Respiratory and gastric functions monitor

Abstract

The invention is a system for continuous monitoring of respiratory and gastric functions using wearable devices. Each device includes a three-axis accelerometer and optionally a gyroscopic sensor. They feature wireless communication for data transmission and means for attachment to the skin. A computation device analyzes the data, using a machine learning model to detect coughing, swallowing and respiration. The method involves placing sensors on the skin of the cricoid region and epigastric region, filtering and transmitting data, training the model, and plotting coughing frequency. This system enhances detection accuracy and facilitates early medical interventions, benefiting chronic condition management and postoperative care.

Description

[0001] RESPIRATORY AND GASTRIC FUNCTIONS MONITOR

[0002] FIELD OF THE INVENTION

[0003] The field of the invention pertains to medical monitoring systems, specifically to a system and method for continuous, non-invasive monitoring of respiratory and gastric functions. The invention utilizes wearable devices equipped with three-axis accelerometers and optionally gyroscopic sensors to detect and distinguish swallowing, coughing, and other respiratory-related actions. The system is designed for use in various medical settings, including intensive care units, regular hospital wards, and home monitoring, providing real-time data through wireless communication to a computation device that processes the data using machine learning algorithms. This technology is also applicable in sports physiology, high- altitude activities, fitness, fall detection, and veterinary medicine.

[0004] BACKGROUND

[0005] Oropharyngeal swallowing involves a rapid, highly coordinated set of neuromuscular actions beginning with lip closure and terminating with the opening of the upper esophageal sphincter. Dysphagia, a condition affecting up to 40 million EU citizens, can result in the entry of food or liquid below the true vocal cords, known as aspiration, which is a common consequence of this condition. The clinical consequences of dysphagia are directly linked to the patient's overall prognosis and may include aspiration pneumonia, malnutrition, and dehydration. Early identification of dysphagia in stroke survivors has been shown to reduce morbidity and mortality. Current methods for screening dysphagia primarily involve clinical examination using a test swallow on demand, and there is no consensus on the best practice for populations other than stroke survivors. Instrumental assessments like Video- fluoroscopic Swallowing Study (VFSS) and Fiberoptic Endoscopic Evaluation of Swallowing (FEES) do not target spontaneous swallowing, which is a significant indicator for dysphagia and pneumonia detection. Despite the importance of spontaneous swallowing frequency (SSF) as an indicator, there is no validated bedside screening tool for automated screening of dysphagia and aspiration risk. Existing methods to measure SSF involve models combining audio and electromagnetic signals, which are more complex and susceptible to interference, making them less reliable in settings like Intensive Care Units (ICUs) or residential care. Additionally, coughing, which is related to dysphagia and aspiration pneumonia, is usually evaluated based on subjective patient assessments. Objective measures for coughing are limited and often cumbersome to integrate into clinical settings. Wearable sensors have been developed to continuously monitor respiratory rates and movements, but these devices often face challenges related to discomfort and instability. Currently, there is no non-invasive, continuous monitoring device that integrates the detection of swallowing, respiration, and coughing, nor is there a system that combines accelerometry and gyroscopic sensors to provide accurate monitoring of these functions along with corporal motion detection.

[0006] EP4076176, owned by Koninklijke Philips NV, focuses on monitoring abnormal respiratory events using a sensor system that operates at two different frequencies. A lower frequency for detecting respiration anomalies such as coughs and wheezes. When an anomaly is detected a higher frequency is employed to facilitate a more detailed analysis of the anomaly. The system uses a sternum-worn vibration sensor (accelerometer) that limits data collection to a single point above the sternum. This restricts the accuracy and breadth of data captured. Additionally, the frequencyswitching mechanism adds complexity to the hardware and software, potentially increasing costs and reducing system reliability. Furthermore, this technology is primarily designed for respiratory monitoring, lacking capabilities to track other physiological functions like swallowing.

[0007] EP3806737, held by Strados Labs Inc., describes a method and device for detecting physiological events, particularly focusing on respiratory sounds and movements through a wearable device that captures both motion and audio data. This dual-data approach leads to higher energy consumption, which can significantly affect the device's battery life. Furthermore, the inclusion of audio sensors not only raises the device's complexity and cost but also introduces privacy concerns, especially pertinent in settings where confidentiality is crucial. Moreover, accurately differentiating cough sounds from background noise remains a technical challenge, risking the reliability of the detected data.

[0008] EP3923780, by Societe des Produits Nestle SA, is centered on screening swallowing impairment using an integrated device that receives and analyzes vibrational data from swallowing events. This device, which is positioned on the neck, specifically evaluates swallowing safety and efficiency. However, its focus is narrowly confined to swallowing, without capacity for broader respiratory or cough monitoring. The sensor's placement on the neck requires precise positioning, which can be cumbersome and uncomfortable for continuous use, limiting its practicality in dynamic clinical environments. Additionally, the reliance on a single sensor may not provide a comprehensive overview of the patient's physiological state, and the setup's complexity could hinder its deployment in fast-paced settings.

[0009] US20230190125 describes a method that involves collecting non-invasive signals corresponding to physiological data from a subject by photoplethysmography (PPG) sensor or PPG enabled device and accelerometer. The collected signals are processed to generate physiological data, such as accelerometry data, associated with the subject. Physical acts of coughing are detected from the physiological data by monitoring blood volume changes captured by the PPG sensors. The physical act comprises inhalation, exhalation against dosed glottis, opening of glotti or relaxation. Auxiliary data is incorporated to provide context on conditions under which the signals are collected.

[0010] WO2019241674 describes a method that involves receiving motion data from a sensor of a wearable device worn by a user. An audio data is received from the sensor of the wearable device, where the audio data representative of sounds emanates from the users respiratory system. The motion data is compared to a motion data criteria. The audio data is compared to an audio data criteria. Determination is made to check whether the user is coughed based on the comparison of the motion data to the motion data criteria and the comparison of the audio data to the audio data criteria.

[0011] US20230118304 describes a system that has a sensory instrumentation unit. Abdominal contraction sensor and acoustic sensor collect and transmit raw sensory signals to the sensory instrumentation unit. The sensory instrumentation unit affects the conditioning of respective signals from the abdominal contraction sensor and from the acoustic sensor. A data acquisition unit is obtained by the sensory instrumentation unit and the acoustic sensor activation unit of the microcontroller are interrelated. One data flow unit is selected from the group consisting of a data transmission and reception unit and a data storage unit. The data flow unit is controlled by the microcontroller.

[0012] ON 109498228 describes a device that has an acoustic monitoring module connected with a data integration component and a data processing unit. A feedback treatment module is provided with the data processing unit and a data application component. The data integration component is connected with an acoustic detection component. An acoustic transducer is connected with an acceleration sensor. The data processing unit is connected with a normalized sample base. The data application component is connected with a feedback information display component.

[0013] The present invention aims to resolve at least some of the problems and disadvantages mentioned above.

[0014] SUMMARY OF THE INVENTION

[0015] The invention relates to a system and method for monitoring respiratory and gastric function using wearable devices and a computation device.

[0016] The system comprises multiple wearable devices, each equipped with at least one three-axis accelerometer, and optionally, a three-axis gyroscopic sensor. These devices are designed to be worn on the skin of a patient and are capable of wireless communication with a central processing unit. The computation device includes a processing section, a communication section to receive data from the wearable devices, and a memory section containing a trained machine-learning model that detects when a patient coughs or swallows.

[0017] The method involves placing at least two sensors on the patient's skin, specifically on the skin of the cricoid region and the epigastric region, collecting raw data on acceleration and gyroscopic signals, filtering this data to remove noise, and transmitting it to the computation device. The machine learning model is then trained with the filtered data to distinguish coughing and swallowing from other anomalies in breathing patterns, such as humming, scraping, speaking, and patient movements.

[0018] The system offers several advantages: the machine learning model can be tailored to specific patient needs, allowing for precise monitoring and management of respiratory and gastric functions. Additionally, the model's continuous improvement and adaptability enable predictive analytics, facilitating early identification of deteriorating health conditions. This system provides a non-invasive, continuous monitoring solution that enhances patient comfort and adherence, improves early detection of complications, and supports timely medical interventions. Optionally, the system can detect falls and track respiratory rates, offering a comprehensive health monitoring tool. Furthermore, in a preferred embodiment, the invention includes real- time data processing and display capabilities and can send emergency messages or alarms when critical thresholds are breached, ensuring rapid response to emergencies. The system is suitable for use in various healthcare settings, including home care, postoperative care, pediatric care, and remote or underserved locations, making it a versatile and valuable addition to modern healthcare.

[0019] DESCRIPTION OF FIGURES

[0020] The following description of the figures of specific embodiments of the invention is merely exemplary in nature and is not intended to limit the present teachings, their application or uses. Throughout the drawings, corresponding reference numerals indicate like or corresponding parts and features.

[0021] Figure 1 shows a schematic representation of a human body and illustrates the position of two wearable devices on the human body.

[0022] DETAILED DESCRIPTION OF THE INVENTION

[0023] The invention concerns a wearable system designed to monitor respiratory and gastric functions, capable of accurately detecting, among other activities, both coughing and swallowing.

[0024] Unless otherwise defined, all terms used in disclosing the invention, including technical and scientific terms, have the meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. By means of further guidance, term definitions are included to better appreciate the teaching of the present invention.

[0025] As used herein, the following terms have the following meanings:

[0026] "A", "an", and "the" as used herein refer to both singular and plural referents unless the context clearly dictates otherwise. By way of example, "a compartment" refers to one or more than one compartment.

[0027] "Comprise", "comprising", and "comprises" and "comprised of" as used herein are synonymous with "include", "including", "includes" or "contain", "containing", "contains" and are inclusive or open-ended terms that specify the presence of what follows e.g. component and do not exclude or preclude the presence of additional, non-recited components, features, element, members, steps, known in the art or disclosed therein.

[0028] Whereas the terms "one or more" or "at least one", such as one or more or at least one member(s) of a group of members, is clear per se, by means of further exemplification, the term encompasses inter alia a reference to any one of said members, or to any two or more of said members, such as, e.g., any >3, >4, >5, >6 or >7, etc. of said members, and up to all said members.

[0029] The term "system" refers in the present invention to a combination of hardware and software components designed to work together to monitor respiratory and gastric functions.

[0030] The term "wearable devices" is meant in the present invention devices that can be comfortably worn on the skin of a patient for continuous monitoring. These devices typically include sensors and communication units.

[0031] The term "three-axis accelerometer" refers to a sensor capable of measuring acceleration along three Cartesian axes (x, y, and z). This sensor provides data on the movement and orientation of the device.

[0032] The term "cricoid region" refers to the region of skin located in the neck, below the Adam's apple.

[0033] The term "epigastric region" refers to the upper central part of the abdomen, located just below the sternum.

[0034] The term "dead zone" refers to a predefined period immediately following certain actions, such as a cough or scrape, during which the system does not register new swallowing events to prevent false positives.

[0035] The term "Real-time" refers to systems or processes that respond immediately to input, executing tasks or updating information without perceptible delay.

[0036] Unless otherwise defined, all terms used in disclosing the invention, including technical and scientific terms, have the meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. By means of further guidance, definitions for the terms used in the description are included to better appreciate the teaching of the present invention. The terms or definitions used herein are provided solely to aid in the understanding of the invention.

[0037] Reference throughout this specification to "one embodiment" or "an embodiment" means that a particular feature, structure or characteristic described in connection with the embodiment is included in at least one embodiment of the present invention. Thus, appearances of the phrases "in one embodiment" or "in an embodiment" in various places throughout this specification are not necessarily all referring to the same embodiment, but may. Furthermore, the particular features, structures or characteristics may be combined in any suitable manner, as would be apparent to a person skilled in the art from this disclosure, in one or more embodiments. Furthermore, while some embodiments described herein include some but not other features included in other embodiments, combinations of features of different embodiments are meant to be within the scope of the invention, and form different embodiments, as would be understood by those in the art. For example, in the following claims, any of the claimed embodiments can be used in any combination.

[0038] In a first aspect, the invention concerns a system for monitoring respiratory and gastric function, comprising: a plurality of wearable devices, each configured to be worn on the skin of a patient, and each comprising: at least one three-axis accelerometer; optionally, a gyroscopic sensor; a communication unit configured to establish wireless communication with a processing unit; means for attachment to the skin of a patient, a computation device, comprising: a processing section; a communication section configured to receive data from each of the sensor devices; a memory section comprising a trained machine learning model configured to detect when a patient coughs and swallows, whereby at least one wearable device is suited to be placed on the skin of the cricoid region, and whereby at least one wearable device is suited to be placed on the skin of the epigastric region.

[0039] By integrating a plurality of wearable devices, each outfitted with a three-axis accelerometer, the system provides continuous monitoring of minute motions and vibrations corresponding to respiratory and gastric functions. This setup allows for a nuanced collection of physiological data, which is particularly beneficial for detecting and analysing the intricacies of patient-specific coughing and swallowing patterns. One of the benefits of this system is its capability for early detection of potential complications and due risk stratification. Conditions such as dysphagia, which can lead to severe outcomes like aspiration pneumonia if not managed properly, and various pulmonary, neurological and cardiac diseases can be monitored more effectively. The data collection helps in identifying changes in swallowing patterns promptly, thus enabling timely medical interventions.

[0040] The combination of two sensor locations, namely the cricoid region and the epigastric region, allows for the discrimination of swallowing from coughing, speaking, or unrelated torso movement, based on their distinct biomechanical profiles. The invention thereby achieves dual-function monitoring— of both cough and swallow events— through a minimally obtrusive and sensor-efficient design.

[0041] Importantly, the use of only two wearable sensors placed at the cricoid and epigastric regions makes the system particularly suitable for long-term and continuous use. Both positions allow for comfortable and stable attachment. Because the sensors are configured to be positioned where they do not interfere with daily movement or speech, patients can wear the system for extended periods without significant discomfort, making it well-suited for continuous ambulatory monitoring in hospital or home settings. This stands in contrast to existing systems that require bulky sensor arrays or audio capture devices, which limit wearability and raise privacy concerns.

[0042] The ability to monitor coughing and swallowing simultaneously and continuously is of critical clinical importance, as the temporal and functional relationship between these two actions serves as a key biomarker in the detection and management of several high-risk conditions. In particular, disorders such as dysphagia and aspiration are associated to swallowing dysfunction and an increased or abnormal coughing, where coughing and swallowing also influence each other. Patterns are not detectable by isolated monitoring of either function. Therefore, a system capable of reliably capturing both swallowing and coughing activity, and their timing relative to one another, is essential for the early detection of aspiration risk, the diagnosis of swallowing disorders, and the monitoring of disease progression or therapeutic response in vulnerable patient populations, such as individuals with neurogenic dysphagia, post-stroke complications, or degenerative neuromuscular conditions. The simultaneous detection of both coughing and swallowing using a single measurement system is therefore essential for preventive healthcare. The identification of sensor locations capable of delivering sufficient discriminatory signal features for both swallowing and coughing is technically complex. Accelerometric signals are inherently sensitive to a wide range of motion artefacts, and signal quality is highly dependent on both sensor orientation and anatomical coupling. As such, the effectiveness of detection is tightly linked to the precise placement of the sensors on the body. Small deviations from optimal placement may result in signal contamination or insufficient resolution of the targeted physiological events. The system disclosed in the present invention provides a technically elegant and robust solution to this challenge by defining two fixed, empirically validated sensor locations that yield reproducible and discriminative data for both targeted functions.

[0043] This is in contrast to prior art documents, which fail to disclose or suggest such an optimized sensor configuration. US20230190125, for instance, provides no instruction concerning specific sensor placement, and even states that the accelerometer may be positioned near, rather than on, the body. Furthermore, US20230190125 teaches that cough detection is primarily performed using a photoplethysmography (PPG) sensor, with the accelerometer serving a supplementary role. It offers no indication— explicit or implicit— that coughing could be reliably detected using solely the data from an accelerometer, nor that sensor location on the body surface is critical.

[0044] WO2019241674, while referring to a single accelerometer placed on the anterior / superior chest wall for cough detection, fails to propose alternative locations and does not address the possibility of using multiple accelerometers for multiparameter detection. The focus remains limited to respiration monitoring with no suggestion that swallowing detection could be combined within the same system. WO2019241674 does not suggest nor imply that placement on both the cricoid and epigastric regions would improve or even enable simultaneous detection of both cough and swallow events. WO2019241674 concludes that only a single location on the chest wall is suitable, which is counter to the objective of the current invention.

[0045] US20230118304 concerns the detection of cough using two electromyography (EMG) sensors, both positioned at the epigastric region. Accelerometers are only employed to estimate posture and orientation and are not considered viable for cough detection. Furthermore, US20230118304 emphasizes that both sensors are co-located, and it does not suggest that sensor diversity in placement could enhance event detection performance. Thus, the document does not provide any incentive to use accelerometers alone, nor to place them at anatomically distinct locations such as the cricoid and epigastric regions. US20230118304 goes against the objective of this invention by mentioning that both sensors are co-located and by using the EMG sensor for cough detection.

[0046] CN 109498228 describes cough detection using an accelerometer in combination with an audio sensor. The implication in CN109498228 is that accelerometry alone is insufficient, reinforcing the reliance on multi-modal sensing strategies that increase system complexity. Instead an acoustic sensor is placed on the manubrium of the sternum, the suprasternal notch, or the anterior midline of the cricoid cartilage. Acoustic sensors are not preferred for continuous monitoring, as they are intrusive on the privacy patient due to the capture of speech and environmental sounds. CN 109498228 does not mention or suggest positioning at least two accelerometer, one at the cricoid region, and one at epigastric region, nor does it provide teaching related to the combination of cough and swallow detection. Instead, CN109498228 uses a single accelerometer, placed on thoracic spine on the back or the anterior chest wall, to detect movement of the body during a cough, is inadequate for the aims of the current invention.

[0047] By providing a system comprising at least two wearable accelerometer sensors configured for placement at the cricoid and epigastric regions, the present invention achieves a system for the simultaneous detection of swallowing and coughing with sufficient accuracy to be effectively used in early detection of signs of dysfunction in coughing and swallowing patterns. The ability to monitor both events continuously and in parallel contributes to a more comprehensive assessment of conditions such as dysphagia and aspiration, where the temporal relationship between swallowing and coughing is diagnostically relevant. As such, the invention offers a meaningful advancement over existing systems.

[0048] In an embodiment, a wireless communication unit in each device facilitates immediate data transmission to a central processing unit, allowing for long-term continuous monitoring and data analysis simultaneously. This aspect is particularly crucial in settings where immediate data analysis can lead to rapid decision-making and adjustment of care protocols. In an embodiment, multiple wearable devices are connected by a wire to a wireless communication unit that is suited to be placed on the skin, whereby the wearable devices measure simultaneously, sequentially or separately. This limits the number of wireless communication units needed for communicating with a computation device.

[0049] In an embodiment, each wearable device comprises, or is connected to, a memory for storing accelerometer, and preferably gyroscopic, data. Said data is stored in the memory in the case that the wireless communication unit fails to establish communication with the central processing unit. After communications have been reestablished the data stored in the memory is sent to the central processing unit and removed from the memory. This prevents important data from getting lost if the communication signal is unstable.

[0050] Additionally, the memory section of the computation device, which houses a trained machine-learning model, allows for sophisticated data processing. Unlike traditional systems that might rely on manual interpretation of data, this machine learning model can automatically detect nuances in coughing and swallowing, reducing the likelihood of human error and increasing the accuracy of assessments.

[0051] Furthermore, the use of small wearable devices ensures minimal intrusion and greater comfort for the patient, promoting longer periods of monitoring compared to traditional methods that might require more invasive procedures or intermittent assessments. This continuous monitoring is crucial for patients with chronic conditions like COPD and asthma, where fluctuations in symptoms are common and can occur unexpectedly. The comfort and wearability of the invention significantly expand its potential applications across various healthcare settings. The system can be seamlessly integrated into the daily lives of patients, ensuring that the monitoring does not disrupt their normal activities or reduce their quality of life. For example, elderly patients or those with mobility issues can wear these devices without feeling burdened, which is crucial for adherence to treatment protocols and consistent data collection.

[0052] Furthermore, the wearability aspect of this technology significantly transforms postoperative care. Patients recuperating from surgeries related to the respiratory or digestive systems, gain from the continuous, monitoring that these devices offer, eliminating the need for continuous clinical oversight. This capability could notably shorten hospital stays and facilitate a more comfortable, home-based recovery while still being closely monitored by healthcare providers who can view and respond to data sent from the devices.

[0053] In pediatric care, the non-invasive and comfortable nature of the devices encourages cooperation from young patients who might be averse to more cumbersome or intimidating medical equipment. This can lead to better compliance and more accurate health monitoring in children suffering from conditions like asthma or congenital respiratory anomalies.

[0054] The adaptability of the system to non-clinical environments also opens up possibilities for its use in remote or underserved locations where traditional medical equipment may not be feasible. This can significantly enhance healthcare delivery in rural or developing areas, providing crucial data that can be used to make informed medical decisions from a distance.

[0055] In another preferred embodiment, the wearable devices comprise a three-axis gyroscopic sensor. Incorporating a three-axis gyroscopic sensor alongside a three- axis accelerometer in each wearable device enhances the monitoring of respiratory and gastric functions by capturing both angular movements and orientations. This dual-sensor setup significantly improves the accuracy in detecting coughs and swallows, crucial for managing conditions like dysphagia and chronic respiratory diseases.

[0056] In another preferred embodiment, the wearable device is a MoveSense MD sensor with a male ECG-electrode connected to it by means of a connection piece from a 3M red dot electrode.

[0057] In another preferred embodiment, the invention comprises two wearable devices. Using two sensors in the wearable monitoring system strikes an optimal balance between comprehensive data collection and minimizing device complexity. Fewer sensors simplify the manufacturing and maintenance of the devices, lowering costs and improving reliability. Consequently, the dual-sensor configuration offers detailed health monitoring without the complexity of managing more than two sensors, making the system more user-friendly and accessible for continuous, everyday use. In another embodiment, the system is integrated in clothing, so that the wearable devices are placed in the right location on the skin when said clothing is worn. This makes the system more user-friendly by simplifying its use and increasing its comfort.

[0058] In another preferred embodiment, the trained machine learning model is configured to detect a fall. Integrating a trained machine learning model that detects falls into the dual-sensor wearable system greatly enhances its utility, especially for the elderly or those with mobility issues. The capability for immediate fall detection reassures users and caregivers and ensures timely assistance and support. By broadening its functionality to include safety monitoring, the device becomes a more comprehensive tool for health and safety management in both clinical and home environments.

[0059] In another preferred embodiment, the trained machine learning model is configured to detect respiratory rate. Tracking respiratory rate is beneficial in various medical, sports, and high-altitude activities due to its ability to provide early warnings of potential health issues and aid in health monitoring and management. Continuous monitoring of respiratory rates can lead to early detection of conditions like pneumonia, apnea, and respiratory distress. Furthermore, in sports physiology, it helps in assessing an athlete's performance and recovery, while at high altitudes, it serves as an essential tool for detecting acute mountain sickness and aiding acclimatization.

[0060] In another preferred embodiment, the system comprises a portable display in wireless communication with the computation device, configured to display detected quantities. The communication with a portable display ensures the detected information is accessible anywhere. For healthcare professionals, this means they can make informed decisions more swiftly and monitor the effectiveness of interventions wherever they are. Furthermore, portable devices can leverage existing health monitoring apps and platforms, integrating with other health data to provide a comprehensive view of the patient's overall health. Additionally, for patients, particularly those in home care settings, the portable display enhances their awareness and understanding of their health status, fostering greater engagement in their care process. Preferably the computation device and the display are integrated into a single device, such as a smartphone or a smartwatch, reducing the number of devices needed. In a second aspect, the invention concerns a method for monitoring respiratory and gastric function using a system of wearable devices and a computation device, said method comprising the following steps: placing at least two sensors on the skin of a patient; receiving raw data from the at least one sensor, said data comprising information on acceleration and gyroscopic signals along three cartesian axes; filtering said data using noise filtering components; transmitting the filtered data to a computation device; within the computation device: training a machine learning model with the received filtered data; using the machine learning model to detect and distinguish coughing and swallowing from other action causing anomalies in a normal breathing pattern such as humming, scraping, speaking, and patients moving; plotting coughing frequency based on the output of the machine learning model, characterized in that, at least one sensor device is placed on the skin of the cricoid region and at least another sensor device is placed on the epigastric region.

[0061] The cricoid region's rigidity ensures that the sensors can accurately capture the mechanical vibrations and accelerations associated with various physiological activities such as swallowing, coughing, and breathing without being influenced by extraneous movements. The cricoid region is also relatively free from large muscle groups, which helps in obtaining clear and precise data.

[0062] The epigastric region, located just above the stomach and below the ribcage, covers the area of the diaphragm and upper abdominal muscles. This region is characterized by its significant involvement in both respiratory and digestive processes, experiencing pronounced movements during breathing, coughing, and speaking. The physical properties of the epigastric area, including its flexibility and range of motion, make it an excellent site for capturing dynamic changes. The diaphragm's contractions and the movement of the abdominal wall during these activities provide rich data that is crucial for accurate monitoring.

[0063] One of the primary advantages of incorporating a machine learning model is its ability to continuously improve and adapt over time. As more data is collected, the model refines its algorithms to increase accuracy in distinguishing between different actions. This capability is essential because it addresses a significant limitation of existing methods, which often struggle with high false positive rates and difficulty in accurately identifying specific activities, especially when the signals are similar or when multiple actions occur simultaneously. In a preferred embodiment, the training process involves using advanced filtering techniques to remove noise from the raw data received from the wearable devices, preferably a wavelet filter. This preprocessing ensures that the data fed into the model is clean and consistent, thereby enhancing the model's performance. Additionally, in an embodiment, the data from the wearable devices is normalized using an Euclidean norm. This removes the directional dependency of the wearable devices from the data, allowing the wearable device to be worn in any orientation.

[0064] In a preferred embodiment, the machine learning model is trained with labelled data from examples of various actions performed by participants. By analysing patterns in the data, the model learns to recognize the unique signatures of coughing, swallowing, talking and scraping movements. When the model is trained it is tested using a different dataset. By using a different dataset any biases in the model caused by biases in the training data can be discerned.

[0065] In a preferred embodiment, two models are trained using a different dataset. One dataset consists of annotated cough, swallow, scrap and speaking data from two wearable devices: one placed on the skin of the cricoid region and one placed on the epigastric region. The second dataset consists of annotated scrape, cough and speaking data from only one sensor placed on the epigastric region. The models are trained using cross-validation with several subsets of each corresponding dataset. Datasets are preferably divided into at least 10 subsets, and more ideally into 15 to 25 subsets

[0066] In a preferred embodiment, the datasets for training the models are based on at least 10 participants, and more preferably at least 25 participants. The datasets for testing the models are based on at least 5 participants, and more preferably on at least 10 participants. In another embodiment, a dataset acquired from a participant consist of data of at least two of each performed action, whereby the performed actions comprise coughing, swallowing, scrapping, and speaking.

[0067] In another preferred embodiment, a method is provided wherein the sensors are configured to record raw accelerometric and gyroscopic data at a sampling frequency of at least 100 Hz, more preferably at or above 200 Hz, and preferably at no more than 1000 Hz, and more preferably at no more than 600 Hz. The method comprises the placement of one sensor on the cricoid region and a second sensor on the epigastric region, and more preferably on the epigastric region of the thorax. The wearable devices are configured to wirelessly transmit raw data to a computation device comprising a custom-built dashboard, wherein time-stamped events may be monitored and annotated in real-time or retrospectively. In one embodiment, recordings comprise a plurality of action types including, but not limited to, dry swallowing, wet swallowing, coughing, speaking, breathing, and patient movement such as body turns. Preferably, at least five repetitions of each action are recorded, and each repetition is separated by at least a 10-second interval. In a preferred embodiment, dry and wet swallows are annotated as distinct classes, thereby improving granularity and clinical relevance of swallowing event detection.

[0068] In an preferred embodiment, the method further comprises synchronizing incoming raw signals from both sensors using Bluetooth time-stamping on the computation device side. In one embodiment, signals are resampled using linear interpolation to a uniform frequency, preferably to 100 to 300 Hz, in order to facilitate time-aligned processing.

[0069] In a preferred embodiment, the method comprises a step of signal preprocessing, which comprises multi-stage filtering including soft wavelet thresholding, more preferably with a Daubechies wavelet basis, and optionally supplemented with high- pass filtering of the cricoid accelerometric signal. In a further embodiment, the high- pass component is extracted by subtracting a low-pass Gaussian-filtered signal with a standard deviation of between 0.15 and 0.25 seconds, preferably 0.20 seconds. In a preferred embodiment, only wavelet filtering is applied to the gyroscopic channels of the cricoid sensor. For the thoracic sensor, a Euclidean norm is computed separately for the accelerometric and gyroscopic signals to emphasize global motion patterns while reducing dimensionality.

[0070] In a further preferred embodiment, the method comprises transforming the preprocessed signal into a structured form suitable for machine learning by applying a sliding window segmentation scheme. In a preferred embodiment, windows have a duration of at least 0.5 seconds and at most 2 seconds, and more preferably at least 0.8 seconds and at most 1.5 seconds, with a sliding interval of at most 0.1 seconds, more preferably at most 0.07 seconds. Each window is assigned a class label based on a majority vote over the constituent samples. In a preferred embodiment, for each signal channel and each window, a feature vector is extracted using a feature extraction library such as for example tsfresh. The feature extraction module computes at least 2000, preferably more than 4000, statistical, temporal, and frequency-domain descriptors per window. In a more preferred embodiment, feature selection is performed using a repeated subsampling strategy over a plurality of patient subsets, more preferably with at least 5 iterations, and wherein a feature is retained only if it is selected in a majority of iterations, more preferably in at least 4 out of 6 iterations.

[0071] In another preferred embodiment, the trained machine learning model is a gradient- boosted decision tree model, preferably a CatBoost classifier. In a preferred embodiment, the classifier is configured to output a probability distribution across at least five distinct classes: background, coughing, speaking, swallowing of liquid (wet), and swallowing of saliva (dry). In another preferred embodiment, the classifier is trained with a tree depth between 4 and 8, a learning rate between 0.01 and 0.05, and a number of iterations between 300 and 600, most preferably between 450 and 550. To address class imbalance, class weights may be assigned inversely proportional to class frequency.

[0072] In another preferred embodiment, the method further comprises a post-processing step for assigning temporally localized predictions to individual samples. In this embodiment, the probability distributions produced by the machine learning model for each windowed segment are assigned to the center time point of each corresponding window. Preferably, the probabilities are then spread across a symmetric time interval equal to the stride length of the window segmentation. This results in a continuous, time-resolved prediction profile.

[0073] In a preferred embodiment, the time-continuous probability signals are smoothed using a temporal Gaussian filter to reduce high-frequency noise. In one embodiment, a Gaussian kernel with a standard deviation between 0.05 and 0.2 seconds, more preferably between 0.08 and 0.13 seconds, is applied to each class-specific probability stream. Following smoothing, a final class label is assigned to each individual time sample based on the class with the highest smoothed probability at that time point. This process ensures temporally coherent event labeling and minimizes spurious class transitions. In another preferred embodiment, the system is configured to evaluate the impact of different sensor signal combinations on classification performance. In one embodiment, a plurality of sensor configurations is defined, each comprising different combinations of accelerometer and gyroscope signals from the cricoid and epigastric region sensors. In a preferred embodiment, at least four configurations are evaluated, comprising: (i) cricoid accelerometry alone; (ii) cricoid accelerometry and gyroscopy; (iii) cricoid and epigastric accelerometry; and (iv) all combined signals, including epigastric gyroscope data. In each configuration, a sensor modality is considered included if its derived features are incorporated into the model's feature input vector.

[0074] In another preferred embodiment, model evaluation is performed using a leave-one- patient-out cross-validation protocol across data collected from a plurality of patients, preferably at least 10, more preferably at least 25 patients. In each fold of the evaluation, the model is trained on data from all but one patient and tested on the held-out patient's data.

[0075] In a preferred embodiment, sample-based evaluation is performed by comparing predicted and ground truth class labels on a per-sample basis. Preferably, precision, recall, and Fl-score are computed separately for each class and averaged across all cross-validation folds. In a further embodiment, a single confusion matrix is constructed by aggregating predictions across all folds, and normalized such that diagonal entries represent class-specific recall scores.

[0076] In another preferred embodiment, event-based evaluation is applied to assess the temporal coverage of predicted action segments. In a preferred embodiment, for each annotated event, the proportion of its duration correctly predicted as the corresponding class is counted as a true positive. Any misaligned segment is counted as a false negative. This allows the construction of an event-based confusion matrix, wherein recall is computed as the fraction of each event class's duration that overlaps with a correctly predicted segment.

[0077] In a further preferred embodiment, the importance of signal sources is evaluated using the PredictionValuesChange metric provided by the CatBoost model. In one embodiment, the importance of each extracted feature is computed per fold, grouped by its originating sensor and signal type (e.g., cricoid gyroscope, epigastric accelerometer), and averaged across all folds. This allows for a quantitative assessment of which sensors contribute most significantly to the classifier's predictions, and thus informs future design choices for minimal or optimal sensor configurations.

[0078] In another preferred embodiment, the system is configured to be operable in clinical environments where actions occur spontaneously rather than as externally prompted tasks. In such use cases, the classification method may for example include a "background" category representing non-specific periods without targeted activity. Preferably, the machine learning model is trained on both annotated action events and prolonged background segments to improve robustness in realistic, continuous- use scenarios. For example, background segments may include patient rest, spontaneous posture adjustments, or breathing-related motion.

[0079] In a preferred embodiment, inter-individual variability is mitigated through baseline adaptation, wherein a short sequence of patient-specific recordings is used to normalize or adjust model behavior.

[0080] In an preferred embodiment, the system includes an annotation interface that supports time-synchronized labeling of recorded data by multiple clinicians or analysts. In a preferred embodiment, annotations made by one reviewer are crosschecked by another to minimize labeling bias and improve consistency. For instance, an analyst may tag the start of a swallowing event using sensor data patterns, and a second reviewer may confirm the boundary based on visual inspection of accelerometer and gyroscope signals.

[0081] In another preferred embodiment, the system supports deployment with varying combinations of sensing modalities depending on the target application. In a preferred configuration, a tri-axial accelerometer positioned on the epigastric region is included for robust detection of coughing, due to its sensitivity to chest wall motion. For example, cough events are typically characterized by high-amplitude, broadband bursts across all three acceleration axes. In another embodiment, a cricoid-placed accelerometer is used to capture laryngeal vibrations during speech and swallowing, which are typically low in amplitude and localized. In a further embodiment, a gyroscope co-located at the cricoid region may be optionally included to improve detection sensitivity to head and neck motion during swallowing, though its overall contribution may be context-dependent. In another preferred embodiment, the classification model is configured to either distinguish between different types of swallowing or aggregate them into a single general class, depending on clinical objectives. For instance, wet swallowing involving a water bolus can be differentiated from dry swallowing of saliva in research applications examining oropharyngeal function. Alternatively, in general monitoring use, both are grouped under a single "swallowing" label to reduce misclassification due to overlap in biomechanical features. Preferably, the model architecture supports both configurations with minimal retraining effort.

[0082] In another preferred embodiment, the system includes a calibration protocol, wherein a short set of task-specific recordings is collected from each new patient to improve personalization. For example, during a two-minute setup, the patient may be asked to perform three swallows, two coughs, and a standard phrase to generate baseline signal templates. This data is used to adjust normalization parameters or fine-tune model thresholds, thereby enhancing accuracy during subsequent monitoring.

[0083] In another preferred embodiment, the system is configured to provide real-time biofeedback for use during swallowing rehabilitation. In this embodiment, the system comprises a visual feedback interface, preferably in the form of a dashboard, configured to display one or more parameters related to the detected swallowing events. Examples of such parameters include the frequency of successful swallows, duration of each swallow, consistency in timing across repetitions, or the amplitude of the accelerometric signal. The feedback interface may be presented on a portable device, such as a tablet or smartphone, accessible to both the patient and the therapist.

[0084] In a further preferred embodiment, the system comprises a gamification module that allows patients to practice swallowing exercises through interactive feedback. The module may present targets or scores based on the number of correctly performed swallows or encourage repetition through visual or auditory cues. For example, a patient may be challenged to complete five swallows within a certain time frame, or to maintain a consistent swallowing pattern to progress in a level-based interface. Such gamified interactions promote therapy adherence and engagement, especially in long-term rehabilitation trajectories. In one embodiment, progress over time is visualized to motivate patient participation and to assist therapists in evaluating progress. In another embodiment, the system is configured for use in monitoring laryngeal function during speech rehabilitation, particularly after surgery involving the vocal folds or surrounding structures. In this embodiment, the system functions as a followup tool, enabling periodic assessment of phonatory behavior without the need for continuous therapist presence. Preferably, the system logs speech-related vibration events captured by the cricoid sensor and visualizes these events over time in a trend-based format. For instance, the number of voice-initiating events, the stability of phonation, or the amplitude variability in speech episodes may be tracked over days or weeks. This enables healthcare providers to evaluate vocal recovery trajectories and make informed decisions regarding ongoing therapy.

[0085] In a preferred embodiment, both swallowing and speech rehabilitation modes are configurable via software presets, enabling seamless switching between interactive training and passive monitoring. This dual-mode operation ensures that the system can be used across different phases of recovery, from intensive re-education to longterm follow-up. For example, a hospital-based therapist may enable real-time feedback during inpatient sessions, while a patient at home receives daily trend updates to monitor ongoing improvement.

[0086] In a preferred embodiment, the models are trained using a Random Forest classifier with a preferred tree depth of 4 to 8 and at least 800 estimators.

[0087] In another preferred embodiment, postprocessing techniques further enhance the model's accuracy, such as a median filter to smooth out small periods of incorrect predictions. In another embodiment, a dead zone is introduced to prevent false positives immediately following certain actions, after which neither the same nor any other action can occur. For example, immediately after a swallow, another swallow is unlikely to follow. In another embodiment, an action time of 1 second is defined for each action, during which no other action can occur. In another embodiment, periods during which a sensor on the epigastric region detects an action are set as periods when swallowing cannot occur.

[0088] In another preferred embodiment, the method further comprises processing and displaying all data in real-time. Real-time data processing ensures that any changes in the patient's respiratory or gastric functions are immediately detected and communicated. This instant feedback loop is crucial for early intervention, allowing healthcare providers to respond promptly to potential health issues such as respiratory distress, abnormal swallowing patterns, or sudden changes in vital signs. For patients with chronic conditions, continuous real-time monitoring can lead to better management of their health, and so reduce the risk of complications and hospital readmissions. For instance, real-time detection of increased coughing frequency can indicate a respiratory infection or exacerbation of a chronic condition, prompting immediate medical attention and adjustment of treatment plans. Moreover, displaying data in real-time on portable devices such as smartphones or smartwatches enhances accessibility and convenience. Patients and caregivers can monitor health metrics on-the-go, ensuring that critical information is always at hand. This feature also promotes patient engagement and self-management, as individuals can observe the effects of their activities, medications, or dietary changes in realtime, making it easier to adhere to treatment regimens and lifestyle modifications. In clinical settings, real-time data processing and display facilitate more efficient workflow for healthcare providers. It allows for continuous monitoring without the need for manual data collection and analysis, freeing up medical staff to focus on direct patient care.

[0089] In another preferred embodiment, the method further comprises the step of sending an emergency message and / or an alarm to at least one device external to the system when a value is detected within or outside a predetermined range. This function provides immediate notification to healthcare providers, caregivers, or family members when critical thresholds are breached, enabling rapid response to emergencies such as respiratory distress, aspiration, or abnormal swallowing patterns and potentially preventing severe health complications.

[0090] DESCRIPTION OF FIGURES

[0091] With as a goal illustrating better the properties of the invention figure 1 presents, as an example and limiting in no way other potential applications, a schematic representation of the human body and illustrates a preferred position of two wearable devices on the human body. A first wearable device (1), is positioned on the skin of the cricoid region (3). A second wearable device (2), is positioned on the epigastric region(4).

Claims

CLAIMS1. A system for monitoring of respiratory and gastric function, comprising: a plurality of wearable devices, each configured to be worn on the skin of a patient, and each comprising:- at least one three-axis accelerometer;- optionally, a gyroscopic sensor; a communication unit configured to establish wireless communication with a processing unit; means for attachment to the skin of a patient, a computation device, comprising:- a processing section; a communication section configured to receive data from each of the sensor devices; a memory section comprising a trained machine learning model configured to detect when a patient coughs or swallows and characterized by that at least one wearable device is suited to be placed on the skin of the cricoid region, and whereby at least one wearable device is suited to be placed on the skin of the epigastric region.

2. The system according to claim 1, whereby the wearable devices comprise a three-axis gyroscopic sensor.

3. The system according to any of the preceding claims, whereby the trained machine learning model is configured to detect a fall.

4. The system according to any of the preceding claims, whereby the trained machine learning model is configured to detect respiratory rate.

5. The system according to any of the preceding claims, further comprising a portable display in wireless communication with the computation device, configured to display detected quantities.

6. The system according to claim 6, whereby the portable display and the computation device are integrated on a single device, preferably on a smartphone or smartwatch.

7. The system according any of the previous claims for use in treating and / or preventing dysphagia and / or respiratory diseases.

8. A method for monitoring of respiratory and gastric function using a system of wearable devices and a computation device, said method comprising the following steps: placing at least two sensors on the skin of a patient; receiving raw data from the at least one sensor, said data comprising information on acceleration and gyroscopic signals along three cartesian axes;- filtering said data using noise filtering components;- transmitting the filtered data to a computation device; within the computation device:- training a machine learning model with the received filtered data; using the machine learning model to detect and distinguish coughing and swallowing from other action causing anomalies in a normal breathing pattern such as humming, scraping, speaking, and patients moving; plotting coughing frequency based on the output of the machine learning model, characterized in that, at least one sensor device is placed on the skin in the cricoid region and at least another sensor device is placed on the skin of the epigastric region.

9. The method according to claim 9, further comprising the step of detecting respiratory rate.

10. The method according to claim 9 or 10, further comprising the step of detecting falls.

11. The method according to any of the claims 9 to 11, further comprising the step of displaying detected values on at least one portable device, said portable device preferably being a smartphone or a smartwatch.

12. The method according to any of the claims 9 to 12, further comprising the step of sending an emergency message and / or an alarm to at least one device external to the system when a value is detected within or outside a predetermined range.

13. The method according to any of the claims 9 to 13, whereby all data is processed and displayed in real-time.

14. The method according to any of the claims 9 to 14, used in the treatment of patients between 60 and 90 years old, wherein said patients have a respiratory disease and / or dysphagia.

Images