements.

From a total of 63,341 paired RR observations collected across 400 subjects, the algorithm demonstrated a mean bias of − 0.39 BrPM with a standard deviation of 2.13 BrPM. The paired effect size, Cohen’s d = 0.18, was small, indicating minimal systematic deviation from the reference measurement. Based on the observed effect size and available sample size, the post-hoc statistical power exceeded 0.999 for α = 0.05, confirming that the study was adequately powered to detect even small deviations.

Subgroup performance analysis

Table 3 breaks down performance by each individual study (1 through 11). Table 4 presents a detailed breakdown of algorithm performance across various RR ranges, demographic groups, and clinical subgroups.

Table 4.

Summary of RR measurement algorithm performance across different demographic sub-group.

Demographic variables		Reference datapoints	Measured datapoints	MAE	DR (%)
RR (BrPM)	5–15	14,360	13,235	1.54	92.17
	15–25	50,271	46,872	1.21	93.24
	25–35	3536	3065	1.65	86.68
	35–45	126	121	1.58	96.03
	> 45	49	48	1.31	97.96
Age (years)	18–24	13,825	13,075	1.27	94.58
	24–30	17,614	16,645	1.23	94.5
	30–36	11,336	10,928	1.12	96.4
	36–42	6893	5828	1.54	84.55
	42–48	3751	3247	1.37	86.56
	>=48	14,830	13,539	1.4	91.29
Gender	Female	29,667	27,300	1.22	92.02
	Male	38,675	36,041	1.34	93.19
BMI (kg/m²)	< 18.5	4765	4612	1.16	96.79
	18.5–24.9	35,690	32,665	1.26	91.52
	25–45	27,794	25,985	1.35	93.49
Race/ethnicity	Asian-American	1122	1046	1.47	93.23
	Black/ African American	378	348	1.44	92.06
	Indian	62,200	57,612	1.24	92.62
	Mixed	241	230	1.81	95.44
	White/Caucasian/unknown	4401	4105	1.92	93.27
Study subjects	Healthy Volunteers	42,413	40,214	1.21	94.82
	ICU Subjects	7071	6518	1.09	92.18
	Sleep Apnea Subjects	2977	2846	1.15	95.6
	Ward Patients	15,881	13,763	1.66	86.66
Comorbidities	Hypertension	5788	4759	1.36	82.22
	Diabetes	5677	4781	1.95	84.22
	Cardiovascular Disease	2723	2314	2.14	84.98
	Obesity	2655	2262	2.04	85.2
	COPD	1713	1342	1.33	78.34
	Apnea	2959	2834	1.13	95.78
	Hypothyroidism	1234	1066	1.24	86.39
	Hyperthyroidism	1769	1455	1.37	82.25
	Trauma	1372	1352	0.75	98.54
	Hemorrhage	1194	1145	0.86	95.9
	Others	4296	3737	1.35	86.99
	No Comorbidities reported	50,214	47,285	1.25	94.17

Individual study (Table 3)

Consistently low MAE values were observed in all studies, ranging from 0.96 BrPM up to 1.81 BrPM. DR per study ranged from about 85 to 96%.

Study 1, conducted in the USA, showed a slightly higher MAE = 1.8 BrPM with a DR = 93.3%. Although the overall DR was strong, the relatively higher error could be attributed to factors such as occasional patient movement or signal quality variability due to environmental or bed-related conditions. Nevertheless, the high DR indicates that the algorithm was still able to track most respiratory cycles accurately.

The highest accuracy was seen in Study 2, a sleep study, with MAE = 0.96 BrPM and DR = 95.8%​. This study involved overnight monitoring of healthy volunteers in a controlled sleep laboratory setting. The controlled environment likely ensured stable sensor contact, and also aided by the high confidence in reference RR measurements due to concordance across all three reference channels.

Study 9 and Study 10 both sleep apnea study in a hospital setting similarly showed high detection rate (~ 93–96%) and low error (~ 1.1 BrPM), reinforcing that the algorithm handles disordered breathing well. In these cases, reference RR would drop to very low during apnea events and then spike during recovery breaths; the RR algorithm reliably captured these dynamics. As mentioned above the low error can be attributed to high confidence RR measurements.

Study 11, conducted in an ICU, had MAE = 1.09 BrPM and DR = 92.2%​. The ICU subjects often had higher RR and some motion (e.g., nurses tending to them). The strong performance here indicates the algorithm’s robustness against typical ICU disturbances (bed adjustments, etc.). Notably, (Fig. 7) illustrates a segment from an ICU patient in which the BCG-derived RR is overlaid on the capnography RR trace: the two coincide closely over time, even as the rate varies from ~ 12 up to 45 BrPM, evidencing the algorithm’s ability to follow rapid respiratory changes.

Fig. 7.

Example of continuous RR tracing from BCG-derived RR vs. capnography RR in an ICU patient (time-series overlay). The trace demonstrates the BCG-derived RR following rapid changes including bradypnea (12–14 BrPM) and tachypnea (> 45 BrPM). The missing segments in BCG RR occur because the algorithm removes low-quality signal portions.

Studies 6 and 7, which were general ward patients at a hospital in Hyderabad, India, showed slightly higher errors (MAE= 1.59 and 1.65 BrPM) and detection rate (86.6% and 85.0%, respectively)​. These represent some of the lower-end performances of this algorithm. Possible reasons include more frequent patient movement or differences in bed padding that attenuated signals. Even so, an error ~ 1.6 BrPM is still quite low, and a ~ 85–87% DR, while below the average, means the system was capturing most breaths.

Studies 3–5 and 8 involved a mix of ward patients and healthy volunteers at another site (Bangalore, India). These had MAEs mostly between ~ 1.4–1.7 BrPM and detection ~ 92–96%, aligning well with the overall average. Healthy volunteers (Studies 3–5) yielded a high detection rate (92–97%)​, as expected when subjects remain relatively still.

Location (Table 3 )

Splitting by location, data collected in the USA vs. India showed virtually no difference in performance trends, aside from a slightly higher MAE in the smaller USA cohort (1.8 vs.  1.2 BrPM in the larger India cohort). Both groups’ DR were ~ 93%, indicating consistency across regions. This supports the algorithm’s generalizability across different healthcare environments and patient populations.

RR range (Table 4)

The algorithm maintained good accuracy across the full spectrum of RR encountered. In normal RR ranges (15–25 BrPM, the largest bin of ~ 50k data points), MAE was 1.2 BrPM with > 93% DR​. Even at low rates (5–15 BrPM, e.g., during deep sleep or opioid-induced depression), MAE remained 1.54 BrPM, and DR > 92%, indicating the system can still track slow breathing well. At moderately high rates (25–35 BrPM), error was 1.65 BrPM, and DR dipped to 86.7% ,this slight drop likely reflects that some episodes of rapid breathing coincided with movement (e.g., patients might be more restless when breathing heavily). Importantly, in the most extreme tachypnea range (> 45 BrPM, which had only 49 reference points, indicating rare events of very fast breathing), the algorithm still captured 98% of those instances with MAE of 1.31 BrPM.

Age (Table 4)

Performance was broadly consistent across age brackets. Young adults (18–30 years) and middle-aged groups had MAEs ~ 1.1–1.3 BrPM and DR of ~ 94–96%. Older adults (e.g., ≥ 48 years group) showed a slight increase in error (~ 1.4 BrPM) and a small drop-in detection rate (91%), but notably the 36–42 and 42–48-year bins had somewhat lower detection (84–86%). No systematic degradation is seen with aging, which is encouraging for use in elderly populations.

Gender (Table 4)

The algorithm performed similarly for females (MAE = 1.22 BrPM, DR = 92.0%) and males (MAE = 1.34 BrPM, DR = 93.2%). This indicates that physiological differences in body shape or breathing patterns between sexes did not materially affect the sensor’s ability to pick up respirations.

BMI (Table 4)

Underweight (BMI < 18.5), normal (18.5–24.9), and overweight/obese (BMI ≥ 25 up to ~ 45) categories were analyzed. Interestingly, the lowest error was in the underweight group (MAE =1.16 BrPM) with a very high detection ~ 96.8%. Smaller body habits likely allow clearer transmission of BCG signals. Normal BMI individuals had near-average performance (MAE =  1.26, DR ~ 91.5%). The combined overweight/obese group (BMI 25–45) had MAE = 1.35 and DR ~ 93.5%. These results show that even with higher body weight (which can dampen bed vibrations), the algorithm still works effectively. There was no dramatic drop-off in obese patients; although extremely high BMI (> 45) was not represented, the range up to 45 kg/m² suggests robustness for most typical hospital populations.

Race/ethnicity (Table 4)

Most of the dataset (≈ 91%) were Indian subjects, reflecting the greater number of studies in India, for which MAE ~ 1.2 and DR ~ 92.6%. Smaller subgroups of other ethnicities (from the US study) included Asian American, Black, Caucasian, and a few mixed-race participants. Their sample sizes are modest but generally performance was comparable: e.g., Black/African American subjects had MAE ~ 1.4, DR ~ 92.1%; Caucasian/White had MAE ~ 1.9, DR ~ 93.3%. Overall, no clear race-related performance issues emerged, suggesting the algorithm is race-neutral in accuracy.

Clinical subject (Table 4)

Healthy volunteers: (n  = 42413 data points) MAE =  1.21, DR = 98.9%. The error here is low and DR is high as expected when subjects remain relatively still.

ICU patients: (n = 7071 points) MAE =  1.09, DR = 92.2%. ICU data had the lowest error, possibly due to many ICU patients being monitored in stable supine positions with minimal motion.

Sleep apnea subjects: (n =  2977 data points) MAE=  1.13, DR = 95.8%. The error here is low due to high confidence reference RR measurements as mentioned before.

Ward patients: ( n= 15881 points) MAE = 1.66, DR = 86.7%. This group had the highest challenges, likely because ward patients are more prone to leaving bed or moving (no sedation or monitoring constraints as in ICU). Even so, an 86% detection over many hours of unsupervised ward time is a strong result for an unobtrusive monitor, and the error of ~ 1.6 BrPM is within acceptable clinical range.

Comorbidity (Table 4)

Hypertension (HTN)

Patients with hypertension (n = 5788 reference points from) had MAE = 1.36, DR = 82.2%. DR here is on the lower side, possibly because hypertensive patients might have been older and more restless.

Diabetes

MAE =  1.95 (slightly higher), DR = 84.2%. The slightly higher error could be coincidental or due to a smaller sample; diabetics might overlap with other conditions.

Cardiovascular disease

(e.g., heart failure or CAD patients) MAE =2.14, DR=  84.98%. This was one of the higher errors noted. Such patients might have concurrent arrhythmias or edema that affect signal quality.

Obesity

(BMI criteria) MAE =  2.04, DR =  85.2%. Interestingly, while the combined BMI category didn’t show much degradation, specifically flagged “obesity” as a comorbidity did show a 2 BrPM error and only 85% detection. This suggests extremely high BMI individuals (or those with obesity-related breathing issues) pose some challenge, possibly requiring sensor repositioning or algorithm tuning for very low signal-to-noise scenarios.

COPD

Chronic obstructive pulmonary disease patients (n =  1713 points) had DR only 78.3% – the lowest of any subgroup – though MAE remained fairly low at 1.33 BrPM. COPD patients can have erratic breathing patterns (including coughing or breath-holding) and may shift frequently to alleviate discomfort, explaining the increased data dropouts. Notably, when data was captured, the accuracy did not suffer much (error ~ 1.3).

Apnea

Patients with diagnosed sleep apnea (overlap with the apnea studies) unsurprisingly did very well: MAE =  1.13, DR = 95.8%.

Hypothyroidism, hyperthyroidism had intermediate results (MAE ~ 1.2–1.37, DR ~ 82–86%). Trauma patients (monitored during recovery) interestingly had one of the highest DR (98.5%) and very low error (0.75 BrPM), likely because these patients were bedridden with minimal motion, allowing excellent signal capture.

Others

Subjects categorized under “Others” (n =4296 reference points) exhibited a MAE of  1.35 BrPM and a DR of  87.0%. This group likely includes patients with less common or unspecified conditions not covered under primary diagnostic categories. While the MAE remains well within clinically acceptable limits, the slightly lower DR may be attributed to heterogeneity within this subgroup,potentially including cases with intermittent movement, postoperative instability, or transient physiological abnormalities. Nevertheless, the algorithm maintained stable accuracy across this diverse category, reinforcing its resilience even when faced with atypical or less-characterized patient profiles.

No Comorbidities Reported: Subjects without any reported comorbidities (n ≈ 50,214 reference points) demonstrated robust algorithm performance, with a MAE of  1.25 BrPM and a DR of  94.2%. This high DR underscores the system’s reliability in individuals without underlying chronic conditions, reflecting more stable respiratory patterns and less confounding physiological noise. The accuracy in this subgroup closely mirrors the overall cohort, reaffirming the algorithm’s core precision. Notably, this group represents the largest data volume and serves as a benchmark for expected performance in low-risk patients, where unobstructed signal acquisition and minimal motion artifacts contribute to consistently high-fidelity measurements.

Overall, aside from certain chronic conditions the algorithm’s accuracy stayed within ~ ± 2 BrPM across all subgroups. No subgroup saw an average error exceeding 2.1 BrPM, underscoring consistent performance. When the device did miss data (lower DR), it was typically in scenarios where any non-contact sensor would struggle (significant movement artifact). Overall, these results confirm that the algorithm generalizes across healthcare settings and populations without loss of accuracy.

Discussion

In this multi-study evaluation, it was demonstrated that an unobtrusive BCG-based algorithm can monitor RR continuously with accuracy on order of less than 2 BrPM and with high reliability across a variety of patient populations. To our knowledge, this is one of the most extensive validations of a BCG respiratory monitoring system, aggregating data from hospital wards, ICUs, and sleep labs with hundreds of subjects. The results are significant given the longstanding challenges in practical RR monitoring on general hospital wards.

Significance of findings

The ability to accurately track RR without patient contact has important clinical implications. Early detection of respiratory abnormalities can be life-saving – for instance, detecting an upward trend in RR could herald sepsis or cardiac failure³¹, whereas a downward trend could indicate impending respiratory arrest or oversedation⁹​. Our findings show that the BCG device captures these changes comparably to capnography, which is considered a gold standard. Notably, even when patients had irregular breathing patterns (such as in sleep apnea or some COPD cases), the algorithm usually obtains correct reading. This is critical for clinical trust: a ~ 93% DR with validated accuracy implies the device can be relied upon to alert clinicians to true respiratory changes most of the time. In contexts like post-operative opioid monitoring, continuous RR has been shown to reduce incidents of unrecognized respiratory depression​³². By delivering continuous data with minimal burden, a system like this could integrate into early warning score protocols (e.g., MEWS or NEWS) to enhance their sensitivity, especially since RR is a heavily weighted component of such scores​². Furthermore, the strong performance in low-RR ranges is valuable for patient safety: bradypnea (RR < 10) was detected reliably, which is crucial for alarm systems to catch narcotic-induced respiratory depression. High-RR accuracy ensures that conditions like fever or pain-induced hyperventilation are also tracked.

Advantages over traditional methods

Compared to intermittent manual measurement, the benefits are clear, nearly continuous data versus a single snapshot every few hours​⁷. Continuous monitoring can reveal trends or transient events that manual checks would miss​¹⁵.

ECG-derived or impedance RR is a commonly used technique. Our unobtrusive method avoids the problem of electrode misplacements and false alarms from patient movement that plague impedance pneumography¹⁶​. By using mechanical vibration, which inherently stops when true breathing stops, the BCG is less likely to report a false breath (whereas impedance can be fooled by motion artifact).

Capnography outside ICU is hampered by patient non-compliance (nasal cannulas are often removed by patients). An unobtrusive system solves the compliance issue, nothing is required of the patient. Additionally, capnographs require consumables (cannulas) and can incur cost and nursing effort to maintain, whereas a BCG pad is a one-time placement and can work for months without intervention. In terms of accuracy, our MAE ~ 1.3 BrPM is on the order of magnitude of differences one might even see between two gold-standard capnographs themselves​. This suggests that the BCG system’s accuracy is comparable enough to serve as a practical alternative when capnography is not feasible.

Many wearable RR sensors or devices (chest bands, stickers, etc.) can be uncomfortable long-term and may not be tolerated by elderly or fragile patients. Under-mattress sensors like BCG are imperceptible to the patient¹⁹, improving adherence to monitoring. The trade-off is that BCG is only effective when the patient is in bed, but hospitalized patients generally are in bed at least at night, if not most of the day when unwell. For times when they are ambulatory, nurses would still need to be vigilant, but the system immediately picks back up when the patient returns to bed as seen by the quick recovery of detection after movement.

Comparison to other unobtrusive RR measurements

Recent studies on other unobtrusive RR monitoring have included radar-based systems¹​ and camera-based algorithms. Radar systems, for instance, have shown impressive accuracy (sub-1 BrPM error) in short-term recordings of resting patients¹​. Our BCG system achieved slightly higher error overall, but our testing included many non-resting scenarios and a much longer duration. In exchange, BCG sensors are passive and do not radiate signals unlike radar and are relatively low-cost. Camera-based approaches using computer vision can monitor breathing by detecting chest movements, but they face challenges in low light, privacy concerns, and line-of-sight requirements. In contrast, the BCG sensor works in darkness and under covers and is not video based maintaining patient privacy.

The performance of this BCG system is comparable to other state-of-the-art unobtrusive BCG systems. These studies span a wide range of sensing BCG modalities from traditional bed-embedded BCG pads, accelerometers, and piezo-force sensitive resistor (FSR) sensors to fiber-optic and RFID-based systems, and even vehicle seats, as well as diverse algorithms (Hilbert transforms¹⁹, empirical mode decomposition with thresholding³³, frequency-domain FFT³⁴ and wavelet analyses, etc.) and environments (home bedrooms, sleep labs, general wards, ICUs, and moving vehicles). In a controlled sleep-lab study¹⁹ , et al., achieved an MAE of ~ 0.7 BrPM on nine subjects by applying a Hilbert-transform method to BCG signals¹⁹. The classical NAPS passive BCG bed system yielded about 2.10 BrPM error on 20 subjects in a sleep laboratory³⁵. Other modalities show varied performance: a fiber-optic “smart mattress” in a cardiac ICU (CICU) achieved ~ 1.97 ± 2.12 BrPM error using wavelet analysis; a home-use accelerometer under-mattress system reached ~ 1.52 BrPM error in 23 subjects after filtering³⁶; and force-sensitive resistor (FSR) strips under the bed attained ~ 2.32 BrPM error on 20 subjects with wavelet-based processing³³. An EMD-based approach with adaptive thresholding in an in-home setting showed ~ 1.43 BrPM error on 16 subjects³⁷. Notably, a cutting-edge RFID mattress sensor has demonstrated exceptionally low error (~ 0.6 BrPM) in breathing rate estimation, while even automotive seat sensors using BCG and FFT frequency analysis can measure respiration within about ± 5 BrPM on 11 patients as reference³⁴. Unlike these prior works that validated under tightly controlled conditions or with relatively small cohorts, our approach achieves competitive error rates while demonstrating robust generalizability and reliability over a large dataset in heterogeneous clinical and sleep environments.

Limitations

Despite the positive results, certain limitations of our study and technology should be acknowledged. First, the BCG sensor requires that the patient remains on a bed or chair with the sensor in place; if a patient frequently gets up or is never in bed (e.g., ambulatory patients), the utility is reduced. Thus, this system is best suited for inpatients, post-surgical recovery, or sleep monitoring, and not for patients fully on their feet.

Second, periods of very high activity or gross movement artifacts are still a challenge. While our algorithm smartly refrains from outputting during those times, it means there are brief gaps in continuous monitoring. Encouragingly, in our data even the worst-case subgroup (COPD) had ~ 78% coverage, meaning the monitor was active most of the time. Nevertheless, future algorithm enhancements could focus on faster re-initialization after movement and perhaps sensor fusion (e.g., combining BCG with an accelerometer to distinguish motion vs. breathing).

Third, our reference standard was not uniform across all studies; while we primarily used capnography or PSG, some variation in reference accuracy could influence results. However, only confident RR measurements were taken from PSG as mentioned above and given the good number of capnography RR datapoints, such effects likely even out and our overall error is small, indicating we were close to the true value.

Another consideration is that our population, though large, did not include pediatric patients or those with very severe motion (e.g., delirious patients thrashing). Further validation in those populations would be valuable to confirm generalizability. Additionally, while we did include ICU patients, most were breathing spontaneously; testing the algorithm in mechanically ventilated patients (where chest vibrations might differ) could be a next step or at least filtering out ventilator respirations.

Future directions

Building on this work, future studies could explore integrating multiple sensors (e.g., multiple BCG sensors under different bed locations) to see if redundancy can improve detection during movement (one sensor might pick up signal if a patient shifts weight off the other). The algorithm could also incorporate additional machine learning approaches using our extensive dataset to further improve detection of breath waveforms amidst noise. Another interesting direction would be to combine HR and respiratory information from BCG to potentially detect patterns like respiratory sinus arrhythmia or to cross-verify the vital signs (the same BCG signal gives both heart and respiratory data). Finally, clinical outcome studies are warranted: for example, deploying this continuous RR monitoring in a ward and measuring if it indeed leads to earlier intervention for deterioration (e.g., comparing code blue rates or ICU transfer rates before vs. after implementing BCG monitors). Such studies will ultimately quantify the patient safety benefit hinted at by our technical validation.

Conclusion

Continuous RR monitoring is increasingly recognized as a crucial component of patient safety and effective clinical care. Our study demonstrates that an unobtrusive BCG-based RR algorithm can fulfill this role by providing accurate, continuous, and comfortable monitoring of respiration. The algorithm achieved a mean absolute error of about 1 BrPM and over 90% DR in diverse clinical scenarios, matching the performance of many contact-based devices. This level of accuracy means clinicians can trust the readings for decision-making. The presented BCG-based unobtrusive RR monitoring algorithm validated in this work represents a significant innovation in vital sign monitoring. Our findings support the use of BCG as a reliable means to achieve widespread continuous RR monitoring, ultimately advancing patient care toward more proactive and preventive models. Further monitoring of real-world performance is needed to assess cases where the current algorithm may have challenges.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Kumar Chokalingam was responsible for the study design and conceptualization, manuscript drafting, and critical revision of the manuscript. Muthukumarasamy Saravanan and Ashish Kaushal contributed to algorithm development and participated in manuscript review. Ashwathi Nambiar also contributed to algorithm development and critically reviewed the manuscript. Inam Ur Rahman and Siva Bhavana were responsible for data collection and data analysis. Mudit Dandwate, Ravi Mahajan, Kunal Sarkar, and Yogesh Kothari contributed to study conceptualization and critically reviewed the manuscript. Gaurav Parchani contributed to the study design, conceptualization, and manuscript review. All authors read and approved of the final manuscript.

Data availability

The datasets generated and/or analyzed during the current study are not publicly available due to patient consent restrictions. However, they are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.