System and method for automatic detection of respiratory events

Abstract

A method for training a system for automatic detection of respiratory events is provided. The method comprises capturing data from users using sensors on a wearable garment, annotating the collected data while the users are performing coughs, processing the annotated data and extracting selection features from the processed data having most importance on prediction of detection of respiratory events. A smart garment automatically detecting respiratory events comprising a respiratory sensor, an electrocardiogram sensor, an accelerometer, a processor and a mobile power source. The processor being configured to receive data from the respiratory sensor, an electrocardiogram sensor and an accelerometer, use a machine learning trained model to detect a cough event and a non-cough event based on the received data, the trained model comprising a plurality of selection features having most importance on prediction of detection of respiratory events.

Description

SYSTEM AND METHOD FOR AUTOMATIC DETECTION OF RESPIRATORY EVENTSCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] The present patent application claims the benefits of priority of United States Provisional Patent Application No. 63 / 619,135, entitled “SYSTEM AND METHOD FOR AUTOMATIC DETECTION OF RESPIRATORY EVENTS” and filed at the United States Patent and Trademark Office on January 9, 2024, the content of which is incorporated herein by reference.FIELD OF THE INVENTION

[0002] The present invention generally relates to systems and methods for respiratory activity classification comprising the use of body measurements collected with wearable physiological sensors. More precisely, the present invention relates to systems and methods for respiratory activity classification comprising adhesives, harnesses, and smart garments such as, but not limited to, Respiratory Inductance Plethysmography (RIP), accelerometry, sound, vibration, and electrocardiogram (ECG) via supervised machine learning classification models.BACKGROUND OF THE INVENTION

[0003] Respiratory activity classification is useful to assess general health and wellbeing. Respiratory activity reflects pathology severity, mental states, internal organs health, infectious diseases symptoms, etc. and are also useful to diagnose health conditions such as cystic fibrosis, tuberculosis, COPD, asthma, chronic cough, influenza, COVID- 19, colds, etc.

[0004] Methods for respiratory activity classification are varied. Main approaches are: manual annotation of events reviewed by trained medical personnel, algorithmic solutions, statistical models, or machine-learning models.

[0005] Sensor modalities for respiratory activity classification are varied. Most approaches documented are based on the detection of sound or mechanical vibrations (Acoustic monitoring system, monitoring method, and monitoring computer program, Patent number: US10898160B2, Filing date: Nov 27, 2015 Issue date: Jan 26 2021; Wearable system for autonomous detection of asthma symptoms and inhaler use, and for asthma management, Patent number: US11179060B2, Filing date: Jul 7, 2016. Issue date: Nov 23, 202; Compressed cough automatic detection method and embedded device. Patent number: CN112687290B Filing date: Dec 30, 2020. Issue date: Sep 20, 2022).

[0006] Systems using accelerometer data are also common (Device and method for detecting and monitoring cough. Patent number: CN112687290B. Filing date: Dec 30, 2020. Issue date: Sep 20, 2022; Cough detection using frontal accelerometer. Patent number: US11793423B2, Filing date: May 3, 2021. Issue date: Oct 24, 2023).

[0007] Other signals such as ECG sensors (Wearable medical monitoring device. Patent number: US9737262B2, Filing date: Jan 29, 2016, Issue date: Aug 22, 2017) and respiratory sensors (Band-like garment for physiological monitoring. Patent number: US9504410B2. Issue date: Oct 24, 2006, Filing date: Nov 29, 2016) have also been harnessed for respiratory activity classification.

[0008] Cough detection is a very important respiratory activity (event) for health and wellbeing assessment. Coughing is a forceful, quick expulsion of air from the lungs, often associated with respiratory tract infections and diseases.

[0009] Cough can also be a symptom of allergies, including those that involve the nose or sinuses, asthma, cystic fibrosis, pulmonary fibrosis and COPD, including emphysema or chronic bronchitis, the common cold, flu, COVID-19, and other viral infections including lung infections such as pneumonia or acute bronchitis, sinusitis, gastroesophageal reflux disease (GERD), exposure to pollutants, cigarette smoking or exposure to secondhand smoke, lung cancer and other lung diseases.

[0010] Methods to analyze sensor data for respiratory activity classification are generally either algorithm, statistical, or model-based. For example, the Lifeshirt (no longer available) used a proprietary algorithm housed within the software to identify cough from the physiologic recordings. PulmoTrack-CC automatically finds potential coughs via an algorithm based on energy characteristics of the signal. Expert reviewers then confirm the class assignment. The Leicester Cough Monitor uses a statistical model (Hidden Markov models) for cough detection. Model-based approaches include The Hull Automated Cough Counter (artificial neural networks).

[0011] Validation of respiratory activity classification systems, common endpoints. Systems are validated against manually identified coughs or other respiratory events. Some systems only validate the total number of events (bouts) while others validate each event timing / location.

[0012] Audio and Privacy Traditionally, microphones and videos are the main source of reliable cough information and have been the focus of most of the research and development efforts for automatic cough detection. However, these methods are rarely used in clinical practice, in hospitals, at home, at work, etc. because of obvious privacy issues related tocontinuously recording personal audio and or video information. Some groups have proposed on-device pre-processing of audio using filters, ensuring that recorded sounds are intelligible. Methods to automatically detect cough and estimate cough frequency without invading a user’s privacy the way audio recordings do are desirable and useful.

[0013] All self-assessment questionnaires are subjective, sensitive to user effort to report, and not practical during sleep or physical activity. Automatic cough detection is desirable.

[0014] Increasingly, automatic detection of coughs via machine learning algorithms is gaining popularity. Pahar et al. [8] developed machine learning algorithms to predict coughs based on acceleration sensor data placed in the headboard of a hospital bed. In a population of 14 patients (6,000 coughs and 68,000 non-coughs), they obtained an Area Under the Curve (AUC) value of 98% using a residual -based architecture model known as Resnet50 [9], Although model performance is strong, the approach delimits use to a pre-instrumented bed. Mohammadi et al.

[0010] developed a system with an accelerometer placed in the neck region with strong performance (sensitivity and specificity of 91 and 89%, respectively) for classifying coughs vs other events (swallowing, speaking, head movements). Performance is unknown for a larger class of respiratory events and their approach may be impractical for long term monitoring given the placement of the sensor. More recently, a smart watch-based device demonstrated strong cough detection performance in an uncontrolled, free-living environment (AUC 85.5%); however, it relies on both movement and audio data, the latter being potentially problematic as previously discussed. A wearable, unobtrusive device is required to achieve remote monitoring of cough symptoms in ambulatory settings.

[0015] Coughing is a forceful, quick expulsion of air from the lungs, often associated with respiratory tract infections and diseases [1], In cases of chronic coughs, i.e. coughing over a period of 8-weeks or greater [2], coughing behaviour may severely affect quality of life and is associated with comorbidities, such as sleep apnea, asthma, and chronic obstructive pulmonary disorder (COPD). In Canada, the prevalence of chronic cough in middle age and older adult populations (45-85 years) is estimated at 16-18%. Moreover, management of chronic cough has been identified as a significant burden on health care systems [3], Questionnaires for assessing the impact of coughs on quality of life (e.g. Leicester Cough Questionnaire [4], Cough-Specific Quality -of-Life Questionnaire [5]) have been developed; however, the evaluation of coughing frequency and severity is also relevant for populations suffering from both acute and chronic cough. For example, an accurate and reliable count of coughs pre- and post- use of a pharmacological cough suppressant, is necessary to evaluate the effectiveness ofthe intervention (c.f. [6]). In that context, coughing quantification is often performed via questionnaires or manual review of audio recordings. Validated questionnaires exist for assessing cough frequency and severity (e.g. Cough Severity Diary [7]); however, they are dependent on patient compliance or recall, which may be especially difficult for nocturnal coughs. Manual review of audio recordings is potentially more accurate, but raises privacy concerns due to the collection and review of sensitive audio-data by a human annotator.

[0016] For example, an accurate and reliable count of coughs pre- and post- use of a pharmacological cough suppressant, is necessary to evaluate the effectiveness of the intervention (c.f. [6]).

[0017] Sensor modalities for respiratory activity classification

[0018] Sensors to monitor body measurements useful to respiratory activity classification can be body worn (e.g. embedded in a smart watch, glasses, clothing, or another accessory, or held on the body with a band or other contraption, or held on the skin or clothes using adhesives or clips) or in the environment of the person (e.g. microphone, camera, RF sensors, laser-based sensors, ultrasound sensors, Lidar, etc.).

[0019] Respiratory activity classification can be done using one of more of these sensors, and also if these sensors sometimes do not produce usable data or data at all, a method can be designed to account for changing availability of these sensors over time.

[0020] For example, the Hexoskin™ garment (Carre Technologies, Inc. Montreal, Canada) is a commercial smart shirt that comprises a three-dimensional accelerometer placed laterally on the trunk above the hip, two respiratory inductance plethysmography sensors at the thorax and abdomen, and a single channel electrocardiogram (ECG) sensor.

[0021] One advantage of using sensors in garments is the high level of social acceptability for clothing in general, and the possibility of placing sensors in various location on the body.

[0022] Physiological sensors have long been known and widely used for medical and health related applications. Various physiological sensors embedded in textile or garments, sometimes called portable or wearable sensors, have been described before in publications and patents (Portable Blood Pressure, U.S. Pat. No. 4,889,132, Filing date: Sep. 26, 1986 Issue date: Dec. 26, 1989; Portable device for sensing cardiac function, U.S. Pat. No. 4,928,690, Filing date: Apr. 25, 1988, Issue date: May 29, 1990). The term “wearable sensors” is now commonly used to describe a variety of body-worn sensors to monitor activity, environmental data, body signals, biometrics, health related signals, and other types of data.

[0023] Electrocardiogram (ECG) electrodes made of conductive textile, conductive polymer, metal and other materials used in wearable sensors have been described in patents such as (Textile-based electrode, U.S. Pat. No. 7,970,451, Filing date: Dec. 31, 2008, Issue date: Jun. 28, 2011).

[0024] Textile-based Respiratory Inductive Plethysmography sensors have been described in patents such as (Method and apparatus for monitoring respiration, U.S. Pat. No. 4,308,872 Issue date: Jan. 5, 1982).

[0025] Multi-parameter wearable connected personal monitoring systems (Zephyr BioHamess™, Qinetiq's Traintrak™, Nuubo's nECG™) are already available on the market.

[0026] There is thus a need for automatic extraction of physiological parameters to obtain breathing metrics such as but not limited to respiratory rate, tidal volume, minute ventilation and fractional inspiratory time.SUMMARY OF THE INVENTION

[0027] The shortcomings of the prior art are generally mitigated by a method to train a system for automatic detection of respiratory events and by a system for automatically detecting respiratory events.

[0028] Preferably, the physiological parameters extracted by the application of the present method are breathing metrics such as, but not limited to respiratory rate, tidal volume, minute ventilation and fractional inspiratory time.

[0029] According to an embodiment, the method may further comprise the step of detecting and characterizing physical conditions such as, but not limited to talking, laughing, crying, hiccups, coughing, asthma, apnea, sleep apnea, stress related apnea, relaxation exercise, breathing cycle symmetry, and pulmonary diseases.

[0030] In an aspect of the present invention, a method to train a system for automatic detection of respiratory events is provided. The method comprises capturing data from users using sensors on a wearable garment, annotating the collected data while the users are experiencing respiratory events, processing the annotated data and reducing selection features from the processed data maximizing information in the selected features for discrimination of respiratory events and training a model with the reduced features using a statistical machine learning algorithm.

[0031] The data which is collected may be accelerometric data, respiratory data and electrocardiographic (ECG) data. The processing of the data may further comprise modifyingfrequency of signals of the accelerometric data and ECG data to match frequency of a signal of respiratory data.

[0032] The processing of the data further may comprise splitting the annotated data into equal strips. The processing of the data may comprise splitting the data into training and test sets. The data may be manually annotated.

[0033] The respiratory events may be coughs and the coughs may be any of the followings: single, double, soft, normal and loud coughs. The respiratory events may further be selected from talking, laughing, crying, hiccups, coughing, asthma, apnea, sleep apnea, stress related apnea, relaxation exercise, breathing cycle symmetry, and pulmonary diseases.

[0034] The selection features may be reduced by a factor of at least 10. The selection features from the processed data maximizing impact on prediction of the detection of respiratory events may be the selection features associated with the highest correlation and lowest variance.

[0035] In another aspect of the invention, a system for automatically detecting respiratory events is provided. The system comprises a smart garment comprising a respiratory sensor, an electrocardiogram sensor and an accelerometer. The system further comprises a processor configured to receive data from the respiratory sensor, the electrocardiogram sensor and the accelerometer, a machine learning classifier configured to detect a cough event and a non-cough event based on the data collected by the smart garment, the machine learning classifier using a trained model comprising a plurality of selection features maximizing impact on prediction of detection of respiratory events, wherein the trained model is trained with the reduced features using a statistical machine learning algorithm and with annotated data collected from a plurality of users are experiencing respiratory events and a mobile power source powering the garment.

[0036] The selection features from the collected data of the system may be the selection features associated having the highest correlation and lowest variance.

[0037] The respiratory events of the system may be selected from talking, laughing, crying, hiccups, coughing, asthma, apnea, sleep apnea, stress related apnea, relaxation exercise, breathing cycle symmetry, and pulmonary diseases. The system may further comprise a cough database used to train the machine learning classifier.

[0038] Other and further aspects and advantages of the present invention will be obvious upon an understanding of the illustrative embodiments about to be described or will be indicated in the appended claims, and various advantages not referred to herein will occur to one skilled in the art upon employment of the invention in practice.BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The above and other aspects, features and advantages of the invention will become more readily apparent from the following description, reference being made to the accompanying drawings in which:

[0040] FIG. 1 is schematic representation of a general automatic input / output model for detection of respiratory events based on sensor inputs in accordance with to the principles of the present invention.

[0041] FIG. 2 is a schematic representation of method to validate and / or train a model for automatic detection of respiratory events based on sensor inputs in accordance with to the principles of the present invention.

[0042] FIG. 3 is a schematic representation of a method for selecting general automatic input feature for generating an optimal feature subset for discriminating between respiratory events and non-respiratory events in accordance with to the principles of the present invention.

[0043] FIG. 4 is a schematic representation of a general automatic machine learning model training using preprocessed sensor inputs and event class outputs based on annotations by an operator in accordance with to the principles of the present invention.

[0044] FIG. 5 is a schematic representation of a general automatic input data conditioning based on input and output data distribution and statistical characteristics in accordance with to the principles of the present invention.

[0045] FIG. 6 is schematic representation of a general automatic input / output model for detection of respiratory events using input conditioning and preprocessing in accordance with to the principles of the present invention.

[0046] FIG. 7 is a schematic representation of an embodiment of a respiratory event detection system based on a biometric garment in accordance with to the principles of the present invention.DETAILED DESCRIPTION OF THE INVENTION

[0047] A novel system and method Wearable biometric garment for automatic detection of respiratory events will be described hereinafter. Although the invention is described in terms of specific illustrative embodiments, it is to be understood that the embodiments described hereinare by way of example only and that the scope of the invention is not intended to be limited thereby.

[0048] In some embodiments, a smart garment, such as but not limited to a garment developed by Hexoskin, Carre Technologies Inc., Montreal, Canada, is used to capture data of a user. In such embodiment, the smart garment typically comprises an accelerometer, a respiratory (inductance plethysmography) sensor and an ECG (single-lead) sensor. In a preferred embodiment, the smart garment comprises a three-dimensional accelerometer positioned laterally on distal trunk (64 Hz), two respiratory sensors (respiratory inductive plethysmography bands) (128 Hz) placed across the abdomen and thorax, and a single lead electrocardiogram (ECG) sensor (256 Hz).

[0049] A method to train a system for automatic detection of respiratory events is provided. The method to train typically comprises collecting data from users, annotating the said data, processing the annotated data and extracting features from the processed data. Data is collected using the smart garment worn by a user while performing coughs, such as but not limited to single, double, soft, normal, loud and respiratory tasks, such as but not limited to breathing, talking, throat clearing, laughing, sniffing. In some embodiment, an audio signal is recorded using a recording device, such as but not limited to a smartphone. The audio signal is synchronized with the sensors data of the smart garment. In further embodiments, the data may be collected while the user is in supine and sitting (trunk 30° from horizontal). The collected data are then annotated, such as by an observer, and are processed.

[0050] In some embodiments, the annotation of data comprises each task being performed by the user being manually annotated based on the corresponding recorded audio signal recorded. In some embodiments, the processing of the data may comprise modifying the frequency of the signals from the sensors to match the frequency of the respiration sensor. As such, a signal outputted by the acceleration sensor signals may be up sampled and a signal outputted by the ECG may be down sampled to match the respiration sensor frequency (128 Hz). Understandably, any other method to match the frequency of the respiration sensor may be used within the scope of the present invention. The signal may be split into equal strips. In some embodiments, the signals may be split into strips having a predetermined length, such as 1 second and may be binarized into coughs (ex: n=l,207) and non-coughs (ex: n = 4,069).

[0051] The data may be split into training and test sets. The splitting between training and test sets generally ensures that different trials from the same user are not found in both sets and therefore improves generalizability of results. In some embodiments, the data split is performedusing 80:20 inter-subject (subject- wise) split scheme. The data may further be scaled to the range [0, 1] using Scikit-leam’s MinMaxScaler

[0013] , The scaler may be fit on the train set and applied to the test set.

[0052] The selection of features may comprise extracting features from for each strip using computer program. The computer program may comprise using a feature generation toolkit designed for wearable sensors, such as but not limited to FLIRT. The feature generation toolkit generally comprises feature extraction using statistical (e.g. mean, maximum, minimum) and frequency -domain features (e.g. Direct Fourier Transform coefficients) methods. The frequency-domain features may allow extraction of features based on the L2-norm (magnitude) of the x-, y-, and z- acceleration such as, but not limited to, entropy, skewness, and kurtosis. The number of features being computed may be up to 256 features. Understandably, the present invention is not limited to a maximum of 256 features.

[0053] In some embodiment, the extraction method may further comprise a step to reduce the number of features as the extraction method may produce duplicate, non-informative, or similar features across feature extraction. The step for reducing the number of features may comprise removing duplicate features, removing features with low-variance across labels, such as variance < 0.01, and / or removing highly correlated features, such as r > 0.95. The importance of each feature for the best performing model may be based on the mean decrease in impurity.

[0054] In yet other embodiments, the method for training a system for automatic detection of respiratory events may further comprise increasing the number of cough instances in the training set. To perform the increase of cough events, the synthetic minority class augmentation technique (SMOTE) may be used.

[0055] Using the method for training a system for automatic detection of respiratory events, a classifier model is developed. In a preferred embodiment, Random Forest classifier models are developed, such as using Scikit-leam. The classifier models may be tuned using a 5 -fold-cross validation of the training data. In some embodiments, the model is tune using the SkLeamTuner class available. The models are generally tuned to obtain an optimized Area Under the Curve (AUC) value.

[0056] An exemplary list of tunable hyperparameters in the models is shown in Table 1. The method may comprise identifying a final model using hyperparameters minimizing the cost function 403 and applied to the test set for each model type. Thus, hyperparameter values may differ across models. As such, in some embodiment, the performance of a model being developed may be evaluated an implementation of the cost function 403 using the AUC metric.The development of the model may further comprise computing accuracy and sensitivity of the developed model, such as but not limited to computing a Fl -score. The above step may be repeated on models being developed using different sensor inputs (acceleration, respiration, ECG, acceleration and respiration, all sensors). Understandably, any cost function 403 may be used within the scope of the present invention. Some examples of cost functions 403 includeAUC, Fl, and accuracy and sensitivity.Table 1 : Example of hyperparameters may be used for tuning the Random Forest Model based on a manually annotated datasetTable 2: Example of effect of sensor input on Random Forest Classifier model performance (per cent) based on a manually annotated datasetUsing a model trained using the above-described method, the data captured by the sensors of the smart garment allows predicting coughs from non-cough events.

[0057] Referring now to FIG. 1, an embodiment of a general automatic input / output model for detection of respiratory events based on sensor 100 inputs is illustrated. The model 100comprises sensors 102, a respiratory event model 104 adapted to detect respiratory events and to present the said results 106. The sensors 102 are adapted to capture data relating to a condition of the user and to input the same in the respiratory event model 104.

[0058] Referring to FIG. 2, a method to validate and / or train a model for automatic detection of respiratory events 200 based on sensor inputs 102 is illustrated. The method 200 typically comprises equipping a subject with wearable sensors 201, the subject producing coughing events. The method 200 further comprises recording the subject while coughing using any preferred sensor 202, such as the wearable sensors, a microphone, sound capturing device, a video sensor, such as a video camera, and / or a vibration detector. Understandably, any type of sensor adapted to gather data regarding the coughing subject. The method 200 further comprises an observer annotating and storing observed coughing events 203 while the sensors record data on the subject 202. The method 200 further comprises storing a timestamp 204 for each data captured from the sensors 202 or annotated by the observer 203. All the collected data are synchronized 205 based on the timestamp.

[0059] Referring to FIG. 3, a method for selecting general automatic input feature for generating an optimal feature subset for discriminating between respiratory events and non- respiratory events 300 is illustrated. The method 300 comprises gathering data from wearable sensors on a subject 301 and detecting event category or class 302. The method further comprises selecting features from a list of known event features 303, such as but not limited to frequency spectrum power, signal statistics (integral, kurtosis, skewness, etc.) and wavelet coefficients. The method 300 further comprises conditioning input data and applying preprocessing functions over the said input data 304 based on the selected features.

[0060] Referring to FIG. 4, a method for general automatic machine learning model training using preprocessed sensor inputs and event class outputs based on manual annotations by an operator 400 is illustrated. The machine learning model trainer 405 generally uses a cost function 403 to optimize the model according to desired characteristics, e.g. to minimize false positives in a detection task where positive events are not very common. The method 400 comprises using the data conditioned and preprocessed 304 as input 401 to the model trainer 405. The method 400 further comprises using the event class outputs as input to the model trainer 402. The method 400 comprises inputting a cost function 403 to optimize the trainer model 405 according to desired characteristics, e.g. to minimize false positives in a detection task where positive events are not very common. The method 400 thus comprises applying the cost function and the data conditioned and preprocessed as input to the model trainer 405 tooutput an optimized machine learning model 406.

[0061] Referring now to FIG. 5, an embodiment of an automatic method of input data conditioning based on input and output data distribution and statistical characteristics 500 is illustrated. The method 500 comprise extracting data from an input / output data source or database 501. The extracted data is then conditioned and preprocessed 502. The conditioned and preprocessed data is stored in a normalized input / output data source or database 503.

[0062] Referring now to FIG. 6, an embodiment of an automatic input / output model for detection of respiratory events with input conditioning and preprocessing 600 is illustrated. The model 600 comprises one or more sensors 601, such as wearable sensors, used as input. The model 600 further comprises a module for conditioning and preprocessing the said input 602.

[0063] Referring to FIG. 7, an embodiment of a respiratory event detection system 700 based on a biometric garment 710 is illustrated. The garment 710 comprises a body 711, an accelerometer 712, a respiratory sensor 713 and an ECG sensor 714. The sensors are connected to a processing device (not shown) typically integrated to the garment 710. The system 700 further comprises a cough data source 720, such as a relational database. The cough data source 720 is configured to store events captures from the sensors 712, 713 and 714 of the garment 710 from a plurality of users. The system 700 is configured to select channel 701 from the accelerometer 712, respiration sensor 713 and / or electrocardiogram 714. The selected channel is stored in the cough data source 720. The system 700 is further configured to switch between a feature based model or a parametric model 702.

[0064] In cases using a feature-based model, the system 700 is configured to compute features from the cough data source 720 and organize the computed features in a matrix of trials and features 703. The system 700 further selects the best features from the computed and organized features 704. The selection 704 may be performed using correlation and variance. The features identified as the best 704 are inputted to a machine learning classifier 740 trained to determine if data provided by the sensors of the garment 710 are associated to a cough event 708 or a noncough event 707.

[0065] In cases using a non-feature-based model, the system 700 is configured to organize raw data from the sensors as (trials x frames x channels) tensor 705. The tensor from the organized data is inputted in a deep learning classifier 730 trained to determine if data provided by the sensors of the garment 710 are associated to a cough event 708 or a non-cough event 707.

[0066] While illustrative and presently preferred embodiments of the invention have been described in detail hereinabove, it is to be understood that the inventive concepts may beotherwise variously embodied and employed and that the appended claims are intended to be construed to include such variations except insofar as limited by the prior art.

Claims

Claims1) A method to train a system for automatic detection of respiratory events comprising: capturing data from users using sensors on a wearable garment; annotating the collected data while the users are experiencing respiratory events; processing the annotated data; and reducing selection features from the processed data maximizing information in the selected features for discrimination of respiratory events; training a model with the reduced features using a statistical machine learning algorithm.2) The method of claim 1, the data being collected being accelerometric data, respiratory data and electrocardiographic (ECG) data.3) The method of claim 2, the processing of the data further comprising modifying frequency of signals of the accelerometric data and ECG data to match frequency of a signal of respiratory data.4) The method of claim 1, the processing of the data further comprising splitting the annotated data into equal strips.5) The method of claim 1, the processing of the data further comprising splitting the data into training and test sets.6) The method of claim 1, the data being manually annotated.7) The method of claim 1, the respiratory events being coughs.8) The method of claim 7, the coughs being any of the followings: single, double, soft, normal and loud coughs.9) The method of claim 1, the respiratory events being selected from talking, laughing, crying, hiccups, coughing, asthma, apnea, sleep apnea, stress related apnea, relaxation exercise, breathing cycle symmetry, and pulmonary diseases.10) The method of claim 1, wherein the selection features are reduced by a factor of at least 10.11) The method of claim 1 wherein the selection features from the processed data maximizing impact on prediction of the detection of respiratory events are the selection features associated with the highest correlation and lowest variance.12) A system for automatically detecting respiratory events comprising: a smart garment comprising: a respiratory sensor;an electrocardiogram sensor; an accelerometer; a processor configured to receive data from the respiratory sensor, the electrocardiogram sensor and the accelerometer; a machine learning classifier configured to detect a cough event and a non-cough event based on the data collected by the smart garment, the machine learning classifier using a trained model comprising a plurality of selection features maximizing impact on prediction of detection of respiratory events, wherein the trained model is trained with the reduced features using a statistical machine learning algorithm and with annotated data collected from a plurality of users are experiencing respiratory events; a mobile power source powering the garment.13) The system of claim 12, the selection features from the collected data are the selection features associated having the highest correlation and lowest variance.14) The system of claim 12, the respiratory events being selected from talking, laughing, crying, hiccups, coughing, asthma, apnea, sleep apnea, stress related apnea, relaxation exercise, breathing cycle symmetry, and pulmonary diseases.15) The system of claim 12 further comprising a cough database used to train the machine learning classifier.

Images