Description:SYSTEM AND METHOD OF DETECTING OBSTRUCTIVE SLEEP APNEA USING PHOTOPLETHYSMOGRAPHY SIGNALS

FIELD OF INVENTION
[0001] The present invention generally relates to data computing. More specifically, the present invention is related to a deep-learning-based approach for detecting obstructive sleep apnea (OSA) using single-channel photoplethysmography (PPG).
BACKGROUND OF THE INVENTION
[0002] The subject matter discussed in the background section should not be assumed to be prior art merely as a result of its mention in the background section. Similarly, a problem mentioned in the background section or associated with the subject matter of the background section should not be assumed to have been previously recognized in the prior art. The subject matter in the background section merely represents different approaches, which in and of themselves may also correspond to implementations of the claimed technology.
[0003] Obstructive Sleep Apnea (OSA) is a prevalent sleep disorder characterized by repetitive obstruction or narrowing of the upper airway during sleep, leading to reduced or completely halted airflow. The narrowing results in a significant drop in blood oxygen levels, triggering physiological responses such as micro-arousals and abrupt shifts in sleep patterns. Common symptoms of OSA include excessive daytime sleepiness, morning headaches, mood disturbances, forgetfulness, and frequent nighttime awakenings. Individuals with risk factors such as obesity, large tonsils, or hormonal imbalances are more susceptible to developing OSA. If left untreated, OSA can lead to severe complications, including heart failure, atrial fibrillation, other arrhythmias, and non-alcoholic steatohepatitis.
[0004] Polysomnography (PSG) is widely recognized as the gold standard for detecting OSA, as it provides a comprehensive evaluation of sleep-disordered breathing using multiple physiological signals. The PSG procedure involves simultaneous monitoring of several parameters, including airflow measurement through nasal pressure transducers and oronasal thermal sensors, respiratory effort tracking via thoracic and abdominal belts, oxygen saturation monitoring using pulse oximetry, and cardiac activity assessment via electrocardiography (ECG). Additionally, sleep architecture is analyzed using electroencephalography (EEG), electrooculography (EOG) for eye movements, and electromyography (EMG) for muscle tone. While PSG offers a detailed analysis, its extensive monitoring requirements present several challenges. The PSG procedure is highly intrusive, requiring patients to be attached to multiple sensors, which can cause discomfort and disrupt natural sleep patterns. Discomfort can result in altered sleep behavior, leading to data that may not accurately reflect the sleep state of a patient.
[0005] Moreover, PSG tests are conducted in specialized sleep laboratories, requiring trained professionals to administer and interpret the results. Dependence on dedicated facilities and professionals makes PSG expensive and limits its accessibility. Another major drawback is that PSG is not suitable for continuous long-term monitoring, as it requires a controlled environment and cannot be easily adapted for home-based or ambulatory use. This limitation makes it impractical for tracking variations in OSA severity over extended periods.
[0006] Additionally, patients with conditions that necessitate frequent movement during sleep, such as Overactive Bladder Syndrome or Restless Leg Syndrome, may experience additional challenges, as movement can disrupt signal acquisition, leading to incomplete or poor-quality data. The complex setup, high cost, and inconvenience of PSG contribute to a significant proportion of undiagnosed OSA cases, emphasizing the need for a more accessible and user-friendly diagnostic solution.
[0007] In response to the limitations of PSG, alternative methods leveraging deep-learning and machine-learning techniques must be explored for OSA detection. However, a majority of such techniques may rely on multiple physiological signals, such as nasal airflow, thoracic movement, and oxygen saturation, necessitating the use of multiple sensors. Multiple sensor requirement limits the practicality of such techniques, as multiple sensors make them less user-friendly and more difficult to integrate into wearable or home-based monitoring systems. Moreover, accuracy of such techniques remains lower than that of PSG due to incomplete feature extraction from a single physiological signal.
[0008] One of the most significant limitations of these techniques is their failure to properly account for the inherent time delay between an apneic event and the corresponding changes in PPG signals. Apneic events often induce physiological changes, such as variations in heart rate and blood oxygen levels, but such changes do not manifest instantaneously in PPG signals. Many existing models do not consider this delay, leading to poor feature representation, misclassification of apnea events, and reduced overall detection accuracy.
[0009] A critical challenge in PPG-based OSA detection is determining an appropriate time window for analyzing apneic events. Traditionally, fixed-duration windows are employed, such as 90 seconds or 5 minutes, to segment PPG data for analysis. However, fixed-window approaches introduce significant issues. Apneic events vary in duration and often occur at irregular intervals. When fixed time windows are used, an event may be split between two adjacent windows, causing the loss of crucial event data. This fragmentation affects the ability of deep-learning models to recognize clear patterns related to the onset, duration, and resolution of apnea events.
[0010] Furthermore, splitting events across multiple windows creates inconsistencies in ground truth labeling, confusing machine-learning models during training and reducing classification accuracy. Another issue with fixed-window approaches is their inability to account for the time delay between apnea onset and its corresponding physiological changes in PPG signals. If the chosen window length does not align properly with these delayed changes, crucial post-event patterns may be missed, further reducing model performance.
[0011] In light of the above-mentioned challenges, there is a long-felt need for an improved approach to OSA detection that ensures accurate event classification while maintaining the feasibility and comfort of long-term monitoring.
SUMMARY OF THE INVENTION
[0012] This summary is provided to introduce aspects related to a system and a method of detecting obstructive sleep apnea using photoplethysmography signals, and the aspects are further described below in the detailed description. This summary is not intended to identify essential features of the claimed subject matter, nor is it intended for use in determining or limiting the scope of the claimed subject matter.
[0013] In an embodiment, a method of detecting obstructive sleep apnea (OSA) using photoplethysmography (PPG) signals is disclosed. The method includes receiving PPG signals from a wearable device. Further, the method includes de-noising the PPG signals using a Butterworth bandpass filter. The method further includes segmenting the PPG signals into one or more windows using an adaptive windowing technique. Apnea events are centrally positioned within the one or more windows.
[0014] Furthermore, the method includes extracting temporal and spatial features from the PPG signals in the one or more windows using a machine learning (ML) model, and classifying the one or more windows into an apnea event and a non-apnea event, based on presence of an apnea event in each window. Further, the method includes providing an output indicating presence of the apnea event within each window.
[0015] In an aspect, segmenting further includes applying a fixed window size of 60 seconds to the one or more windows. Further, segmenting includes positioning an apnea event at center of each window to capture delays in a PPG signal that happen before and after the apnea event. Furthermore, segmenting includes adjusting the one or more windows to prevent fragmentation of the apnea event.
[0016] In another aspect, the ML model includes a Multivariate Long Short-Term Memory - Fully Convolutional Network (MLSTM-FCN) architecture. The MLSTM-FCN architecture includes a long short-term memory (LSTM) layer to capture the temporal features in the PPG signals. Further, the MLSTM-FCN architecture includes a fully convolutional network (FCN) branch to extract spatial features using a sequence of convolutional layers. Furthermore, the MLSTM-FCN architecture includes one or more squeeze and excitation (SE) blocks to adjust feature importance. The MLSTM-FCN architecture further includes a global average pooling (GAP) layer to reduce dimensionality while retaining important PPG signal features.
[0017] In another aspect, classifying further includes concatenating temporal features captured by the LSTM layer and the spatial features extracted by the FCN branch. Further, classifying includes mapping the temporal features and the spatial features to an apnea classification using a fully connected layer. The fully connected layer includes a sigmoid activation function. Further, classifying includes utilizing the sigmoid activation function to provide the output. The output classifies the one or more windows into an apnea event and a non-apnea event, based on presence of an apnea event in each window.
[0018] In yet another aspect, a 4th-order Butterworth bandpass filter is utilized to de-noise within a frequency range of 0.5–8 Hz.
[0019] In an embodiment, a system to detect OSA using PPG signals is disclosed. The system includes a processor and a memory. The memory is coupled with the processor. Further, the memory stores program instructions configured to receive PPG signals from a wearable device and de-noise the PPG signals by using a Butterworth bandpass filter. The memory further stores program instructions configured to segment the PPG signals into one or more windows. Apnea events are centrally positioned within the one or more windows.
[0020] Further, the memory stores program instructions configured to extract temporal and spatial features from the PPG signals in the one or more windows using an ML model. The memory further stores program instructions configured to classify the one or more windows into an apnea event and a non-apnea event, based on presence of an apnea event in each window. Further, the memory stores program instructions configured to provide an output indicating presence of the apnea event within each window.
[0021] In an aspect, to segment the PPG signals into one or more windows, the memory further stores program instructions configured to apply a fixed window size of 60 seconds to the one or more windows. Further, the memory stores program instructions configured to position an apnea event at center of each window to capture delays in a PPG signal that happen before and after the apnea event. Furthermore, the memory stores program instructions configured to adjust windows to prevent fragmentation of the apnea event.
[0022] In another aspect, the ML model includes a Multivariate Long Short-Term Memory - Fully Convolutional Network (MLSTM-FCN) architecture. The MLSTM-FCN architecture includes a long short-term memory (LSTM) layer to capture the temporal features in the PPG signals. Further, the MLSTM-FCN architecture includes a fully convolutional network (FCN) branch to extract spatial features using a sequence of convolutional layers. Furthermore, the MLSTM-FCN architecture includes one or more squeeze and excitation (SE) blocks to adjust feature importance. Additionally, the MLSTM-FCN architecture includes a global average pooling (GAP) layer reduces dimensionality while retaining important PPG signal features.
[0023] In one aspect, to classify the one or more windows into an apnea event and a non-apnea event, the memory further stores program instructions configured to concatenate temporal features captured by the LSTM layer and the spatial features extracted by the FCN branch and map the temporal features and the spatial features to an apnea classification using a fully connected layer. The fully connected layer includes a sigmoid activation function. The memory further stores program instructions configured to utilize the sigmoid activation function to provide the output. The output classifies the one or more windows into an apnea event and a non-apnea event, based on presence of an apnea event in each window.
[0024] In another aspect, a 4th-order Butterworth bandpass filter is utilized to de-noise within a frequency range of 0.5–8 Hz.
[0025] Other aspects and advantages of the invention will become apparent from the following description, taken in conjunction with the accompanying drawings, illustrating by way of example the principles of the invention.
BRIEF DESCRIPTION OF THE DRAWINGS
[0026] The accompanying drawings constitute a part of the description and are used to provide a further understanding of the present disclosure. In the drawings:
[0027] Figure 1 illustrates a working architecture of a system for detecting obstructive sleep apnea (OSA) using photoplethysmography (PPG) signals, in accordance with an embodiment of the present disclosure;
[0028] Figure 2 illustrates a block diagram of the system for detecting OSA using PPG signals, in accordance with an embodiment of the present disclosure;
[0029] Figure 3 illustrates an architecture of a machine learning model of the system for detecting OSA using PPG signals, in accordance with an embodiment of the present disclosure; and
[0030] Figure 4 illustrates a flow chart of a method of detecting OSA using PPG signals, in accordance with an embodiment of the present disclosure.
[0031] A more complete understanding of the present disclosure and its embodiments thereof may be acquired by referring to the following description and the accompanying drawings.
DETAILED DESCRIPTION OF THE INVENTION
[0032] Exemplary embodiments now will be described with reference to the accompanying drawings. The disclosure may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey its scope to those skilled in the art. The terminology used in the detailed description of the particular exemplary embodiments illustrated in the accompanying drawings is not intended to be limiting. In the drawings, like numbers refer to like elements.
[0033] It is to be noted, however, that the reference numerals used herein illustrate only typical embodiments of the present subject matter, and are therefore, not to be considered for limiting its scope, for the subject matter may admit to other equally effective embodiments.
[0034] The detailed description includes specific details for the purpose of providing a thorough understanding of the present invention. However, it will be apparent to those skilled in the art that the present invention may be practiced without these specific details.
[0035] The present invention aims to provide a solution for the above-mentioned challenges by introducing a deep-learning-based method for detecting Obstructive Sleep Apnea (OSA) using a single-channel Photoplethysmography (PPG) signal. Unlike traditional approaches that require multiple physiological sensors, this invention leverages PPG data, which is easily accessible through wearable devices, to provide an accurate and non-intrusive method for apnea detection.
[0036] Additionally, the invention incorporates an adaptive windowing strategy designed to preserve integrity of apnea events within a window for analysis while ensuring inclusion of sufficient pre-event and post-event data. The windowing strategy eliminates event fragmentation, enhances accuracy of ground truth labels, and minimizes model misclassification. The present invention provides a practical and cost-effective alternative to traditional polysomnography, enabling continuous long-term monitoring in home environments without the need for specialized sleep laboratories or trained personnel. By improving the feasibility and accessibility of OSA detection, the invention has the potential to address the significant percentage of undiagnosed cases and contribute to better management of sleep-disordered breathing conditions.
[0037] Figure 1 illustrates a working architecture of a system 100 for detecting obstructive sleep apnea (OSA) using photoplethysmography (PPG) signals, in accordance with an embodiment of the present disclosure. The system 100 may be connected to a wearable device 110 (labelled as PPG Sensor in Figure 1) via a communication network 120. The wearable device 110 may be configured to monitor physiological signals, particularly PPG signals of a user, for the detection of OSA. Further, the wearable device 110 may transfer the PPG signals to the system 100, via the communication network 120, for detecting OSA.
[0038] The wearable device 110 may include one of, a wristband, a fingertip sensor, a smartwatch and a ring-type sensor. The wearable device 110 may be configured to ensure user comfort during sleep. Further, the wearable device 110 may include an optical sensor (PPG sensor) that emits and detects light signals to measure changes in blood volume within microvascular tissue, providing a real-time PPG signal. The PPG sensor operates by emitting light into skin of the user and measuring variations in light absorption caused by pulsatile blood flow. The variations correspond to heart rate and blood oxygen levels of the user, which are critical indicators for detecting apnea episodes.
[0039] In an embodiment, the wearable device 110 may capture PPG signals at a predefined sampling rate, ensuring that even minor variations in pulse waveform morphology associated with apneic events are accurately recorded. Additionally, the wearable device 110 may include sensors, including but not limited to an accelerometer and a gyroscope, to track body movements and detect potential artifacts caused by excessive motion. In an embodiment, the wearable device 110 may enhance the reliability of the PPG signals by enabling signal correction and noise reduction techniques.
[0040] In an embodiment, to ensure optimal functionality, the wearable device 110 may include a processor capable of executing basic preprocessing tasks, such as filtering raw PPG signals to remove noise and artifacts. In an embodiment, filtering may be executed using a Butterworth bandpass filter implemented in the wearable device 110. The Butterworth bandpass filter may be an Analog or Digital filter. In an implementation, the Butterworth bandpass filter may be located within signal processing circuits of the wearable device. In one embodiment, the Butterworth bandpass filter may be implemented using resistors, capacitors, and operational amplifiers (op-amps) in low-pass, high-pass, band-pass, or band-stop configurations.
[0041] Further, the wearable device 110 may include a memory for temporary storage of data related to PPG signals before transmission. Additionally, the wearable device 110 may include a wireless communication module (e.g., Bluetooth, Wi-Fi) and a battery unit.
[0042] The communication network 120 may utilize network components to establish a connection between the system 100 and the wearable device 110. The network components may include hubs, switches, routers, bridges, and repeaters. The routers may be of different types, such as Provide Edge (PE) routers, Customer Edge (CE) routers, and intermediate routers.
[0043] The communication network 120 may be a wired and/or a wireless network. The communication network 120 may be implemented using communication techniques such as Visible Light Communication (VLC), Worldwide Interoperability for Microwave Access (WiMAX), Long Term Evolution (LTE), Wireless Local Area Network (WLAN), Infrared (IR) communication, Public Switched Telephone Network (PSTN), Radio waves, and other communication techniques known in the art.
[0044] The system 100 may be a remote processing unit for receiving, storing, analyzing, and processing PPG signals transmitted from the wearable device 110 using edge computing. In one embodiment, the system 100 may be implemented as a standalone physical computing device, such as a server, or as a cloud-based server infrastructure, enabling scalable and distributed data processing. The system 100 may utilize an adaptive windowing technique to ensure that each apnea event is fully contained within a single analysis window while capturing pre-event and post-event variations in the PPG signal. Traditional fixed-duration windowing approaches often result in fragmented apnea events across multiple windows, leading to misclassification and reduced model performance.
[0045] Further, the system 100 may be configured to execute a machine learning (ML) model trained to identify patterns associated with sleep apnea events based on variations in the PPG signal. Furthermore, the system 100 provides a binary output to a user device 130, indicating presence of an apnea event in an analysis window.
[0046] The user device 130 includes but is not limited to a mobile phone, a local computer, a tablet, a laptop, a smart watch, and a smart ring. Further, the user device 130 may provide a user interface on a display, which can be accessed by a user 140. The user 140 may include but is not limited to healthcare professionals, researchers and the user wearing the wearable device 110. The user device 130 may include real-time alerts, historical trend analysis, and data export capabilities for further clinical assessment, based on the output provided by the system 100.
[0047] Figure 2 illustrates a block diagram of the system 100 for detecting OSA using PPG signals, in accordance with an embodiment of the present disclosure. The system 100 may be a remote processing unit for receiving, storing, analyzing, and processing PPG signals transmitted from the wearable device 110 (as shown in Figure 1) using edge computing. In one embodiment, the system 100 may be implemented as a standalone physical computing device, such as a server, or as a cloud-based server infrastructure, enabling scalable and distributed data processing.
[0048] Further, the system 100 may include one or more network interfaces 102 (e.g., wired, wireless, etc.), at least one processor 104, and a memory 106. The one or more network interfaces 102, the at least one processor 104, and the memory 106 may be interconnected by a system bus and a power supply (not shown in the Figure 2). Further, the processor 104 is communicatively coupled with the memory 106.
[0049] The one or more network interfaces 102 may be used to provide input to or fetch output from, the system 100. The one or more network interfaces 108 may be implemented as a Command Line Interface (CLI) or a Graphical User Interface (GUI). Further, Application Programming Interfaces (APIs) may also be used for remotely interacting with edge systems and cloud servers.
[0050] The processor 104 may include one or more general purpose processors (e.g., INTEL® or Advanced Micro Devices® (AMD) microprocessors) and/or one or more special purpose processors (e.g., digital signal processors or Xilinx® System On Chip (SOC) Field Programmable Gate Array (FPGA) processor), MIPS/ARM-class processor, a microprocessor, a digital signal processor, an application specific integrated circuit, a microcontroller, a state machine, or any type of programmable logic array.
[0051] The memory 106 may include, but is not limited to, non-transitory machine-readable storage devices such as hard drives, magnetic tape, floppy diskettes, optical disks, Compact Disc Read-Only Memories (CD-ROMs), and magneto-optical disks, semiconductor memories, such as ROMs, Random Access Memories (RAMs), Programmable Read-Only Memories (PROMs), Erasable PROMs (EPROMs), Electrically Erasable PROMs (EEPROMs), flash memory, magnetic or optical cards, or other type of media/machine-readable medium suitable for storing electronic instructions.
[0052] The memory 106 may include a plurality of storage locations that are addressable by the processor 104 and the network interfaces 102 for storing software programs and other necessary information (program instructions and machine learning model 160) associated with the embodiments described herein. The processor 104 detects OSA by executing program instructions stored in the memory 106. The processor 102 may comprise hardware elements or hardware logic adapted to execute the software programs and manipulate data structures.
[0053] It will be apparent to those skilled in the art that other processor and memory types, including various computer-readable media, may be used to store and execute program instructions pertaining to the techniques described herein. Also, while the description illustrates various processes, it is expressly contemplated that various processes may be embodied as modules configured to operate in accordance with the techniques herein (e.g., according to the functionality of a similar process). Further, while the processes have been shown separately, those skilled in the art will appreciate that processes may be routines or modules within other processes.
[0054] The processor 104 of the system 100 is configured to receive PPG signals from the wearable device 110, via the communication network 120 (as shown in Figure 1). Further, the processor 104 utilizes the one or more network interfaces 102 of the system 100 to receive the PPG signals from the communication network 120.
[0055] In one embodiment, the processor 104 may be configured to de-noise the PPG signals by using a Butterworth bandpass filter to enhance signal quality and improve the accuracy of apnea detection. The Butterworth bandpass filter may be a 4th-order Butterworth bandpass filter configured to de-noise within a predefined frequency range. Furthermore, the Butterworth bandpass filter may be at least one of, a hardware based Analog Butterworth filter, a software-based Digital Butterworth filter.
[0056] The predefined frequency range may be 0.5–8 Hz. The frequency range of 0.5–8 Hz is selected based on spectral characteristics of PPG signals related to sleep apnea detection. The lower cut-off frequency of 0.5 Hz eliminates low-frequency drift and baseline wandering, which can result from sensor movement, respiration, or slow variations in blood volume. The upper cut-off frequency of 8 Hz removes high-frequency noise, such as power line interference, muscle artifacts, and random electronic noise from the sensor. By restricting the PPG signal to this specific range, the filter may ensure that essential cardiovascular and respiratory components relevant for apnea detection are retained while non-informative or disruptive signal components are suppressed.
[0057] In an embodiment, the processor 104 may be configured to adaptively adjust filtering process based on signal quality metrics. For instance, if excessive motion artifacts or poor-quality signals are detected, the processor 104 may apply additional filtering techniques or request re-acquisition of the PPG signals to maintain integrity of apnea detection.
[0058] In an embodiment, the filtering process ensures that only high-quality signals contribute to apnea detection by selecting minimum 6 hours of recorded data to capture sufficient sleep patterns, and signal quality rating of at least 95% (based on 'qupleth5' parameter) to reduce noise interference, and an ‘outstanding’ overall rating in the ‘overall5’ quality metric to eliminate unreliable recordings.
[0059] Further, the processor 104 may be configured to segment the PPG signals into one or more windows using an adaptive windowing technique to facilitate accurate apnea detection while preserving temporal relationship between apnea events and corresponding changes in the PPG signal. The adaptive windowing technique ensures that apnea events are centrally positioned within each window.
[0060] Conventional windowing techniques use fixed segmentation intervals, where the PPG signal is divided into consecutive windows without regard to event boundaries. An apnea event may be split between two consecutive windows, leading to ambiguity in classification and increased confusion for a machine learning model. Further, changes in PPG signals due to an apnea event do not occur instantaneously but rather exhibit a delay. If an apnea event is positioned towards the end of a window in a conventional approach, corresponding delayed PPG variations may appear in next window, resulting in a misalignment between input data and ground-truth labels.
[0061] To address the above-mentioned challenges of conventional approaches, the processor 104 implements the adaptive windowing technique incorporating 60-second windows, ensuring that every apnea event is centrally located within its respective window. The rationale behind the 60-seconds duration is based on a statistical analysis of apnea events, which revealed that 41.7% of apnea events and 10% of hypopnea events last longer than 30 seconds, with a mean apnea duration of 21.8 seconds. A 60-second window may provide a sufficient temporal buffer before and after the apnea event, allowing the system 100 to capture the apnea event and delayed physiological changes in the PPG signal.
[0062] The processor 104 may adjust window boundaries to ensure complete apnea events remain within a single segment/window. If an apnea event of length x seconds is detected, the processor 104 calculates a pre-event and post-event buffer of y=(60−x)/2 seconds to maintain symmetrical padding around the apnea event. The adaptive windowing technique prevents fragmentation of apnea events across multiple windows and ensures that all relevant PPG variations associated with apnea are contained within the same segment/window.
[0063] By centralizing apnea events within windows, the system 100 may improve reliability of ground-truth labels, reducing risk of misclassification and improving accuracy of machine-learning-based apnea detection. The adaptive windowing technique ensures that normal windows contain only normal breathing patterns, while apnea windows consistently feature a clearly defined apneic event at their center, creating a well-structured dataset for training and inference of machine learning models.
[0064] The processor 104 may execute a machine learning (ML) model 160 stored in the memory 106 to detect OSA in the one or more windows. The ML model 160 may be a Multivariate Long Short-Term Memory - Fully Convolutional Network (MLSTM-FCN) model 160, as illustrated in Figure 3.
[0065] In an embodiment, the ML model 160 may be trained and validated using Multi-Ethnic Study of Atherosclerosis (MESA) dataset. The MESA dataset comprises PPG recordings from about 2056 subjects of diverse ethnic backgrounds, including Black, White, Hispanic, and Chinese-American individuals, aged between 45-84 years. To ensure the quality of data used for training, a rigorous data pre-processing pipeline may be implemented. The PPG signals from the MESA dataset may be merged with their corresponding annotation files, after which the signals may be labeled and categorized into three classes: obstructive apnea, hypopnea, and normal. However, for the present invention, obstructive apnea and hypopnea may be treated as a single class, referred to as apnea.
[0066] The MLSTM-FCN framework is specifically designed to leverage temporal dependencies and local spatial patterns in PPG signals, making it accurate for apnea detection. The MLSTM-FCN model 160 employs a dual-branch architecture that includes an LSTM layer 162 to extract temporal features from the PPG signals. The LSTM layer 162 may identify long-term variations in the PPG signal due to apnea-related disturbances. In an embodiment, the LSTM layer 162 may be configured with 8 units, captures dependencies across time, ensuring that even delayed responses in the PPG signal are recognized.
[0067] Further, the LSTM layer 162 may include an attention mechanism to assign higher importance to critical segments of the PPG signal, allowing the ML model 160 to focus on relevant patterns related to OSA events. The attention mechanism may function by analyzing sequential data and identifying portions of the PPG signal that exhibit patterns indicative of apnea events. The attention mechanism ensures that the ML model 160 focuses on most relevant time intervals, rather than treating all input data equally. Given that apneic events often result in delayed changes in PPG signals, the attention mechanism improves ability of the ML model 160 to distinguish between normal and abnormal patterns by weighting critical segments more heavily. By doing so, the attention mechanism enhances the learning process of the MLSTM-FCN architecture, allowing it to capture both short-term and long-term dependencies more effectively.
[0068] Further, the MLSTM-FCN model 160 employs a 1D fully convolutional network branch 164 (FCN) for local spatial feature extraction. In an embodiment, the FCN 164 includes a first convolutional layer with 128 filters and a kernel size of 8, a second convolutional layer with 256 filters and a kernel size of 5, and a third convolutional layer with 128 filters and a kernel size of 3.
[0069] To further refine feature selection, squeeze-and-excitation (SE) blocks 166 are incorporated after each convolutional layer 164. The SE blocks 166 reweight feature maps and adjusts feature importance, emphasizing most informative aspects of the PPG signal while suppressing less relevant noise. Following convolution, the ML model 160 may apply a global average pooling layer 168 (GAP), which reduces feature dimensionality while retaining crucial apnea-related information to prevent overfitting.
[0070] Outputs of both the LSTM layer 162 and FCN 164 branches are concatenated to form a comprehensive feature representation, combining long-term dependencies with localized spatial signal variations. Further, a fully connected output layer 170 with a sigmoid activation function predicts and provides an output indicating a probability of an apnea event occurring within the one or more windows and classifies each window based on the probability, as an apnea event (A) or a normal event (N), as shown in Figure 3. In an embodiment, the ML model 160 achieves a test accuracy of 93.44%, outperforming traditional methods by 11.3%. The ML model 160 further records a precision of 0.94, recall of 0.91, specificity of 0.95, and an F1-score of 0.93.
[0071] The processor 104 may be further configured to provide the output indicating presence of the apnea event within each classified window to the user device 130, allowing the user 140 to remotely track sleep patterns of patient. The real-time transmission enables early intervention, reducing health risks associated with undiagnosed or untreated sleep apnea.
[0072] In an embodiment, the system 100 may incorporate an adaptive thresholding mechanism. The adaptive thresholding mechanism may adjust a detection threshold based on historical sleep data of the user, improving personalized accuracy over time. By continuously learning from past apnea patterns, the system 100 minimizes false positives and false negatives, making it highly reliable for long-term use.
[0073] Figure 4 illustrates a flow chart of a method 400 of detecting OSA using PPG signals, in accordance with an embodiment of the present invention. In this regard, each block may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that in some alternative implementations, the functions noted in the blocks may occur out of the order noted in the drawings.
[0074] For example, two blocks shown in succession in Figure 4 may in fact be executed substantially concurrently or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. Any process descriptions or blocks in flow charts should be understood as representing modules, segments, or portions of code which include one or more executable instructions for implementing specific logical functions or steps in the process, and alternate implementations are included within the scope of the example embodiments in which functions may be executed out of order from that shown or discussed, including substantially concurrently or in reverse order, depending on the functionality involved. In addition, the process descriptions or blocks in flow charts should be understood as representing decisions made by a hardware structure such as a state machine.
[0075] The order in which the method is described is not intended to be construed as a limitation, and any number of the described method blocks can be combined in any order to implement the method. Additionally, individual blocks may be deleted from the methods without departing from the scope of the subject matter described herein.
[0076] Furthermore, the method can be implemented in any suitable hardware, software, firmware, or combination thereof. Furthermore, the above-mentioned methods may be implemented in suitable hardware, computer-readable instructions, or a combination thereof. The steps of such methods may be performed by either a system under the instruction of machine-executable instructions stored on a non-transitory computer-readable medium or by dedicated hardware circuits, microcontrollers, or logic circuits. The method may include the following steps.
[0077] At step 402, PPG signals may be received from a wearable device at a system for detecting OSA. The wearable device may include one of, a wristband, a fingertip sensor, a smartwatch and a ring-type sensor. The wearable device may be configured to ensure user comfort during sleep.
[0078] Further, the wearable device may include an optical sensor (PPG sensor) that emits and detects light signals to measure changes in blood volume within microvascular tissue, providing a real-time PPG signal. The PPG sensor operates by emitting light into skin of a user and measuring variations in light absorption caused by pulsatile blood flow. The variations correspond to heart rate and blood oxygen levels of the user, which are critical indicators for detecting apnea episodes.
[0079] In an embodiment, the wearable device may capture PPG signals at a predefined sampling rate, ensuring that even minor variations in pulse waveform morphology associated with apneic events are accurately recorded. Additionally, the wearable device may include sensors, including but not limited to an accelerometer and a gyroscope, to track body movements and detect potential artifacts caused by excessive motion. In an embodiment, the wearable device may enhance the reliability of the PPG signals by enabling signal correction and noise reduction techniques.
[0080] Further, the wearable device may include a memory for temporary storage of data related to PPG signals before transmission. Additionally, the wearable device may include a wireless communication module (e.g., Bluetooth, Wi-Fi) and a battery unit. Further, the wearable device may transfer the PPG signals to the system, via the communication network, for detecting OSA.
[0081] The system may be a remote processing unit for receiving, storing, analyzing, and processing PPG signals transmitted from the wearable device using edge computing. In one embodiment, the system may be implemented as a standalone physical computing device, such as a server, or as a cloud-based server infrastructure, enabling scalable and distributed data processing.
[0082] At step 404, the PPG signals may be de-noised using a Butterworth bandpass filter. In an embodiment, to ensure optimal functionality, the wearable device may include a processor capable of executing basic preprocessing tasks, such as filtering raw PPG signals to remove noise and artifacts. In an embodiment, filtering may be executed using a Butterworth bandpass filter implemented in the wearable device.
[0083] The Butterworth bandpass filter may be an Analog or Digital filter. In an implementation, the Butterworth bandpass filter may be located within signal processing circuits of the wearable device. In one embodiment, the Butterworth bandpass filter may be implemented using resistors, capacitors, and operational amplifiers (op-amps) in low-pass, high-pass, band-pass, or band-stop configurations.
[0084] At step 406, the PPG signals may be segmented into one or more windows using an adaptive windowing technique to facilitate accurate apnea detection while preserving temporal relationship between apnea events and corresponding changes in the PPG signal. The adaptive windowing technique ensures that apnea events are centrally positioned within each window.
[0085] Conventional windowing techniques use fixed segmentation intervals, where the PPG signal is divided into consecutive windows without regard to event boundaries. An apnea event may be split between two consecutive windows, leading to ambiguity in classification and increased confusion for a machine learning model. Further, changes in PPG signals due to an apnea event do not occur instantaneously but rather exhibit a delay. If an apnea event is positioned towards the end of a window in a conventional approach, corresponding delayed PPG variations may appear in next window, resulting in a misalignment between input data and ground-truth labels.
[0086] To address the above-mentioned challenges of conventional windowing techniques, the system implements the adaptive windowing technique incorporating 60-second windows, ensuring that every apnea event is centrally located within its respective window. The rationale behind the 60-seconds duration is based on a statistical analysis of apnea events, which revealed that 41.7% of apnea events and 10% of hypopnea events last longer than 30 seconds, with a mean apnea duration of 21.8 seconds. A 60-second window may provide a sufficient temporal buffer before and after the apnea event, allowing the system to capture the apnea event and delayed physiological changes in the PPG signal.
[0087] The system may adjust window boundaries to ensure complete apnea events remain within a single segment/window. If an apnea event of length x seconds is detected, the system calculates a pre-event and post-event buffer of y=(60−x)/2 seconds to maintain symmetrical padding around the apnea event. The adaptive windowing technique prevents fragmentation of apnea events across multiple windows and ensures that all relevant PPG variations associated with apnea are contained within the same segment/window.
[0088] At step 408, temporal and spatial features from the PPG signals in the one or more windows may be extracted using a machine learning (ML) model. The ML model may be a Multivariate Long Short-Term Memory - Fully Convolutional Network (MLSTM-FCN) model.
[0089] In an embodiment, the ML model may be trained and validated using Multi-Ethnic Study of Atherosclerosis (MESA) dataset. The MESA dataset comprises PPG recordings from about 2056 subjects of diverse ethnic backgrounds, including Black, White, Hispanic, and Chinese-American individuals, aged between 45-84 years. To ensure the quality of data used for training, a rigorous data pre-processing pipeline may be implemented. The PPG signals from the MESA dataset may be merged with their corresponding annotation files, after which the signals may be labeled and categorized into three classes: obstructive apnea, hypopnea, and normal. However, for the present invention, obstructive apnea and hypopnea may be treated as a single class, referred to as apnea.
[0090] The MLSTM-FCN framework is specifically designed to leverage temporal dependencies and local spatial patterns in PPG signals, making it accurate for apnea detection. The MLSTM-FCN model employs a dual-branch architecture that includes an LSTM layer to extract temporal features from the PPG signals. The LSTM layer may identify long-term variations in the PPG signal due to apnea-related disturbances.
[0091] In an embodiment, the LSTM layer may be configured with 8 units, captures dependencies across time, ensuring that even delayed responses in the PPG signal are recognized. Further, the LSTM layer may include an attention mechanism to assign higher importance to critical segments of the PPG signal, allowing the ML model to focus on relevant patterns related to OSA events. The attention mechanism may function by analyzing sequential data and identifying portions of the PPG signal that exhibit patterns indicative of apnea events. The attention mechanism ensures that the ML model focuses on most relevant time intervals, rather than treating all input data equally. Given that apneic events often result in delayed changes in PPG signals, the attention mechanism improves ability of the ML model to distinguish between normal and abnormal patterns by weighting critical segments more heavily. By doing so, the attention mechanism enhances the learning process of the MLSTM-FCN architecture, allowing it to capture both short-term and long-term dependencies more effectively.
[0092] Further, the MLSTM-FCN model may employ a 1D fully convolutional network branch (FCN) for local spatial feature extraction. In an embodiment, the FCN includes a first convolutional layer with 128 filters and a kernel size of 8, a second convolutional layer with 256 filters and a kernel size of 5, and a third convolutional layer with 128 filters and a kernel size of 3.
[0093] To further refine feature selection, squeeze-and-excitation (SE) blocks may be incorporated after each convolutional layer. The SE blocks reweight feature maps and adjusts feature importance, emphasizing most informative aspects of the PPG signal while suppressing less relevant noise. Following convolution, the ML model may be configured to apply a global average pooling layer (GAP), which reduces feature dimensionality while retaining crucial apnea-related information to prevent overfitting.
[0094] At step 410, the one or more windows are classified into an apnea event and a non-apnea event, based on presence of an apnea event in each window. To classify the windows, outputs of both the LSTM layer and FCN branches are concatenated to form a comprehensive feature representation, combining long-term dependencies with localized spatial signal variations. Further, a fully connected output layer with a sigmoid activation function predicts and provides an output indicating a probability of an apnea event occurring within the one or more windows. The output classifies each window based on the probability, as an apnea event or a normal event.
[0095] At step 412, the system may provide the output indicating presence of the apnea event within each classified window to the user device, allowing the user to remotely track sleep patterns of patient. The real-time transmission enables early intervention, reducing health risks associated with undiagnosed or untreated sleep apnea.
[0096] The user device includes but is not limited to a mobile phone, a local computer, a tablet, a laptop, a smart watch, and a smart ring. Further, the user device may provide a user interface on a display, which can be accessed by a user. The user may include but is not limited to healthcare professionals, researchers and the user wearing the wearable device. The user device may include real-time alerts, historical trend analysis, and data export capabilities for further clinical assessment, based on the output provided by the system.

Technical Advancement and Economic Significance
[0097] The system and the method disclosed in the present invention of detecting obstructive sleep apnea (OSA) using photoplethysmography (PPG) signals may have the following advantages over conventional art:
- Accurate OSA detection is achieved using a Multivariate Long Short-Term Memory - Fully Convolutional Network (MLSTM-FCN) model, improving classification performance compared to traditional methods.
- Efficient de-noising of PPG signals is performed using a 4th-order Butterworth bandpass filter, ensuring high-quality input data for OSA detection and reduction in PPG signal processing overhead.
- Adaptive windowing technique ensures that apnea events remain centrally positioned within windows, reducing misclassification errors.
- Parallel feature extraction using LSTM and convolutional layers enables the model to capture both temporal and spatial patterns in PPG signals, enhancing detection accuracy.
- Reduced dependency on specialized sleep clinics lowers patient costs and enhances accessibility, making OSA diagnosis widely accessible.
- Cost-effective solution is provided by eliminating the need for complex polysomnography (PSG) setups, making OSA detection more accessible.
[0098] The specification may refer to “an”, “another”, “one” or “some” embodiment(s) in several locations.
[0099] This does not necessarily imply that each such reference is to the same embodiment(s), or that the feature only applies to a single embodiment. Single features of different embodiments may also be combined to provide other embodiments.
[0100] The terms “or” and “and/or” as used herein are to be interpreted as inclusive or meaning any one or any combination. Therefore, “A, B or C” or “A, B and/or C” mean “any of the following: A; B; C; A and B; A and C; B and C; A, B and C.” An exception to this definition will occur only when a combination of elements, functions, steps or acts are in some way inherently mutually exclusive.
[0101] As used herein, the singular forms “a”, “an” and “the” are intended to include the plural forms as well, unless expressly stated otherwise. It will be further understood that the terms “includes”, “comprises”, “including” and/or “comprising” when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof. It will be understood that when an element is referred to as being “connected” or “coupled” to another element, it can be directly connected or coupled to the other element or intervening elements may be present. Furthermore, “connected” or “coupled” as used herein may include operatively connected or coupled. As used herein, the term “and/or” includes any and all combinations and arrangements of one or more of the associated listed items.
[0102] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skills in the art to which this disclosure pertains. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.
[0103] Although implementations of a system and a method of detecting obstructive sleep apnea using photoplethysmography signals have been described in language specific to structural features and/or methods, it is to be understood that the appended claims are not necessarily limited to the specific features or methods described. Rather, the specific features and methods are disclosed as examples of implementations of a system and a method of detecting obstructive sleep apnea using photoplethysmography signals.
[0104] The invention has been described above with reference to numerous embodiments and specific examples. Many variations will suggest themselves to those skilled in this art in light of the above-detailed description. All such obvious variations are within the full intended scope of the appended claims.
, Claims:1. A method (400) of detecting obstructive sleep apnea (OSA) using photoplethysmography (PPG) signals, comprising:
receiving PPG signals from a wearable device;
de-noising the PPG signals by using a Butterworth bandpass filter;
segmenting the PPG signals into one or more windows using an adaptive windowing technique, wherein apnea events are centrally positioned within the one or more windows;
extracting temporal and spatial features from the PPG signals in the one or more windows using a machine learning (ML) model;
classifying the one or more windows into an apnea event and a non-apnea event, based on presence of an apnea event in each window; and
providing an output indicating presence of the apnea event within each window.

2. The method (400) as claimed in claim 1, wherein segmenting further comprising:
applying a fixed window size of 60 seconds to the one or more windows;
positioning an apnea event at center of each window to capture delays in a PPG signal that happen before and after the apnea event; and
adjusting the one or more windows to prevent fragmentation of the apnea event.

3. The method (400) as claimed in claim 1, wherein the ML model comprises a Multivariate Long Short-Term Memory - Fully Convolutional Network (MLSTM-FCN) architecture, and wherein:
a long short-term memory (LSTM) layer captures the temporal features in the PPG signals;
a fully convolutional network (FCN) branch extracts spatial features using a sequence of convolutional layers;
one or more squeeze and excitation (SE) blocks adjust feature importance; and
a global average pooling (GAP) layer reduces dimensionality while retaining important PPG signal features.

4. The method (400) as claimed in claim 1, wherein classifying further comprising:
concatenating temporal features captured by the LSTM layer and the spatial features extracted by the FCN branch;
mapping the temporal features and the spatial features to an apnea classification using a fully connected layer, wherein the fully connected layer comprises a sigmoid activation function;
utilizing the sigmoid activation function to provide the output classifying the one or more windows into an apnea event and a non-apnea event, based on presence of an apnea event in each window.

5. The method as claimed in claim 1, wherein a 4th-order Butterworth bandpass filter is utilized to de-noise within a frequency range of 0.5–8 Hz.

6. A system (100) to detect obstructive sleep apnea (OSA) using photoplethysmography (PPG) signals, comprises:
a processor (104); and
a memory (106) coupled with the processor (104), wherein the memory (106) stores a machine learning (ML) model (160) and program instructions configured to:
receive PPG signals from a wearable device (110);
de-noise the PPG signals by using a Butterworth bandpass filter;
segment the PPG signals into one or more windows, wherein apnea events are centrally positioned within the one or more windows;
extract temporal and spatial features from the PPG signals in the one or more windows using the machine learning (ML) model (160);
classify the one or more windows into an apnea event and a non-apnea event, based on presence of an apnea event in each window; and
provide an output indicating presence of the apnea event within each window.

7. The system (100) as claimed in claim 6, wherein to segment the PPG signals into one or more windows, the memory (106) further stores program instructions configured to:
apply a fixed window size of 60 seconds to the one or more windows;
position an apnea event at center of each window to capture delays in a PPG signal that happen before and after the apnea event; and
adjust windows to prevent fragmentation of the apnea event.

8. The system (100) as claimed in claim 6, wherein the ML model (160) comprises a Multivariate Long Short-Term Memory - Fully Convolutional Network (MLSTM-FCN) architecture, and wherein:
a long short-term memory (LSTM) layer captures the temporal features in the PPG signals;
a fully convolutional network (FCN) branch extracts spatial features using a sequence of convolutional layers;
one or more squeeze and excitation (SE) blocks adjust feature importance; and
a global average pooling (GAP) layer reduces dimensionality while retaining important PPG signal features.

9. The system (100) as claimed in claim 6, wherein to classify the one or more windows into an apnea event and a non-apnea event, the memory (106) further stores program instructions configured to:
concatenate temporal features captured by the LSTM layer and the spatial features extracted by the FCN branch;
map the temporal features and the spatial features to an apnea classification using a fully connected layer, wherein the fully connected layer comprises a sigmoid activation function;
utilize the sigmoid activation function to provide the output classifying the one or more windows into an apnea event and a non-apnea event, based on presence of an apnea event in each window.

10. The system (100) as claimed in claim 6, wherein a 4th-order Butterworth bandpass filter is utilized to de-noise within a frequency range of 0.5–8 Hz.