Screening method and system for sleep apnea disorder based on peripheral oxygen saturation, and, computer readable means

Abstract

The invention relates to a computer implemented screening method for Sleep Apnea Disorder, SA, based on a peripheral oxygen saturation, SpO2, signal comprising the steps of: receiving a SpO2 signal (100) of a user to perform preprocessing operations on the SpO2 signal (100); extracting representative SpO2 features; forming a Features Sequence (303); detecting respiratory events within the Features Sequence (303) using a modified Temporal Convolutional Network, TCN, (400) backbone; generating a Respiratory Events Sequence (402) of probabilities; aggregating the Respiratory Events Sequence (402) to generate an estimation of an Apnea-Hypopnea Index, eAHI, (105). The present invention also relates to a system for screening for Sleep Apnea Disorder, SA, based on a peripheral oxygen saturation, SpO2, signal comprising: a Signal Preprocessing Module (101) to perform preprocessing operations on a SpO2 signal (100); a Feature Extraction Module (102) configured to extract representative SpO2 features and form a Features Sequence (303); a Respiratory Event Detection Module (103) configured to detect respiratory events within the Features Sequence (303) by using a modified Temporal Convolutional Network, TCN, (400) backbone and to generate a Respiratory Events Sequence (402) of probabilities; and a AHI Estimation Module (104) configured to aggregate the Respiratory Events Sequence (402) to generate an estimation of an Apnea-Hypopnea Index, eAHI, (105).

Description

SCREENING METHOD AND SYSTEM FOR SLEEP APNEA DISORDER BASED ON PERIPHERAL OXYGEN SATURATION, AND, COMPUTER READABLE MEANSFIELD OF THE DISCLOSURE

[0001] The present invention relates to the field of Sleep Apnea Disorder (SA) and, more specifically, to a method and system for estimating an Apnea- Hypopnea Index (AHI) using peripheral oxygen saturation (SpO2) measurements.DESCRIPTION OF RELATED ART

[0002] Sleep Apnea Disorder (SA) is a sleep disorder characterized by temporary cessations of breathing during sleep. Undiagnosed, SA results in mild to severe consequences to a person’s health, for instance, daytime sleepiness (which could cause labor and traffic accidents, for example) or increased risk of vascular disease. The Apnea-Hypopnea Index - defined as the average number of apneic events per-hour of REM and non-REM sleep - is used by physicians in both diagnosing and determining the severity of SA on suspected patients. In clinical settings, detection of apneic events to compute the AHI is done by means of a Polysomnography (PSG), an expensive and uncomfortable overnight sleep study.

[0003] Non-invasive, reliable, at-home prediction of the Apnea-Hypopnea Index (AHI) can be a valuable triaging tool for wearable device users who suspect they suffer from a Sleep Apnea Disorder to seek further medical attention. They can be valuable even for general users who do not have any suspicious of having a sleep disorder. More than 20% of the world’s population may suffer from SA and a large portion of this is undiagnosed. While there are methods to estimate AHI from Peripheral-Oxygen Saturation (SpO2), most of them work with golden-standard transmissive sensors for computing the SpO2, and are deployed on expensive, specialized devices. Furthermore, many of such solutions also require other signals (e.g., respiratory signals) to compute a good AHI estimate. The invention describedherein is generic, lightweight and fast, being able to be deployed to a plurality of devices, such as consumer wrist-worn devices (e.g., smartwatches), provided that such devices can compute a SpO2 signal.

[0004] Patent application WO2022221487A1 describes a health monitoring system comprising a wearable band or ring device with a pulse oximetry sensor at the inner surface of a band worn around a user's extremity, collecting data on heart rate and blood oxygen levels at short intervals. This data can be used for sleep apnea screening, detection, diagnosis, or monitoring. The wearable band device collects data rapidly, e.g., every 1-3 seconds or less. An electronic control unit (ECU) connects to the pulse oximetry sensor, receiving collected data. Calculated blood desaturation metrics include the oxygen desaturation index (ODI), average oxygen saturation for a sleep event, nadir point for a night, and total sleep time less than 90% saturation (TST90). Settings adjustable via a mobile app include blood oxygen level and heart rate reminders or alerts triggering haptic feedback if an apnea event is detected, oxygenation level drops below a certain threshold, or heart rate falls outside the desired range.

[0005] Patent application US2022087609A1 offers a technique for measuring a person's photoplethysmogram (PPG) using an in-ear audio device equipped with a PPG sensor. The method involves collecting data from the sensor and applying various techniques to estimate the wearer's PPG. Additionally, it can detect sleep apnea events based on respiratory rate and peripheral capillary oxygen saturation levels. The ear canal's natural isolation from motion and ambient light, combined with rich blood vessel density, allows for more precise PPG readings than wrist-based sensors. Signal processing methods and machine learning algorithms, like recurrent or temporal convolutional neural networks, are employed to extract PPG signals and derive biometric parameters including sleep apnea events.

[0006] Patent application US2023346302A1 focuses on a method for classifying the severity of Obstructive Sleep Apnea (OSA) using a recording-based Peripheral Oxygen Saturation Signal (SpO2 signal) as input to detect four categories of OSA severity. Public datasets provide SpO2 signals for training a four-category OSA severity recognition model, which accepts a recording-based whole SpO2 signal and directly outputs a recognized OSA severity category (normal, mild, moderate, or severe). The model utilizes a convolutional neural network for feature map extraction, followed by a global average pooling method to handle varying input signal lengths and produce consistent output signal lengths.

[0007] Patent application JP2024072968A aims to facilitate test measurements using a pulse oximeter, a program, and a biological information measuring system. To address power consumption concerns, the pulse oximeter comprises a measuring unit for blood oxygen saturation, a clock unit, and a first control unit that manages the device's sleep state based on user input. The first control unit maintains the clock unit's operation even when transitioning from sleep release to sleep mode. The bioinformation measuring system consists of a pulse oximeter and a respiratory sensor that measures a subject's respiratory data, with the pulse oximeter featuring a communication unit for wireless communication with the respiratory sensor and a time setting unit for synchronizing their timing. The respiratory sensor can be attached near the nose, throat, or chest, and is used to assess the subject's breathing state while they sleep.

[0008] Patent application US20210345949A1 estimates the AHI based on 3 wave lengths of PPG light, and accelerometer signals from wearable devices. First, several other signals are derived from the PPG light wave lengths (e.g., SpO2, variation of AC component of PPG, variation of DC component, etc.), and the motion index is derived from the accelerometer. From the derived signals, a set of features are extracted generating sequences of features with sampling rates around 1Hz. The feature sequences are fed to a 1DCNN that estimates the density for each time window. The densities of apnea events for each time window are aggregated through the mean computation to generate an AHI estimate.

[0009] Patent application US20120296182A1 focuses on monitoring the severity of sleep apnea utilizing an oxygen saturation signal, specifically nocturnal oxygen saturation. The technique incorporates both frequency- domain and timedomain features extraction from the entire input signal, and applies dimensionality reduction to the features using techniques such as PCA. The bank of features employed for nocturnal SpO2 consists in several features from the literature on automatic sleep apnea detection, as our work does. The core algorithm consists of a multilinear regression module or a multilayer perceptron network, targeting the regression of AHI as a mean to estimate sleep apnea severity.

[0010] Patent US5891023A uses pulse oximetry to detect desaturations and resaturations in oxygen levels. A phasic desaturation event occurs when the combined duration of desaturation and resaturation is less than 3.5 minutes, and the slopes fall within specific ranges. The number of probable apneic events is determined by the number of phasic desaturation events with specific ascending to descending slope ratios.

[0011] Patent application JP2016007243A is based on expiratory flow rate measurements, and normalizes respiratory signals due to variable individual amplitudes and changes during sleep stages and body movements. The body movements and sleep stages classification are based on heart rate signals. Then, based on apnea definition (breathlessness over 10 seconds) and hypopnea definition (reducing breath intake by 50% or more for over 10 seconds), it calculates the normalized difference every second for ten seconds, identifying hypopnea when this difference is below 50% of the range and apnea when this difference is below 10%.

[0012] Patent US2023122156A1 receives pulse oximetry data from a wearable device, generates oximetry characteristics using parameter thresholds, and employs an apnea and hypoxia machine learning service to extract and select feature, comparing with patient information. Various classification algorithms can used, such as neural networks, gradient boosting, and ensemble algorithms. The oximetry application sends sleep apnea indicators to the machine learning service, which queries a trained model for sleep apnea classifications, like apnea and severity indicators.

[0013] Zhou et al. in the paper “Comparison ofOPPO Watch Sleep Analyzer and Polysomnography for Obstructive Sleep Apnea Screening” proposes OPPO Watch Sleep Analyzer (OWSA) which is an integrated system comprising a smartwatch and the HeyTap Health App for monitoring sleep and assessing the risk of OSA. The smartwatch automatically begins sleep tracking upon detecting sleep onset, measuring vital signs such as sleep stages, pulse rate variability (PRV), respiratory, and SpO2. With the phone's microphone enabled, snoring can also be tracked. The HeyTap Health app then analyzes the collected data to calculate the risk of OSA leveraging deep learning models with CNN and RNN to estimate apnea and hypopnea event densities within a 15-minute window. Statistical features, including total sleep time, mean SpO2, heart rate, and snoring frequency, are derived from continuous feature sequences extracted throughout the night. Finally, a classification model categorizes participants into different OSA severity levels based on consolidated results incorporating extracted features, participant information, and preliminary respiratory event index estimates.

[0014] In the paper “A Single-Center Validation of the Accuracy of a Photoplethysmography -Based Smartwatch for Screening Obstructive Sleep Apnea” by Chen et al., a preliminary screening model for sleep apnea was created using an overnight PPG signal containing green, infrared, and red light sources, applyingmachine learning algorithms. Green light signals were utilized to obtain pulse rate variability (PRV) characteristics, while infrared and red light signals were employed to estimate and extract blood oxygen saturation data. Sleep apnea risk was assessed by building a classification model based on PRV and blood oxygen saturation traits. Furthermore, wrist acceleration signals were used to screen signals effectively and identify unusual events. Respiratory waveforms were subsequently derived from PPG signals, sleep duration was noted, and AHI was computed.

[0015] Papini et al. , in the paper “Wearable monitoring of sleep-disordered breathing: estimation of the apnea-hypopnea index using wrist-worn reflective photoplethysmography” , used wrist-worn reflective photoplethysmography (rPPG) signals and modified lead II ECG signals to extract cardiovascular and respiratory features. The rPPG pulses' time and morphology characteristics were analyzed, selecting high-quality sinus rhythm pulses to derive inter-beat intervals for HRV analysis. Both IBIs and breathing patterns were utilized to compute HRV and respiratory activity features. These features were regularized using transformations before being fed into the deep learning model. The model classified 30-second epochs from overnight recordings as either influenced by a respiratory event or not, providing probability scores for each epoch. The Apnea Hypopnea Index (AHI) was estimated for each participant and corrected using a multiplicative coefficient derived from linear regression with the reference AHI obtained from manual annotations.

[0016] Hay ano et al. “Quantitative detection of sleep apnea in adults using inertial measurement unit embedded in wristwatch wearable devices ' investigated the feasibility of detecting sleep apnea (SA) using an inertial measurement unit (IMU) embedded in a wristwatch wearable device. Acceleration and gyro signals were analyzed separately but similarly. Respiratory signals were extracted from each axis of IMU using a band-pass filter, then combined into a single scalar reflectingrespiratory wrist movement. Moving averages of the envelope were calculated for extracting fast and slow trends. The respiratory event index (REI) was calculated as the average number of SA episodes per hour of total recording time (TRT), and the ratio of the bounded area to the area under the curve (AUC) of the slow trend was used as an index for SA detection.

[0017] The ApneaTrak is a device that records respiratory waveforms for up to three nights on each charge, providing essential data for diagnosing sleep apnea. It features body position, effort, snoring, airflow, and oximetry measurements, and can be used with EnsoData's Al scoring and sleep time analysis for HSATs to qualify for higher reimbursements. The device is designed to prevent accidental turn-off and data loss, with secure, rechargeable batteries. Easy III software supports both PSG and HSAT solutions, ensuring streamlined sleep diagnostic testing. Setup is easy with color-coded connections, anatomical imagery, and patient instructional materials. Non-proprietary connectors allow compatibility with preferred accessories.

[0018] The WatchPAT is a wearable device used for home sleep apnea testing. It is a watch-like device coupled with a finger probe to measure the peripheral arterial signal (PAT), which reflects changes in blood flow during sleep. This allows the device to detect sleep stages, breathing patterns, and movements throughout the night. By analyzing the PAT signal, WatchPAT can identify different types of sleep apnea events and calculate important diagnostic parameters, such as the Apnea Hypopnea Index (AHI), AHIc, RDI, and GDI based upon True Sleep Time and Sleep Staging. WatchPAT

[0015] is designed to be used for a few nights specifically for SA diagnosis through a proprietary hardware comprising of a wrist- worn specific-purpose display and a pulse oximeter.

[0019] The AcuPebble 0x100 system offers an accurate sleep study report by extracting data from two wireless devices: a neck sensor and a finger sensor. Theeight-channel system comprises oximetry, cardiac signal, activity, respiratory phases, airflow, snoring, pulse rate, and body position information. The neck sensor captures essential sounds related to respiratory and cardiac functions, while the finger oximeter records oxygen saturation levels. These signals are transmitted wirelessly to a mobile device and processed in the cloud using advanced algorithms to diagnose sleep apnea. The comprehensive report includes AHIs, and GDIs, all determined using estimated sleep time. AcuPebble 0x100 is designed to be used for a few nights specifically for SA diagnosis through a proprietary hardware comprising of a ring with specific -purpose display and a device that must be attached to the neck.

[0020] The Belun Ring uses spectrophotometry to measure blood oxygen saturation by analyzing light absorption of oxygenated and deoxygenated blood at red and infrared wavelengths. This data is used to calculate arterial oxygen saturation and generate a sleep report assessing sleep apnea risk, sleep efficiency, and stress levels. It also measures SpO2, Pulse Rate (PR), bAHI, sleep stages, Autonomic Nervous System (ANS) response, and actigraphy. Belun Ring is designed to be used for a few nights specifically for SA diagnosis through a proprietary hardware comprising of a ring-like device that, although less invasive than a PSG, is designed to be worn during the night, due to its size possibly hindering activities in the daily life.

[0021] Sleepimage offers two devices for data collection: The Sleepimage Ring, intended for multi-patient use and single- or multi-night testing, and the Sleepimage Tip, intended for single-patient use with three tests over seven nights designed for treatment tracking. These devices analyze blood oxygenation data to generate desaturation events, display graphs, and calculate Sleepimage Apnea Hypopnea Index (sAHI) by combining Cardiopulmonary Coupling (CPC) biomarkers and hypoxic events during sleep time. The sAHI helps cliniciansinterpret and manage sleep disordered breathing (SDB) events concurrently with CPC sleep states to evaluate disease severity and differentiate between obstructive, central, or complex / mixed sleep apnea using pathology biomarkers.OBJECTIVE OF THE INVENTION

[0022] There are several technical problems related to the estimation of Apnea-Hypopnea Index (AHI) from Peripheral Oxygen Saturation (SpO2) signal.

[0023] For example, robust performance regardless of SpO2 acquisition conditions. There are several consumer grade devices to compute SpO2, from golden standard pulse oximeter devices, to less accurate wrist worn devices. Combining a lightweight preprocessing pipeline, a bank of representative features of SpO2, and a sequential deep learning model that captures temporal relationships between respiratory events, the present invention aims to display remarkable robustness to different SpO2 acquisition conditions, including moderate presence of noise and signal loss.

[0024] Another challenge is on-device apneic event detection using machine learning techniques. At the core of the present invention’s Apnea-Hypopnea Index estimation approach lies the use of machine learning techniques to detect the occurrence of apneic events throughout the sleep session. Furthermore, the model must be lightweight and efficient, such that it can be deployed entirely on the wearable device, without need to communicate with an outside server. This last requirement is especially challenging, but of great importance, since it guarantees that the sensitive health data from the user is well protected from leakage. It is therefore an objective of the present invention to provide a solution with a lightweight, yet effective, deep neural network based on single dimensional convolutions, operating over a multi-variate time series of handcrafted features extracted from the signal recorded during sleep.

[0025] Yet another challenge is the precise estimation of AHI using recording time. Reference AHI is computed using total sleep time - the time spent in REM and Non-REM sleep stages, meaning that the time spent in Wake sleep stage is discounted from the computation. It is therefore an objective of the present invention to provide a solution that works with no reliance on sleep stages information, as such relying on the total recorded time to perform the AHI estimation.

[0026] This can lead to underestimation of the AHI, since the recorded time is always greater than or equal to the total sleep time. It is therefore another objective of the present to provide, through careful tuning of the algorithm and employment of adjusting functions, a way to accurately offset this underestimation, leading to a more precise estimate of the AHI.SUMMARY OF THE INVENTION

[0027] In order to solve the limitations of the state of the art, the present invention proposes a computer implemented screening method for Sleep Apnea Disorder, SA, based on a peripheral oxygen saturation, SpO2, signal, comprising the steps of:• receiving, at a Signal Preprocessing Module (101), a SpO2 signal (100) of a user to perform preprocessing operations on the SpO2 signal (100) and generate a Preprocessed SpO2 (203) comprising windows, wherein preprocessing operations comprise: o interpolating (200) missing or invalid measurements of the SpO2 signal (100) based on the valid range of values; o clipping (201) measurements below the lower end of the valid range to the lower end value and measurements above the upper end of the range to the upper end value;o segmenting (202) the SpO2 signal (100) into Preprocessed SpO2 (203) comprising windows with N samples;• extracting, by means of a Feature Extraction Module (102), representative SpO2 features designed to capture desaturation patterns in the SpO2 signal for each window of the Preprocessed SpO2 (203);• forming, by means of the Feature Extraction Module (102), a Features Sequence (303), wherein the Features Sequence (303) comprises a txF shaped feature vector, with t being the period of the window and F being the number of features stacked in a TxF matrix, the value of T corresponding to the number of time intervals, rounded down, of the SpO2 signal (100);• detecting respiratory events within the Features Sequence (303) by means of a Respiratory Event Detection Module (103) using a modified Temporal Convolutional Network, TCN, (400) backbone;• generating a Respiratory Events Sequence (402) of probabilities with shape Txl, such that position i of the sequence has the probability that the i-th time interval of the analysis corresponds to a respiratory event, wherein the Respiratory Events Sequence (402) are binarized by Thresholding (401), with a threshold value selected by optimizing the balanced accuracy of the respiratory event detection task on a validation set;• aggregating, by means of a AHI Estimation Module (104), the Respiratory Events Sequence (402) to generate an estimation of an Apnea-Hypopnea Index, eAHI, (105), wherein aggregating comprises: o computing a ratio of events-per-hour, by means of an Average Events per Hour (500) module, by dividing the number of respiratory events detected by the total recording time;o applying, by means of an AHI Adjustment Function (501) module, an adjustment function to the ratio of events -per-hour to at least reduce or mitigate distortions in AHI estimation; o applying, by means of a Coverage Correction (502) module, a coverage correction to the eAHI, weighting it by the inverse of SpO2 signal (100) coverage C, wherein C is defined by:o wherein SpO2(i) is the SpO2 value at the i-th second of the SpO2 recording before the signal interpolation step (200), N is the total size, in seconds, of the SpO2 recording and valid defines C as a fraction of a valid range of SpO2 values in the recording.

[0028] The present invention also proposes a system for screening for Sleep Apnea Disorder, SA, based on a peripheral oxygen saturation, SpO2, signal, comprising:• a Signal Preprocessing Module (101) configured to receive a SpO2 signal (100) of a user to perform preprocessing operations on the SpO2 signal (100) and generate a Preprocessed SpO2 (203) comprising windows, wherein preprocessing operations comprise: o interpolating (200) missing or invalid measurements of the SpO2 signal (100) based on the valid range of values;o clipping (201) measurements below the lower end of the valid range to the lower end value and measurements above the upper end of the range to the upper end value; and o segmenting (202) the SpO2 signal (100) into Preprocessed SpO2 (203) comprising windows with N samples;• a Feature Extraction Module (102) configured to: o extract representative SpO2 features designed to capture desaturation patterns in the SpO2 signal for each window of the Preprocessed SpO2 (203); and o form a Features Sequence (303), wherein the Features Sequence (303) comprises a txF shaped feature vector, with t being the period of the window and F being the number of features stacked in a TxF matrix, the value of T corresponding to the number of time intervals, rounded down, of the SpO2 signal (100);• a Respiratory Event Detection Module (103) configured to: o detect respiratory events within the Features Sequence (303) by using a modified Temporal Convolutional Network, TCN, (400) backbone; o generate a Respiratory Events Sequence (402) of probabilities with shape Txl, such that position i of the sequence has the probability that the i-th time interval of the analysis corresponds to a respiratory event, wherein the Respiratory Events Sequence (402) are binarized by Thresholding (401), with a threshold value selected by optimizing the balanced accuracy of the respiratory event detection task on a validation set;• a AHI Estimation Module (104) configured to aggregate the Respiratory Events Sequence (402) to generate an estimation of an Apnea-Hypopnea Index, eAHI, (105), wherein the AHI Estimation Module further comprises:o an Average Events per Hour (500) module configured to compute a ratio of events-per-hour by dividing the number of respiratory events detected by the total recording time; o a AHI Adjustment Function (501) module configured to apply an adjustment function to the ratio of events-per-hour to at least reduce or mitigate distortions in AHI estimation; o a Coverage Correction (502) module configured to apply a coverage correction to the eAHI, weighting it by the inverse of SpO2 signal (100) coverage C, wherein C is defined by:o wherein SpO2(i) is the SpO2 value at the i-th second of the SpO2 recording before the signal interpolation step (200), N is the total size, in seconds, of the SpO2 recording and valid defines C as a fraction of a valid range of SpO2 values in the recording.BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The invention is explained in greater detail below and makes references to the drawings and figures, attached herewith, when necessary. Attached herewith are:

[0030] Figure 1 presents an overview of a system according to a preferred embodiment of the present invention.

[0031] Figure 2 presents a Preprocessing Module according to a preferred embodiment of the present invention.

[0032] Figure 3 presents a Feature Extraction Module according to a preferred embodiment of the present invention.

[0033] Figure 4 presents a Respiratory Event Detection Module according to a preferred embodiment of the present invention.

[0034] Figure 5 presents a AHI Estimation Module according to a preferred embodiment of the present invention.DETAILED DESCRIPTION OF THE INVENTIONPreferred Embodiment

[0035] An overview of a preferred exemplary embodiment of a method according to the present invention is illustrated in the block diagram of Figure 1. The preferred method may be described as comprising four main modules which operate on a sequential fashion:• Signal Preprocessing Module (101)• Feature Extraction Module (102)• Respiratory Event Detection Module (103)• AHI estimation module (104)

[0036] The first module takes as input a sleep session recording of a SpO2 signal (100) of a user at a determined sampling rate (for example, 1 Hz). The invention is described by means of an illustrative embodiment wherein the SpO2 signal is at a sampling rate of 1 Hz. However, as it will be clear for a person skilled in the art, the sampling rate depends on the device that is used for recording the SpO2 signal (100). Typically, the devices available in the market use a sampling rate of 1 Hz. Nevertheless, Leppanen, T. et al, in Pulse Oximetry: The Working Principle, Signal Formation, and Applications a PPG, cites that the PPG used for computing the SpO2 is generally in a frequency between 10 to 25 Hz, which results in a SpO2 in the respective frequency. Thus, it would be feasible to use different sampling rates, since there would be no significative changes in the performance. In that case,the differences in the sampling frequency will change the number of samples that will be obtained during recording. For example, in case a frequency of 10 Hz is used, the number of samples obtained in a specific period would be higher when compared to the number of samplings obtained with a frequency of 1 Hz. Thus, more computational processing and time would be required for computing the features for higher frequencies. The SpO2 signal (100) is processed by each subsequent block in order to produce an estimation of the Apnea-Hypopnea Index (AHI), which will also be referred herein as eAHI (105).Module 1 - Signal Preprocessing

[0037] The purpose of a Preprocessing Module (101), with a preferred embodiment illustrated in Figure 2, is to clean up the SpO2 signal (100) and prepare it for the Feature Extraction Module (102) as taught by the present invention. The preferred method according to the present invention works on the assumption that SpO2 signal (100) values are valid if within a determined range, such as 70% to 100% range. Values under the threshold (for example, under 70%), while they may occur in practice, cannot be sustained for long, as it poses severe risk to a person’ s health. So, considering the purposes and application of the present solution, values under the threshold may be considered as artefacts and be interpolated. Those skilled in the art will be able to comprehend that such a range may be chosen arbitrarily so that any ranges may be used without compromising the present solution.

[0038] The preferred embodiment of the Preprocessing Module (101) according to the present invention may apply one or more operations to the SpO2 signal (100), some of are shown in Figure 2 and also discussed below.

[0039] As shown in Figure 2, the Processing Module (101) may comprise an Interpolation (200) process. Missing or invalid measurements are interpolated to be within the range of values (such as 70% to 100% range) considered as valid. This Interpolation (200) process operates using a moving window scheme: the signal ispreferably split in windows of a certain period (such as seconds, minutes or hours), each being interpolated individually. For each window, two scenarios may be considered:• If a portion of the samples (for example, at least 20%) within the window are valid, the Processing Module (101) performs a 1st order spline interpolation using the valid values within the window to interpolate the invalid values.• Otherwise, the invalid values are filled with the median of valid SpO2 signal (100) over the entire recording.

[0040] Using the median guarantees that, when the window does not have enough values for accurate interpolation, a baseline value is used.

[0041] Another process that may be part of the Processing Module (101) is a Clipping (201) process. Interpolation might result in some values slightly outside the valid range, so values under the lower end of the threshold are clipped to the lower end, while values over the upper end of the threshold are clipped to the upper end. For example, if the valid range is 70%-100%, any values under 70% are clipped to 70% while any values above 100% are clipped to 100%.

[0042] One more example of a process that may be part of the Processing Module (101) is a Segmentation (202) process. The proposed invention detects respiratory events at a certain rate (from a few seconds to several minutes). This preprocessing step comprises segmenting the SpO2 signal (100) in windows for feature extraction and, subsequently, respiratory event detection. According to the exemplificative embodiment (shown in Figure 2), the SpO2 signal (100) is segmented in non-overlapping 1 -minute windows. As it will be clear for a skilled person, different sizes of windows may be used. Similarly, the SpO2 signal (100) could be segmented in overlapping and non-overlapping windows.

[0043] Due to the sampling rate, each window will contain a number of samples N equal to the period of the window times the sampling rate. If the last windows do not contain N samples, it is removed from the analysis. Windows resulting from this step are referred as Preprocessed SpO2 (203), which are forwarded to the Feature Extraction Module (102).Module 2 - Feature Extraction

[0044] The next step in the pipeline is the Feature Extraction Module (102). The goal of this module is to extract representative SpO2 features for each window of the Preprocessed SpO2 (203). Features are designed to directly or indirectly capture desaturation patterns in the SpO2 signal, strongly correlated with the occurrence of respiratory events. In the literature on automatic sleep-related breathing disorder classification, several SpO2 features have been proposed with this purpose. As an example, we present a set of SpO2 features proposed by Xie and Minn in the paper "Real-Time Sleep Apnea Detection by Classifier Combination ' , roughly divided in three main groups.

[0045] The first group comprises the Statistical features. They are common statistics of the Preprocessed SpO2 (203), computed on the Preprocessed SpO2 (203) windows. Total of 6 features. Examples: Minimum, Maximum, Mean, Variance, Skewness, Kurtosis of SpO2 values or a combination thereof.

[0046] The second group comprises the Oxygen desaturation features. Inspired by the methodology used to compute indexes of oxygen desaturation, they measure the decrease of oxygen saturation relative to some baseline. Examples:• ODIX: oxygen desaturation index feature, number of times the SpO2 falls X units below a baseline value. Baseline SpO2 value is considered as the last SpO2 value of the last resaturation that happened prior to the desaturation event. Multiple values of X can be used for different parameterizations of this feature.• ODISX: number of times the SpO2 falls X units below the mean of the top 20% values of SpO2 in the 1 minute window. Multiple values of X can be used for different parameterizations of this feature.• ODIXY : number of times the SpO2 falls X units below some baseline value for at least Y seconds. Multiple values of X and Y can be used for different parameterizations of this feature.• TSAX: Cumulative time SpO2 stays below X. Multiple values of X can be used for different parameterizations of this feature.

[0047] The third group comprises Nonlinear / Misc. features. Generic and non-linear operations on time- series taken from several works in the literature on automated sleep apnea disorder detection

[0019] . Examples:• Delta Index: a SpO2 window is split in 12-second sub-windows. For each, the mean SpO2 value is computed. Then, the first-order difference between sub-windows is computed. The Delta Index is the average of the first-order differences.• Approximate Entropy: measures regularity and unpredictability in timeseries through an approximation of its entropy.• Zero Crossing Points: number of times the SpO2 decreases under the window’s mean.

[0048] Subsets of such features set have been used extensively in many works in the literature related to automatic detection of respiratory events from SpO2 signals, with strong performance in a variety of scenarios. The present solution is compatible but not limited to this set of features, being agnostic to the features used as long as they can be computed for the windows of SpO2. Feature extraction functions or different parameterizations of these features can be added or removed from the set used depending on the needs of the target application of our solution. Atradeoff between performance and efficiency can be taken into account when selecting which features to be use.

[0049] Each window of Preprocessed SpO2 (203) results in a txF shaped feature vector, with t being the period of the window and F being the number of features. Feature vectors for each Preprocessed SpO2 (203) window are stacked in a TxF matrix, the value of T corresponding to the number of time intervals (such as one or more minutes), rounded down, of the recorded SpO2 signal (100). For instance, if the recording has a total of 5 hours and 42 minutes and the time interval is 1 minute, then the value of T is 342. This forms a Features Sequence (303) which is analogous to a multi-variate time-series, with a T-sized timestep dimension, and F-sized channel dimension.Module 3 - Respiratory Event Detection

[0050] The Respiratory Event Detection Module (103) is tasked with detecting respiratory events within the windows of the Features Sequence (303). The preferred embodiment of the present invention uses a neural network model with a sequence-to-sequence (seq2seq) architecture (400) to predict a probability that each window has or not a sleep-related breathing disorder event. The neural network’ s input is a Features Sequence (303) with shape TxF, such that T is the number of timesteps, for example, the number of minutes in the recording, and F is the number of features computed for each window. At position i in the Features Sequence (303) is the F-sized feature vector computed for the i-th interval of the recording. One of the major differences from the present invention to other works in the literature is how features are processed. Instead of performing intra-window classification, the network processes the input in an inter-window fashion, by treating the Features Sequence (303) as a multi-channel signal, and detecting patterns in the variations within such signal. The present solution models the respiratory event detection task as a supervised seq2seq binary classification approach, in which the target isclassifying if each window of the recording contains or not a respiratory event. Therefore, the present invention uses a labeled training set of SpO2 recordings, with true / false for the presence of respiratory events for each time interval under analysis.

[0051] The used neural network preferably employs a modified Temporal Convolutional Network (TCN) (400) backbone. This is a seq2seq architecture with residual blocks of 1 -dimensional convolutions. Standard TCN backbones employ causal convolutions to process the input signal in order to capture temporal relationships between each timestep. One of the major modifications employed by the present invention is using standard convolutions instead of causal, to capture general neighborhood information existing in the multi-channel input time-series. Causal convolutions are often a requirement of real-time applications, as it only takes into account past information for the model to learn. The present invention processes the signal offline, so the causality constraint ends up restricting the model’s capability of learning from past and future information, therefore we remove the causality constraint. In this architecture, dilated convolutions in each residual block allow the network to process bigger portions of the signal. Since respiratory events vary significantly in duration, and can be clustered within regions of the recording, it is beneficial to optimize the dilation size to process both smaller portions and bigger portions of the input signal, in order to detect both regions of great respiratory event density as well as isolated events. Dilation rate and other hyperparameters of the architecture have been tuned using hyperparameter optimization techniques like, but not limited to, a Bayesian hyperband optimization approach, with the event detection area under the precision-recall curve on a validation set as target measure for optimization.

[0052] The Temporal Convolutional Network (400) outputs a sequence of probabilities with shape Txl, such that position i of this sequence has the probability that the i-th time interval of the analysis corresponds to a respiratory event. Theprobability output may be obtained by using any type of classification head, such as a dense layer with sigmoid activation, as the output layer of the neural network. Output probabilities are binarized by Thresholding (401), with a threshold value selected by optimizing the balanced accuracy of the respiratory event detection task on the validation set.Module 4 - AHI Estimation

[0053] The goal of AHI Estimation Module (104) is to aggregate the Respiratory Events Sequence (402) to generate eAHI (105), an estimation of the AHI. First, the module Average Events per Hour (500) measures events-per-hour by dividing the number of respiratory events detected by the total recording time, a rough estimation of the AHI. This ratio is then used as input to a AHI Adjustment Function (501), with the goal of at least reducing, mitigating or eliminating distortions in AHI estimation. There are two factors that contribute to distorting the estimation of AHI from counting the respiratory events detected per-minute: (1) using total recording time instead of total sleep time, and (2) using a larger time interval resolution for the detection. The goal of the AHI Adjustment Function (501) is to compensate for such distortions, resulting in a more accurate estimation of the AHI. An example of adjustment function is the hill function proposed by Jung in the paper “Real-Time Automatic Apneic Event Detection Using Nocturnal Pulse Oximetry”'.1254.O6X1 01Adj x) = 0.07 +936 64i.oi+ %i.oi (1)

[0054] Any adjustment function may be used in our solution, including Equation (1) above. Other functions may be used according to the needs of the target application that implements the present solution, including a passthrough identity function to perform no AHI adjustment. Finally, we apply a Coverage Correction(502) to the eAHI, weighting it by the inverse of SpO2 signal (100) coverage C, under the assumption that respiratory events are uniformly distributed along the recording, and thus AHI grows linearly according to the fraction of valid signal. The equations below define the value of C for a valid range of values from a% to b%:

[0055] Where SpO2(i) is the SpO2 value at the i-th second of the SpO2 recording before the signal interpolation step (200), and N is the total size, in seconds, of the SpO2 recording. Since we establish a range a%-b% (such as 70%, 100%) as the range for considering valid SpO2 values, Equation (2) defines C as the fraction of valid SpO2 values in the recording.

[0056] Therefore, after the steps above, the result is one eAHI value for each input.Alternatives (2nd Embodiment) (optional)SA Disorder Estimation

[0057] The invention can be used as the basis for a Sleep-related Breathing Disorder estimation method. Clinically, the AHI is used by physicians to determine if the patient suffers or not from SA, by establishing a threshold on the AHI value. Two accepted rules

[0021] are:• AHI>5 and the patient displays symptoms consistent with SA• AHI>15

[0058] A further embodiment of the present solution may implement either approach. For the first rule, the system implementing this embodiment measures theuser’s eAHI (105) by using, for example, the preferred embodiment of the present invention, and then presents a questionnaire with common symptoms of SA. If eAHI (105) is greater than 5 and the user scores, computed from their questionnaire answers, is over a pre-defined threshold, the system will notify the user that they might suffer from SA. For the second rule, it suffices that the method computes an eAHI (105) greater than or equal to 15 to notify the user and refer them to a medical professional for further investigation. Furthermore, both rules can be combined in an OR fashion:• If the eAHI>15 OR eAHI>5 and the user has symptoms of SA, display the user they show signs of sleep-related breathing disorder• Else, display the user it has no signs of sleep-related breathing disorder.

[0059] An extension of this embodiment is to produce a negative / positive SA response over a certain number of sleep sessions, and notify the user only if a minimum number of sleep sessions within the tested interval have produced a positive result.SA Severity Estimation

[0060] Yet another alternative embodiment of the present invention is as a SA severity estimator and tracking method. There are four clinically accepted

[0021] thresholds on the AHI value to determine a patient’s SA severity:• AHI<5: No SA• 5<AHI<15: Mild SA• 15<AHI<30: Moderate SA• 30< AHI: Severe SA

[0061] Targeting population already diagnosed with SA, the SA severity estimation method proposed may implement the preferred embodiment of this invention to compute the eAHI (105), and use the thresholds above to determine the SA severity for the sleep session. Moreover, it can track the changes in SA severityover several sleep sessions, allowing users to monitor the evolution of the condition. By correlating SA severity with user behavior, it can help users manage their disorder, track its evolution, and give insights to cause positive behavior change that may help lessen the impact of SA on their daily life.Near real-time detection of Respiratory Events

[0062] It is possible to adapt the event detection layer according to the present invention to further produce near real-time detection of respiratory events. The TCN model employed for respiratory event detection is flexible and can use different sequence lengths as its input. By choosing a buffer size B, at some minute i during sleep, the sequence of features from all B-l minutes prior to minute i are buffered, and concatenated to the features from minute i, which are then used as input to the sequential model. The last prediction of the sequential model, corresponding to minute i, is used. For this embodiment, the eAHI estimation module is removed.

[0063] Near real-time detection of respiratory events can be used to give the user insights on how their respiratory events are distributed during sleep, allowing them to adjust behavior that can help manage the occurrence of respiratory events. Furthermore, this tool can be used in conjunction with other devices which can trigger functionality to help alleviate the impact of the respiratory event in course. Hardware implementations

[0001] It is worth mentioning that the example embodiments described herein may be implemented using hardware, software or any combination thereof and may be implemented in one or more computer systems or other processing systems. Additionally, one or more of the steps described in the methods of the example embodiments herein may be implemented, at least in part, by machines. Examples of machines that may be useful for performing the operations of theexample embodiments herein include general purpose portable computers, mobile communication devices, tablets, wearables and / or similar devices.

[0002] For instance, one illustrative example system for performing the operations of the embodiment herein may include one or more components, such as one or more processors, for performing the arithmetic and / or logical operations required for program execution, and storage media, such as one or more disk drives or memory cards (e.g., flash memory) for program and data storage, and a random access memory, for temporary data and program instruction storage.

[0003] Moreover, the aforementioned system of the present invention may also include software residing on a storage media (e.g., a disk drive or memory card), which, when executed, directs the microprocessor(s) in performing transmission, reception and / or processing functions. The software may run on an operating system stored on the storage media and can adhere to various protocols such as the Ethernet, ATM, TCP / IP protocols and / or other connection or connectionless protocols.

[0004] As is well known in the art, microprocessors can run different operating systems, and can contain different types of software, each type being devoted to a different function, such as handling and managing data / information from a particular source, or transforming data / information from one format into another format. The embodiments described herein are not to be construed as being limited for use with any particular type of computer, and that any other suitable type of device for facilitating the exchange and storage of information may be employed instead.

[0005] Software embodiments of the example embodiments presented herein may be provided as a computer program product, or software, that may include an article of manufacture on a machine-accessible or computer-readable medium (also referred to as “machine-readable medium”) having instructions. The instructions on the machine-accessible or machine-readable medium may be used toprogram a computer system or other electronic device. The machine-readable medium may include, but is not limited to, floppy diskettes, optical disks, CD ROMs, magneto-optical disks, electromagnetic signals or any other type of media / machine- readable medium suitable for storing and / or transmitting electronic instructions.

[0006] Thus, the present invention also relates to a computer readable medium which comprises instructions that, when executed by one or more processors, cause the one or more processors to perform an embodiment of the method as disclosed in the present invention.ADVANTAGES OF THE PRESENT INVENTION

[0007] The solution described herein provides a lightweight method to estimate an Apnea-Hypopnea Index (AHI) using only peripheral oxygen saturation (SpO2). It can be deployed entirely on device, making it a perfect fit for scenarios in which privacy over user data is a concern, also eliminating the cost of centralized computation. Furthermore, the estimation has been shown to be accurate, with good correlation to reference values of AHI, the only requirement being that the SpO2 recording is performed during sleep. By design, the proposed solution works well under a wide range of sensor varieties, from the golden-standard transmissive sensors of pulse oximeter devices, to the reflective sensors present in other consumer wearable devices, such as smartwatches. Some of the advantages of the present invention are:• Requires only SpO2 signal, obtained non-invasively from Photoplethysmography (PPG).• Is lightweight and runs entirely on device, mitigating risks of data leakage (i.e., keeping privacy of user data) and requiring no expenditure on servers for inference.• Has tailored preprocessing steps to mitigate effects of PPG sensor errors.• Combines handcrafted features with sequential deep learning model to allow learning from patterns in the variations of features, improving event detection performance under different SpO2 acquisition scenarios.• Produces accurate estimation of AHI in the absence of sleep staging information.• Applies an adjusting function to alleviate distortions caused by using recording time and 1 -minute respiratory event window in the estimation of the AHI.• Applies a compensation function to the AHI to alleviate distortions caused by loss of SpO2 signal throughout the recording.• Has a modularized design which allows to easily modify its building blocks with different bank of handcrafted features and a different event detection module.

[0008] The many advantages of the present invention over the state of the art, some of which are mentioned above, should be evident to those skilled in the art. Nevertheless, reference is made to some of the documents found in the state of the art in order to provide further insights.

[0009] For instance, the present invention is robust to reflective and transmissive PPG signals since it employs a deep learning model to estimate AHI. On the other hand, application WO2022221487A1 disadvantageously relies on handcrafted rules based on SpO2 features, requiring more accurate PPG signals (transmissive).

[0010] While US2022087609A1 specifically focus on the description of an in-ear audio device with PPG sensors and barely mentions sleep apnea estimation, the present invention is designed to work with SpO2 derived from different sources of PPG and is focused in a robust sleep apnea estimation method.

[0011] Concerning US2023346302A1, they estimate only the sleep apnea severity. The present invention, however, estimates the AHI - a more fine-grained measurement of sleep apnea -, which can be further used to estimate the sleep apnea severity. Additionally, the machine learning model described in the present invention uses features extracted from the SpO2 allowing for a more lightweight solution than using the entire SpO2 signal as input to the model as proposed by US2023346302A1.

[0012] The invention JP2024072968A is focuses on a system for data acquisition with power consumption optimizations, while the present invention focuses on a method and system for automatically detecting sleep apnea given a source of SpO2 signal.

[0013] Contrary to US20210345949A1, the present invention is able to estimate the AHI using only the SpO2 signal extracted from a PPG sensor.

[0014] Contrary to US20120296182A1, the present invention extracts features from sliding windows instead of the entire signal, and feeds the temporal sequence of features to a TCN that models temporal correlation between the sequence elements.

[0015] Contrary to US5891023A, the present invention uses machine learning techniques to estimate the AHI, being more robust to quality variations of the input SpO2 signal caused by different means of signal acquisition.

[0016] While not clearly specified in patent application JP2016007243 A, air flow signal often requires nose- or mouth-worn sensors to be detected, making it a more invasive solution than the present invention, which relies only on SpO2, derivable from less invasive sensors. Furthermore, JP2016007243A relies on thresholds over the difference of consecutive normalized signal samples to estimate SA, while the present invention uses machine learning techniques to estimate theAHI, being more robust to quality variations of the input SpO2 caused by different means of signal acquisition.

[0017] US2023122156A1 does not estimate an AHI value, contrarily to the present invention. In addition, the present invention has multiple modules in order to guarantee reliable SpO2 sensor data to use as input for the machine learning techniques, like the signal processing module, the feature extraction module and the respiratory event detection module. US2023122156A1 does not have such preprocessing and data preparation modules, which may not work properly when poor SpO2 signal is received as input, contrarily to our invention.

[0018] Zhou et al., Chen et al., Papini et al., Hayano et al. depend on a combination of devices and sensors in order to be able to estimate sleep apnea disorder risk. The present invention, however, depends only on SpO2 sensor data.

[0019] Regarding the products ApneaTrak, WatchPAT, AcuPebble 0x100, Belun Ring and Sleepimage, they are designed to be used for a few nights specifically for SA diagnosis through a proprietary hardware, while the present invention is deployable to any consumer-grade general-purpose devices, including wearable devices, allowing for continuous monitoring and not hindering the user’s normal activities. As a result, the present invention is more flexible since it does not require any specially designed hardware.

Claims

CLAIMS1. A computer implemented screening method for Sleep Apnea Disorder, SA, based on a peripheral oxygen saturation, SpO2, signal, characterized by comprising the steps of: receiving, at a Signal Preprocessing Module (101), a SpO2 signal (100) of a user to perform preprocessing operations on the SpO2 signal (100) and generate a Preprocessed SpO2 (203) comprising windows, wherein preprocessing operations comprise: interpolating (200) missing or invalid measurements of the SpO2 signal (100) based on the valid range of values; clipping (201) measurements below the lower end of the valid range to the lower end value and measurements above the upper end of the range to the upper end value; and segmenting (202) the SpO2 signal (100) into Preprocessed SpO2 (203) comprising windows with N samples; extracting, by means of a Feature Extraction Module (102), representative SpO2 features designed to capture desaturation patterns in the SpO2 signal for each window of a Preprocessed SpO2 (203); forming, by means of the Feature Extraction Module (102), a Features Sequence (303), wherein the Features Sequence (303) comprises a txF shaped feature vector, with t being the period of the window and F being the number of features stacked in a TxF matrix, the value of T corresponding to the number of time intervals, rounded down, of the SpO2 signal (100); detecting respiratory events within the Features Sequence (303) by means of a Respiratory Event Detection Module (103) using a modified Temporal Convolutional Network, TCN, (400) backbone;generating a Respiratory Events Sequence (402) of probabilities with shape Txl, such that position i of the sequence has the probability that the i-th time interval of the analysis corresponds to a respiratory event, wherein the Respiratory Events Sequence (402) are binarized by Thresholding (401), with a threshold value selected by optimizing the balanced accuracy of the respiratory event detection task on a validation set; aggregating, by means of a AHI Estimation Module (104), the Respiratory Events Sequence (402) to generate an estimation of an Apnea-Hypopnea Index, eAHI, (105), wherein aggregating comprises: computing a ratio of events-per-hour, by means of an Average Events per Hour (500) module, by dividing the number of respiratory events detected by the total recording time; applying, by means of a AHI Adjustment Function (501) module, an adjustment function to the ratio of events-per-hour to at least reduce or mitigate distortions in AHI estimation; applying, by means of a Coverage Correction (502) module, a coverage correction to the eAHI, weighting it by the inverse of SpO2 signal (100) coverage C, wherein C is defined by:wherein SpO2(i) is the SpO2 value at the i-th second of theSpO2 recording before the signal interpolation step (200), N is the total size,in seconds, of the SpO2 recording and valid defines C as a fraction of a valid range of SpO2 values in the recording.

2. The computer implemented method according to claim 1 characterized in that interpolating (200) comprises using a moving window scheme wherein the SpO2 signal (100) is split in windows of a determined period, each window being interpolated individually.

3. The computer implemented method according to claim 2 characterized in that for each window: performing a first order spline interpolation using the valid values within the window to interpolate the invalid values provided a portion of the samples within the window are valid; and otherwise, filling the invalid values with the median of valid SpO2 signal (100) over the entire recording.

4. The computer implemented method according to any one of claims 1-3 characterized in that the valid range of values is 60% to 100%, preferably 70% to 100%.

5. The computer implemented method according to any one of claims 1-4 characterized in that each window contains a N samples, wherein N equals the period of the window times the sampling rate; and removing the last window from the analysis provided it does not contain N samples.

6. The computer implemented method according to any one of claims 1-5 characterized in that the extracted features are divided in groups, wherein: a first group comprises Statistical Features such as Minimum, Maximum, Mean, Variance, Skewness, Kurtosis of SpO2 values or a combination thereof;a second group comprises Oxygen desaturation features such as ODIX, ODISX, ODIXY and TSAX; and a third group comprises Nonlinear / Misc. features such as Delta Index, Approximate Entropy and Zero Crossing Points.

7. The computer implemented method according to any one of claims 1-6 characterized by further comprising estimating an Apnea-Hypopnea Index, AHI by: collecting questionnaire data from a user; checking whether the eAHI value is equal to or greater than a predetermined value, such as 5 or 15; checking the answers of the user to the questionnaire to determine whether the user has symptoms of SA; and provided both are true, displaying to the user that they show signs of a sleep-related breathing disorder; otherwise, display the user it has no signs of a sleep-related breathing disorder.

8. The computer implemented method according to any one of claims 1-7 characterized by further comprising: estimating a severity of a SA of a user based on the eAHI value, wherein:• eAHI<5: No SA;• 5<eAHI<15: Mild SA;• 15<eAHI<30: Moderate SA;• 30<eAHI: Severe SA; the method further comprising: tracking the SA severity over a period to monitor the evolution of the user.

9. The computer implemented method according to any one of claims 1-8 characterized in that the event detection layer is further adapted to produce near real-time detection of respiratory events by: buffering the sequence of features prior to a given instant i; concatenating the buffered features with the features from instant I; inputting the concatenated features in the sequential model; using the last prediction of the sequential model, corresponding to instant I; by-passing the eAHI estimation module.

10. Non-transitory computer readable means characterized by comprising a set of instructions that when executed by a processor, cause the processor to perform the method as defined in any one of claims 1-9.

11. A system for screening for Sleep Apnea Disorder, SA, based on a peripheral oxygen saturation, SpO2, signal, characterized by comprising: a Signal Preprocessing Module (101) configured to receive a SpO2 signal (100) of a user to perform preprocessing operations on the SpO2 signal (100) and generate a Preprocessed SpO2 (203) comprising windows, wherein preprocessing operations comprise: interpolating (200) missing or invalid measurements of the SpO2 signal (100) based on the valid range of values; clipping (201) measurements below the lower end of the valid range to the lower end value and measurements above the upper end of the range to the upper end value; and segmenting (202) the SpO2 signal (100) into Preprocessed SpO2 (203) comprising windows with N samples; a Feature Extraction Module (102) configured to:extract representative SpO2 features designed to capture desaturation patterns in the SpO2 signal for each window of the Preprocessed SpO2 (203); and form a Features Sequence (303), wherein the Features Sequence (303) comprises a txF shaped feature vector, with t being the period of the window and F being the number of features stacked in a TxF matrix, the value of T corresponding to the number of time intervals, rounded down, of the SpO2 signal (100); a Respiratory Event Detection Module (103) configured to: detect respiratory events within the Features Sequence (303) by using a modified Temporal Convolutional Network, TCN, (400) backbone; generate a Respiratory Events Sequence (402) of probabilities with shape Txl, such that position i of the sequence has the probability that the i-th time interval of the analysis corresponds to a respiratory event, wherein the Respiratory Events Sequence (402) are binarized by Thresholding (401), with a threshold value selected by optimizing the balanced accuracy of the respiratory event detection task on a validation set; a AHI Estimation Module (104) configured to aggregate the Respiratory Events Sequence (402) to generate an estimation of an Apnea-Hypopnea Index, eAHI, (105), wherein the AHI Estimation Module further comprises: an Average Events per Hour (500) module configured to compute a ratio of events-per-hour by dividing the number of respiratory events detected by the total recording time; a AHI Adjustment Function (501) module configured to apply an adjustment function to the ratio of events-per-hour to at least reduce or mitigate distortions in AHI estimation;a Coverage Correction (502) module configured to apply a coverage correction to the eAHI, weighting it by the inverse of SpO2 signal (100) coverage C, wherein C is defined by:wherein SpO2(i) is the SpO2 value at the i-th second of the SpO2 recording before the signal interpolation step (200), N is the total size, in seconds, of the SpO2 recording and valid defines C as a fraction of a valid range of SpO2 values in the recording.

12. The system according to claim 11 characterized in that interpolating (200) comprises using a moving window scheme wherein the SpO2 signal (100) is split in windows of a determined period, each window being interpolated individually.

13. The system according to claim 12 characterized in that for each window: performing a first order spline interpolation using the valid values within the window to interpolate the invalid values provided a portion of the samples within the window are valid; and otherwise, filling the invalid values with the median of valid SpO2 signal (100) over the entire recording.

14. The system according to any one of claims 11-13 characterized in that the valid range of values is 60% to 100%, preferably 70% to 100%.

15. The system according to any one of claims 11-14 characterized in that each window contains a N samples, wherein N equals the period of the window times the sampling rate.; and removing the last window from the analysis provided it does not contain N samples.

16. The system according to any one of claims 11-15 characterized in that the extracted features are divided in groups, wherein: a first group comprises Statistical Features such as Minimum, Maximum, Mean, Variance, Skewness, Kurtosis of SpO2 values or a combination thereof; a second group comprises Oxygen desaturation features such as ODIX, ODISX, ODIXY and TSAX; and a third group comprises Nonlinear / Misc. features such as Delta Index, Approximate Entropy and Zero Crossing Points.

Images