Validating a smart bed against polysomnography for sleep apnea detection

Abstract

Sleep-disordered breathing (SDB), including obstructive sleep apnea (OSA) and central sleep apnea (CSA), significantly impairs sleep quality and overall well-being. This study evaluates a novel algorithm, developed and trained by the authors, using ballistocardiography (BCG) data collected from a non-intrusive smart bed platform. The algorithm aims to detect SDB events and estimate whether the apnea-hypopnea index (AHI) is ≥ 15, indicative of moderate to severe apnea. We analyzed data from 104 participants (48 males, 56 females; 21 with AHI ≥ 15 (13 males, 8 females), 83 with AHI < 15) by comparing algorithm-generated AHI estimates with standard polysomnography (PSG)-based AHI measurements. The algorithm achieved an accuracy of 83.3% in identifying individuals with moderate-to-severe apnea (AHI ≥ 15), demonstrating a sensitivity of 76% and specificity of 85%. Visual inspection of signals during apnea episodes, particularly those related to CSA, confirmed the algorithm’s capability to capture meaningful physiological patterns. The unobtrusive design of the smart bed facilitates longitudinal sleep monitoring without requiring cumbersome equipment or specialized technical expertise. Future research will focus on validating the algorithm using multi-night, real-world data to enhance its generalizability. Smart beds show promise for early detection and personalized management of SDB, potentially improving clinical outcomes through improved tracking and targeted intervention.

Introduction

With a growing global prevalence affecting approximately 936 million adults aged 30–69 years, obstructive sleep apnea (OSA) has emerged as a significant public health concern¹. Between 1988 and 2010, the prevalence of mild to severe sleep-disordered breathing increased by approximately 30%, mainly due to an increase in obesity². Despite advancements in understanding its diverse manifestations, sleep-disordered breathing (SDB) remains largely underdiagnosed, with estimates indicating that nearly 92% of women and 82% of men with moderate-to-severe apnea remain undetected³. Such underdiagnosis not only limits opportunities for early intervention but also imposes a substantial economic burden, with undiagnosed OSA alone accounting for costs of approximately $14.6 billion in the United States in 2015⁴. Additionally, untreated OSA is associated with multiple comorbidities, such as hypertension, cardiovascular diseases, metabolic disorders, and impaired functional status, thereby necessitating new and efficient diagnostic approaches^5,6.

Despite the well-documented ramifications of untreated SDB, barriers persist in timely diagnosis and effective management. Negative perceptions surrounding diagnosis and treatment, limited symptom awareness, and fragmented healthcare services have impeded the detection and treatment of SDB⁷. Patients’ perceived stigma associated with SDB, combined with the inconvenience of diagnostic procedures, contributes to their reluctance to seek medical care⁷. Current diagnostic modalities, such as clinical polysomnography (PSG), limit patient throughput and are often perceived as cumbersome and expensive. This highlights the necessity for innovative diagnostic tools offering convenience without compromising accuracy^8,9.

Currently, the diagnosis of OSA primarily relies on in-laboratory PSG, regarded as the gold standard, despite its inconvenience and requirement for specialized technical personnel^10,11. Home sleep apnea tests (HSATs) offer a more convenient alternative, and their usage is widespread. Indeed, the Mayo Clinic reports that approximately 40% of their patients diagnosed with OSA now utilize HSATs¹². However, HSATs are typically not designed for prolonged or long-term monitoring spanning weeks or months. They are particularly effective for diagnosing obstructive sleep apnea in patients with a high pre-test probability of moderate-to-severe OSA, but their utility is mostly realized in short-term diagnostic contexts. This limitation is partly due to the design of the devices, optimized for episodic rather than continuous, long-term use. Moreover, HSATs may have limitations in detecting certain conditions, such as central sleep apnea, and could exhibit lower accuracy in identifying mild OSA cases¹³. Recent research has explored innovative methods for OSA detection, demonstrating the potential of under-the-mattress sleep-analysis devices and portable diagnostic tests^14–16. The study by Edouard et al., 2021, highlighted the efficacy of under-the-mattress devices in identifying moderate-to-severe sleep apnea¹⁵, and the Khor et al., 2023 study underscored the suitability of portable diagnostic tests in detecting mild-to-severe OSA¹⁷.

The smart bed platform described in¹⁸ utilizes ballistocardiography (BCG) sensors to non-intrusively assess sleep patterns, monitor body movements, and even estimate heart rate (HR) and breathing rate (BR)¹⁸. Millions of smart beds are currently deployed in the US, offering an opportunity to quantify undiagnosed sleep apnea cases and raise awareness for earlier diagnosis and treatment. This manuscript presents a non-interventional observational study to validate the accuracy of the smart bed’s SDB detection and single-night apnea-hypopnea index (AHI) estimation against in-laboratory PSG data. Clinically, the AHI is used to classify disease severity by counting the number of complete (apneas) or incomplete (hypopneas) obstructive events per hour of sleep⁶.

Methods

Study design and data collection

This study was approved by an Institutional Review Board (IRB 19–1848, University of Chicago). All procedures and methods were conducted in strict accordance with the guidelines approved by the IRB. It consisted of three separate sleep studies carried out from 2016 to 2024, involving distinct participant cohorts for each trial. These participants were primarily recruited to evaluate smart bed functionality. Subsequently, permission was obtained from the IRB to utilize the deidentified data for secondary analyses. Participants underwent overnight sleep studies using polysomnography (PSG) at the Lakeland Sleep Center in Minneapolis, Minnesota, and the Sleep Health Center at Northwestern University in the western Chicago area, Illinois. The studies targeted recruitment of self-identified ‘normal sleepers.’ Recruitment ensured that participants had no medical, neurological, or psychiatric conditions, verified through a health status questionnaire. All participants were adults who had given their written consent to be involved in research and informed consent was obtained from all the participants. The in-lab sleep sessions utilized smart beds alongside PSG technology and occurred within specific periods: February 21 to March 31, 2016; May 15 to July 21, 2020; and August 2023 to March 2024. Demographic information about the participants is detailed in Table 1.

Table 1.

Participant demographics and characteristics.

Characteristic	Entire data (n = 114)	No apneaa (n = 55, 64%)	Mildb apnea (n = 17, 20%)	Moderatec apnea (n = 6, 7%)	Severed apnea (n = 8, 9%)
Mean age, years (SD)	43.8 (12.2)	41.52 (10)	48.5(14.7)	49.3 (12.7)	44.3 (15.8)
Sex, n (%)
Male	48 (45%)	28 (45%)	10 (41%)	5 (50%)	5 (50%)
Female	56 (55%)	33 (55%)	13 (59%)	5 (50%)	5 (50%)
Mean BMI, kg/m2 (SD)	28.9 (5.9)	27.1 (4.8)	32.4 (7.6)	32.8 (6.9)	31.7 (5.5)
Mean AHIe (SD)	8.59 (15.6)	1.2 (0.2)	6.8 (5.2)	17.2 (18.0)	30.8 (44.9)

^aNo apnea included participants with AHI < 5; ^bMild apnea included participants with AHI ≥ 5 to < 15; ^cModerate apnea included participants with AHI ≥ 15 to < 30; ^dSevere apnea included participants with AHI ≥ 30; ^eAHI was estimated by PSG; ^fApnea severity was determined by AHI estimation based on PSG measurements.

AHI, apnea-hypoxia index; BMI, body mass index; kg/m², kilograms per square meter; SD, standard deviation.

Sleep studies

The participants underwent one or two nights of simultaneous PSG and smart bed data collection within a controlled laboratory environment. Each participant slept alone on a queen-sized (60-inch × 80-inch) smart bed. The same bed size was used consistently across all participants and sleep sessions. The smart bed employs high-resolution, full-body BCG sensing from each side as a non-invasive method to measure movements resulting from motion, positional changes, breathing, and subtle heartbeats. The smart bed continuously collected data at 1 kHz, which were processed to detect bed presence and apneic events¹. The visual diagram illustrating the smart bed and data flow for the AI model is presented in Fig. 1. The image demonstrates the smart bed equipped with pressure sensors capturing the BCG signal. This signal is filtered and down-sampled to a BCG signal at 40 Hz. Specifically, we down-sampled the BCG signal to 40 Hz after applying an 8th-order Chebyshev low-pass filter with a 20 Hz cutoff frequency. This filtering attenuates high-frequency noise and prevents aliasing while preserving the relevant bio-signal information. This processed data feeds into a Deep Neural Network, which detects respiratory events categorized as apnea. These detections are aggregated for the duration of sleep to estimate AHI and the probability of a sleeper having an AHI ≥ 15.

Fig. 1.

Diagram of smart bed monitoring system to estimate AHI. AHI, apnea-hypopnea index; BCG, ballistocardiography; C, convolution layer; DNN, deep neural network; FC, fully connected layer; Hz, hertz; mm: ss, minute: second; Pa, Pascal.

During the sleep sessions, PSG data from when the lights were turned off until they were turned back on were isolated and then securely transferred to a database for subsequent analyses. The timing for lights-off and lights-on aligned precisely with the habitual sleep schedule of each participant, resulting in variation in sleep duration among individuals. To safeguard participant confidentiality, all identifiable information was removed or encrypted, ensuring data anonymity throughout the process.

Overnight PSG

The PSG recordings adhered to a standard setup, including a 6-lead electroencephalogram (EEG) from scalp regions (F3-A2, C3-A2, O1-A2, F4-A1, C4-A1, O2-A1), electrooculogram, electromyogram, electrocardiogram, thoraco-abdominal effort via respiratory impedance plethysmography, oronasal airflow monitored using both a thermistor and nasal pressure-based flow measurement, blood oxygen saturation measured via pulse oximetry, snoring monitored via audio recording, and body position data. This setup was facilitated using SomnoStar and Nihon-Kohden sleep systems.

PSG sleep stages were independently evaluated by three registered sleep technicians following American Academy of Sleep Medicine (AASM) guidelines. To ensure accuracy and consistency, we implemented a per-second majority-voting scheme: any one-second-interval labeled as apneic by two or more annotators was classified as apnea, whereas intervals marked as apneic by only a single annotator were categorized as normal breathing.

Smart bed data were synchronized with PSG data by identifying notable movements and aligning respective datasets based on the largest detected movements, ensuring acceptable alignment considering the resolution of apnea event labels, measured in seconds. All respiratory events—including obstructive apnea, obstructive hypopnea, central apnea, and mixed apnea—were scored by technicians based on AASM guidelines and labeled as respiratory events. The synchronized smart bed data, along with the consensus labels derived from the technicians’ PSG analyses, were utilized to train a deep neural network.

Data processing

The BCG signals collected from the smart bed were processed using a low-pass filter and down-sampled from 1 kHz to 40 Hz. Data were segmented into 10-second episodes of respiratory events for training and testing the model. In this manuscript, annotated apnea and hypopnea events (obstructive, central, or mixed) with variable durations (longer than 10 s) are collectively referred to as “respiratory events.” Apneic segments (“positive samples”) are defined as 10-second segments of signal encompassing a respiratory event lasting at least 5 s, whereas non-apneic segments (“negative samples”) are defined as 10-second segments containing no respiratory events. We then aggregated the detected apneic events according to the AHI definition—dividing the total count of events by the total sleep duration—to produce a predicted AHI for comparison against PSG-derived ground truth.

All sleep studies followed guidelines established by the American Academy of Sleep Medicine (AASM) for scoring and identifying sleep apnea events. Trained and certified AASM technicians performed PSG analyses, adhering to standardized criteria ensuring accurate detection of apnea and hypopnea events. These technicians applied the AASM Manual for the Scoring of Sleep and Associated Events (Version 2.6 or the latest version applicable at the time of the study) to maintain consistency and validity in apnea detection. We combined data from multiple studies to increase the sample size and enhance algorithm validation. The data from none of the subjects involved in these validation studies were used during training of the model, ensuring independence. Data for each sleep session were aggregated to generate AHI estimates consistent with the AASM definition, as well as the probability of a sleeper having AHI ≥ 15. Participants’ AHI values were subsequently assessed using standard AHI categorization methods². According to the AASM criteria, OSA severity is classified based on AHI: an AHI ≥ 5 to < 15 events/hour indicates mild severity, ≥ 15 to < 30 events/hour moderate severity, and ≥ 30 events/hour indicates severe apnea². In this study, an estimated AHI ≥ 15 was used as the cutoff indicating apnea risk³.

Model development and architecture

The apnea detection model developed for the smart bed was trained and validated using data collected from in-laboratory sleep trials involving participants distinct from those used to test the model¹. The dataset used for training and development comprised data from 54 subjects, including both healthy participants and individuals diagnosed with apnea, resulting in a total of 320 h of sleep and 1,150,987 ten-second sleep segments (1,098,975 non-apneic segments and 52,012 apneic segments). To address the resulting class imbalance, apneic epochs were up-sampled, and a class-weighted loss function was applied during training. The model architecture consists of four convolutional layers designed to extract relevant features and identify patterns associated with respiratory event episodes². Additionally, a memory (recurrent) component was integrated to capture contextual relationships across the longitudinal dimension of the time-series data³. The output from this memory component was concatenated and then flattened for further processing. The final component of the model involved a dense layer, which computed the probability that a given input sample represents either normal or abnormal breathing patterns. This output layer served as the classifier, distinguishing between the two categories based on learned features and contextual information. To fine-tune the model’s performance, hyperparameters governing the architecture and neural network functionality were optimized using grid search. This systematic approach explored various hyperparameter combinations to identify the configuration that maximized the model’s predictive performance⁴.

Evaluation of model performance

The model’s performance was assessed using several metrics, including accuracy, sensitivity, specificity, and the F1-score. The overall accuracy measured the correctness of the model in identifying both positive and negative cases (instances with apnea risk or absence thereof). Sensitivity (true positive rate) quantified the model’s ability to correctly predict apnea risk (AHI ≥ 15). Specificity (true negative rate) assessed the model’s accuracy in correctly identifying cases with sub-threshold apnea risk (AHI < 15).

The F1-score combines precision and sensitivity to provide a balanced evaluation of the model’s capability to accurately detect positive instances while minimizing false positives⁵. Specifically, the F1-score represents the harmonic mean of precision and sensitivity, providing a comprehensive assessment of the model’s binary classification performance⁵. Additionally, a confusion matrix comparing the detection performance of the smart bed algorithm against PSG-based AHI was computed for the four categories of apnea severity (no apnea, mild, moderate, and severe).

Results

Participants

The study population used for testing and reporting consisted of 104 individuals: 43 from the 2016 study (single-night), 43 from the 2020 study (single-night), and 18 from the 2023 study (10 out of 18 had two-night studies). The mean age of the study population was 43.8 years (standard deviation [SD] 12.2), mean body mass index (BMI) was 28.8 kg/m² (SD 6.3), and 45% were male (n = 48) (Table 1). In total, 1,066 cumulative hours of sleep were recorded across both sleep studies (10 participants in the 2023 study underwent two-night recordings counted as independent sessions). Of the study population, the mean AHI was 8.59 (SD 15.6). The PSG assessment identified that 66 participants (60%) had no apnea, 28 (24%) had mild apnea, 11 (10%) had moderate apnea, and 9 (9%) had severe apnea.

Apneic event classification performance

The study evaluated the model’s ability to detect individuals with supra- and sub-threshold apnea risk by reporting sensitivity, specificity, and overall accuracy at an AHI threshold of ≥ 15, indicative of moderate-to-severe cases more likely to benefit from medical intervention. Using PSG measurements from 114 studies, 21 (18%) had an AHI ≥ 15, while 93 (82%) had an AHI < 15 (Table 2). The model correctly identified 16 out of 21 (76.1%) participants with AHI ≥ 15, resulting in a sensitivity of 0.76. Furthermore, the model accurately classified 79 out of 93 (85.0%) participants with an AHI < 15, yielding a specificity of 0.85. The macro-average F1-score (average of the F1-score calculated for each class regardless of sample size) was 0.76. The overall accuracy of the smart bed in detecting individuals with an AHI ≥ 15 was 83.3%. ROC curve analysis demonstrated the smart bed’s ability to detect individuals with AHI ≥ 15 with acceptable accuracy (AUC = 0.81)¹.

Table 2.

Results of binary classification of the smart bed versus reference PSG.

		Reference
		Nonea	Apneab
Smart bed	None a	79	5
	Apnea b	14	16
	Total	93	21

^aNo apnea included participants with AHI < 15; ^bApnea included participants with AHI ≥ 15.

AHI, apnea-hypoxia index; PSG, polysomnography.

For a detailed analysis, the confusion matrix for the four classes of sleep apnea is presented in Table 3. The smart bed detected 32 participants with no apnea, 52 with mild apnea, 21 with moderate apnea, and 8 with severe apnea. Based on these detailed results, sensitivity and specificity at an AHI ≥ 5 were 0.92 and 0.43, respectively. Similarly, for AHI ≥ 30, the sensitivity was 0.50, and the specificity was 0.97. The overall macro-average F1-score across the four-class classification was 0.52. Compared to the reference PSG, the model tended to classify a higher number of individuals as having no apnea or sub-threshold apnea and fewer as severe apnea, thus providing higher positive predictive value for moderate-to-severe cases.

Table 3.

Confusion matrix for four classes of sleep apnea comparing the smart bed to the reference PSG.

		Reference
		Nonea	Mildb	Moderatec	Severed
smart bed	None a	28	3	1	0
	Mild b	26	22	2	2
	Moderate c	10	3	6	3
	Severe d	0	1	2	5
	Total	64	29	11	10

^aNo apnea included participants with AHI < 5; ^bMild apnea included participants with AHI ≥ 5 to < 15; ^cModerate apnea included participants with AHI ≥ 15 to < 30; ^dSevere apnea included participants with AHI ≥ 30.

AHI, apnea-hypoxia index; PSG, polysomnography.

In addition to severity classification analysis, metrics were computed for absolute estimation of AHI. The Root Mean Squared Error (RMSE) between estimated and PSG-based ground-truth AHIs was 10.76 events/hour, with errors of 7.8 events/hour for female participants and 16.64 events/hour for male participants. The overall Bland-Altman analysis showed a bias of 0.54 events/hour, with limits of agreement (LoA) ranging from − 20.53 to 21.61 events/hour. Stratified by sex, male participants exhibited a bias of 0.54 events/hour (LoA: −20.53 to 21.61), whereas female participants showed a bias of 6.27 events/hour with narrower LoA (−2.82 to 15.36, standard deviation ± 4.54). Figure 2 (a) illustrates a Bland–Altman plot comparing differences between estimated and PSG-derived ground-truth AHIs across all subjects, with the mean bias (solid blue line) and ± 1.96 SD limits of agreement (dashed lines). The figure (b) depicts the scatter plot of estimated versus PSG-based AHI, including the line of identity (gray) and the least-squares regression line (red dashed), with regression equation indicated. The visual illustration of the signal during apnea episodes is presented in Fig. 3. This figure provides a 90-second view comparing the BCG signal from the smart bed and the PSG signals during CSA episodes. Both raw and filtered BCG signals captured by the smart bed were contrasted with the labeled PSG data. Notably, Fig. 3 indicates that PSG-defined apnea events, marked by gray shading, align closely with detected changes in the BCG signal recorded by the smart bed.

Fig. 2.

AHI estimation Bland-Altman and regression. AHI, apnea-hypopnea index; BCG, ballistocardiography.

Fig. 3.

Example detection of CSA using the smart bed. Ab, abdomen; BCG, ballistocardiography; CSA, central sleep apnea; Ch, chest; Hz, hertz; kHz, kilohertz; O₂, oxygen; PSG, polysomnography.

Discussion

The smart bed tested in this study employs BCG readings to unobtrusively determine sleep duration, motion, heart rate (HR), and breathing rate (BR) using an inflatable bladder embedded within the mattress (Fig. 1)¹. Data collected from in-laboratory sleep studies were processed by a machine-learning algorithm to evaluate the smart bed’s capability in detecting sleep-disordered breathing (SDB) events.

The study involved a comparative analysis of the smart bed’s performance against polysomnography (PSG) in single-night sleep sessions. The findings demonstrated that the smart bed could detect apnea severity with accuracy, especially at an AHI threshold ≥ 15. These results indicate its potential for practical application in respiratory monitoring. However, improvements remain possible regarding severity detection accuracy and the overall F1-score.

At an AHI threshold of 15, the model achieved an overall accuracy of 82% in detecting apnea risk using a single night of sleep data. Nevertheless, the model tended to classify more individuals as having sub-threshold apnea risk, particularly compared to the PSG reference. The differences between the reference and the model’s outcomes could be partially due to the intentional choice of operating point. Specifically, we prioritized minimizing false positives—even if this resulted in missed detection of some milder cases—to prevent overestimating apnea risk. By tuning the model for higher precision at the moderate-to-severe threshold (AHI ≥ 15), we aimed to avoid unnecessary alarms and to reduce flagging individuals unlikely to exceed the threshold.

Further assessments are required to evaluate the model’s performance on larger cohorts, particularly involving participants with higher apnea severity. Unlike single-night observations, multi-night studies help account for variability introduced by factors such as movement (turning, repositioning) and environmental disturbances, which may artificially inflate AHI estimates. The smart bed’s non-intrusive and continuous monitoring approach naturally smooths these one-off artifacts. By averaging nightly AHI values across a week or longer, transient disturbances are reduced, allowing genuine moderate-to-severe apnea patterns to emerge more clearly. Consequently, multi-night monitoring can provide a more reliable home-based screening for clinically significant apnea. Multi-night studies may also enhance the smart bed’s ability to classify apnea severity across diverse clinical scenarios and patient conditions, thereby optimizing predictive accuracy and reliability.

Observed differences in prediction accuracy between female and male participants may originate from distinct physiological, mechanical, and behavioral factors. Typically, male participants exhibit more pronounced respiratory effort and body movements during sleep, increasing noise and variability in the ballistocardiography (BCG) signals recorded by the smart bed sensors. Conversely, female participants often demonstrate subtler respiratory patterns and reduced movement, resulting in clearer, more consistent signals that facilitate easier interpretation by algorithms. Additionally, sex-based differences in body mass distribution, thoracic structure, and respiratory biomechanics influence signal quality and algorithm sensitivity, further contributing to the observed disparity in model performance.

The observed errors primarily stemmed from signal-related rather than model-related issues. Due to the fully-body design of the ballistocardiography (BCG) sensors embedded within the smart bed, the captured signals inherently include physiological and movement-related noise. Unlike traditional intrusive monitoring methods that localize data collection directly to respiratory organs, the smart bed’s non-intrusive design may occasionally result in noisy or ambiguous signals. Consequently, ambiguous respiratory patterns may cause misclassification. Future developments in smart bed technology could specifically target these challenges to improve sleep apnea diagnosis and monitoring.

Continuous tracking of sleep and breathing patterns using the smart bed could raise greater awareness of sleep problems among users. Over time, it could facilitate streamlined healthcare processes. Commercially available sleep trackers have shown potential in promoting education, empowerment, and responsibility for lifestyle choices, thus encouraging better adherence to sleep treatments and interventions². Additionally, these devices support patient engagement in behavior change³. Using smart beds for sleep tracking could provide longitudinal data to identify potential sleep issues, offering healthcare providers actionable information to monitor the effectiveness of apnea treatments. Effortless long-term apnea monitoring, both pre- and post-diagnosis, is crucial for early detection, understanding of sleep health, comprehensive evaluations, and timely interventions. Such tools help healthcare providers identify patterns and severity, enabling informed decision-making and adjustments to treatment plans. This not only improves sleep apnea symptom management but also mitigates long-term health risks.

Commercially available sleep monitoring devices have rapidly gained popularity⁴ and have shown promise for sleep apnea screening. A study evaluating the Somnox mattress topper by Soundisleep reported a sensitivity of 88% and specificity of 91% at an AHI threshold of 15; however, their validation included only 62 subjects (with 2 excluded)⁵. Additionally, the Sleeptracker-AI Monitor by Fullpower Technologies, placed beneath the mattress and compared to PSG, demonstrated sensitivity and specificity for estimating OSA (AHI ≥ 15) of 87.3% (95% CI: 80.8–93.7%) and 85.7% (95% CI: 74.1–97.3%), respectively, with an overall accuracy of 88.1% (95% CI: 80.3–95.8%)⁶. Similarly, the Withings Sleep Analyzer, also positioned beneath the mattress, demonstrated excellent agreement with PSG-derived AHI estimation. At an AHI threshold of ≥ 15 events/hour, the device achieved a sensitivity of 88.0%, specificity of 88.6%, and an area under the receiver operating characteristic curve (AUROC) of 0.926⁷. Consistent with our findings, Edouard and colleagues noted CSA was detected more readily than OSA due to the distinct lack of movement during CSA events. These observations, along with our study’s results, highlight the potential of home-based sleep monitoring devices for unobtrusive detection of sleep apnea.

This study relied on data collected from single-night sleep sessions within a controlled laboratory environment. Future research should incorporate multi-night sleep recordings conducted in participants’ homes. Such an approach would allow a more comprehensive assessment of the smart bed’s performance and its applicability to diverse sleep patterns and real-life conditions, while preserving its non-intrusive data collection capability. Such studies could enhance insights into the model’s generalizability and effectiveness under varied, real-world sleeping conditions, thus producing more reliable detection of sleepers with AHI ≥ 15. Integrating additional sensors, along with subjective questionnaires, may further enhance the evaluation by capturing a more complete picture of sleep quality and user experience. ensuring better patient outcomes and overall health.

Conclusion

Smart beds equipped with sensor technologies offer a promising method for non-invasively assessing sleep quality, breathing patterns, and identifying potential indicators of sleep-disordered breathing (SDB). Leveraging this technology presents an opportunity to bridge the diagnostic gap, enabling early detection, personalized interventions, and improved clinical outcomes for individuals affected by SDB. Our investigation demonstrated the smart bed’s capability to detect SDB at an AHI threshold of ≥ 15, achieving an overall accuracy of 82% in predicting apnea presence. The accuracy of this initial model underscores its promise for respiratory health monitoring applications. Compared to state-of-the-art diagnostic or screening devices for sleep apnea (such as PSG or HSAT), the smart bed provides an appealing alternative because it does not require patients to wear potentially intrusive monitoring devices that might interfere with normal sleep. Moreover, it reduces the dependence on technically trained personnel for device setup or data interpretation, facilitating more frequent and longitudinal monitoring of apnea episodes across multiple nights—a strategy aligned with more informed and effective care.

Author contributions

Farzad Siyahjani: contributed to the conception and design of the research and manuscript; and the acquisition, analysis, and interpretation of data for the manuscript; drafted and reviewed the publication critically for important intellectual content; provided final approval of the version to be published; and agreed to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work were appropriately investigated and resolved.Kostiantyn Kalenyk: contributed to data analysis, result interpretation and reviewed the publication critically.Gary Garcia: contributed to data analysis, result interpretation, and manuscript writing. Saeed Babaeizadeh: contributed to the conception and design; drafted and reviewed the publication critically for important intellectual content; provided final approval of the version to be published.

Funding information

This study was funded by Sleep Number Corporation for consumer research. The study was designed for product research by the Sleep Number Corporation, and it was reviewed by the University of Chicago Institutional Review Board (IRB) for the analysis of de-identified smart bed and PSG data. The IRB gave the study exempt status on 1/24/2020 and it was approved for analysis at IRB19-1848.

Data availability

The datasets generated and/or analyzed during the current study are not publicly available due to limitations stated in the IRB approved consent form, but are available from the corresponding author on reasonable request.

Declarations

Competing interests

Farzad Siyahjani, Saeed Babaeizadeh, Gary Garcia-Molina are employees of Sleep Number Labs, Kostiantyn Kalenyk is employee of GlobalLogic Ukraine, which is a Sleep Number contracting company.