Automatic identification system and identification method for neonatal respiratory abnormalities

Abstract

The application discloses a new-born baby respiratory abnormality automatic identification system and identification method, relates to the technical field of medical devices, and aims to solve the technical problem that mechanical damage appears on the skin of a sick child due to monitoring by an electrode sheet or a sensor, and comprises a monitoring module which can be detachably embedded into a preset installation groove in the top inner wall of an incubator, and comprises an FMCW millimeter wave radar sensor with a working frequency of 60GHz+ / -2GHz and a spatial resolution of 0.1mm+ / -0.02mm, a near-infrared non-contact blood oxygen sensor and a temperature and humidity sensor with an accuracy of + / -0.5 DEG C / + / -3%RH; the detection surface of the millimeter wave radar sensor is kept at a vertical distance of 30-50cm from the chest and abdomen of a premature baby, and the signal transmission power is controlled at-10dBm to-5dBm, so that the electromagnetic radiation does not affect the premature baby. The application has the advantages that no contact with the skin is needed in the whole process, and the skin health of the new-born baby is improved.

Description

Technical Field

[0001] This invention relates to the field of medical device technology, and more specifically, to an automatic identification system and method for abnormal breathing in newborns. Background Technology

[0002] In the field of neonatal intensive care, premature infants, especially those with extremely low birth weight (typically between 1000-1500g) or very low birth weight (typically below 1000g), have a very high incidence of respiratory abnormalities such as apnea, periodic breathing, and tachypnea due to the immaturity of their respiratory system and respiratory center. If these respiratory abnormalities are not identified and effectively intervened in a timely manner, they can easily lead to hypoxia, resulting in serious complications such as hypoxic-ischemic brain injury, and even threatening the infant's life. Therefore, accurate and reliable monitoring of the respiratory status of premature infants, especially those with extremely low / very low birth weight, and timely identification of respiratory abnormalities are key aspects of NICU care and have long been a challenging problem for medical staff. Timely and accurate identification and intervention of respiratory abnormalities are crucial for preventing hypoxic-ischemic brain injury and ensuring the infant's health and life.

[0003] Currently, the gold standard techniques for monitoring respiratory function in preterm infants are chest impedance measurement and pulse oximetry. Both of these methods are invasive contact monitoring techniques. Their working principle is to acquire respiratory-related signals by permanently attaching or contacting electrode pads or sensors with the infant's skin, thereby enabling the monitoring of the infant's respiratory status.

[0004] However, premature infants have thin and delicate skin with extremely low tolerance to external stimuli. During prolonged monitoring with electrodes or sensors, mechanical damage such as redness and breakage can easily occur due to factors such as adhesion pressure, friction between the skin and the electrodes / sensors, and the materials used in the electrodes / sensors themselves. Some infants may also experience allergic reactions to the materials of the electrodes or sensors, resulting in symptoms such as itching and rashes. In more severe cases, skin damage or allergic reactions may lead to local infections, further aggravating the infant's condition and causing additional suffering and health risks. Therefore, we propose an automatic neonatal respiratory abnormality identification system and method. Summary of the Invention

[0005] The purpose of this invention is to provide an automatic identification system and method for neonatal respiratory abnormalities, in order to solve the technical problem that monitoring with electrodes or sensors can cause mechanical damage to the skin of infants.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an automatic identification system for abnormal neonatal breathing, applied to a closed incubator in a neonatal intensive care unit (NICU), comprising a monitoring module, wherein the monitoring module is detachably embedded in a pre-set mounting slot on the inner wall of the top of the incubator, including an FMCW millimeter-wave radar sensor with a working frequency of 60GHz±2GHz and a spatial resolution of 0.1mm±0.02mm, a near-infrared non-contact blood oxygen sensor, and a temperature and humidity sensor with an accuracy of ±0.5℃ / ±3%RH; the detection surface of the millimeter-wave radar sensor is... The vertical distance between the chest and abdomen of the premature infant is maintained at 30-50cm, and the signal transmission power is controlled between -10dBm and -5dBm to avoid the influence of electromagnetic radiation on the premature infant. It is used to collect minute displacement signals in the chest and abdomen range of 0.1-5mm to obtain respiratory movement data. The near-infrared non-contact blood oxygen sensor has a detection wavelength of 660nm±10nm and 940nm±10nm and is used to collect blood oxygen saturation data of the fingertips or soles of the premature infant (sampling rate 1Hz). The temperature and humidity sensor is used to collect the temperature and humidity data of the incubator in real time.

[0007] The signal processing module is electrically connected to the monitoring module via a Type-C interface. It has a built-in signal conditioning circuit and a filtering algorithm unit. The filtering algorithm unit uses a wavelet transform algorithm based on the db4 wavelet base to perform 3-5 layers of decomposition filtering on the respiratory motion data collected by the millimeter-wave radar sensor. This removes interference signals caused by unconscious limb movements of premature infants (amplitude > 5 mm), airflow disturbances in the incubator, and temperature and humidity fluctuations, and extracts a pure respiratory waveform with a frequency range of 0.5-3 Hz.

[0008] The AI ​​analysis module, connected to the signal processing module via Ethernet, includes a multimodal feature model built on a bidirectional LSTM neural network. The input feature set of the multimodal feature model includes: respiratory rate (calculation accuracy ±1 breaths / min) and chest movement amplitude extracted from a clean respiratory waveform (calculation accuracy ±0.05 mm), and the rate of change of blood oxygen saturation (unit % / s) collected by a near-infrared non-contact blood oxygen sensor. The multimodal feature model identifies three types of respiratory abnormalities using the following rules:

[0009] Apnea: Disappearance of displacement signal in respiratory movement data for ≥2 seconds, with synchronous oxygen saturation decrease rate >1% / s and cumulative decrease >5%;

[0010] Periodic breathing: The difference in respiratory rate fluctuation within three consecutive respiratory cycles is >20 breaths / min, and the corresponding difference in chest wall movement amplitude is ≥30%, while there is no significant decrease in blood oxygen saturation (fluctuation ≤2%).

[0011] Rapid breathing: For preterm infants with a gestational age of <28 weeks, a respiratory rate of >75 breaths / min is maintained; for preterm infants with a gestational age of 28-32 weeks, a respiratory rate of >70 breaths / min is maintained and the duration is ≥10 seconds.

[0012] The early warning feedback module, electrically connected to the AI ​​analysis module, includes a local audible and visual alarm unit and a wireless communication unit. The local audible and visual alarm unit triggers within one second after the AI ​​analysis module detects a respiratory abnormality. Specifically, apnea corresponds to a flashing red light (2Hz) + an 80-85dB beep; periodic breathing corresponds to a flashing yellow light (1Hz) + a 70-75dB beep; and rapid breathing corresponds to a flashing orange light (1Hz) + a 70-75dB beep. The wireless communication unit pushes the abnormality type, start time, duration, and real-time blood oxygen data to medical staff's mobile terminals via the hospital's local area network (supporting the 802.11ax protocol), and automatically saves the original radar waveform data (in MAT file format) for the abnormal period (5 seconds before the abnormality ends to 10 seconds after the abnormality ends).

[0013] The data association module, which communicates with the early warning feedback module and the hospital's electronic medical record system (HIS / LIS system) via the HL7FHIR protocol, is used to record medical staff intervention measures (including stimulation methods, oxygen concentration, and medication type), intervention time, and abnormal relief time (defined as respiratory rate returning to normal and blood oxygen stabilizing for 5 minutes). This forms a closed loop of clinical data, encompassing "monitoring-identification-early warning-intervention-backtracking," and supports data retrieval by premature infant medical record number, abnormality type, and intervention method.

[0014] This invention employs a 60GHz±2GHz FMCW millimeter-wave radar sensor, embedded in the top of the incubator, maintaining a safe distance of 30-50cm from the premature infant's chest and abdomen. With an ultra-high spatial resolution of 0.1mm±0.02mm, it accurately captures minute respiratory displacements of 0.1-5mm in the chest and abdomen. The entire process requires no skin contact, fundamentally avoiding skin redness and swelling caused by the adhesive pressure of traditional contact electrodes, stratum corneum abrasion due to long-term contact, and skin damage caused by sensor friction. It perfectly suits the physiological characteristics of premature infants, whose stratum corneum thickness is only 1 / 3 to 1 / 2 that of adults, resolving the core contradiction between "skin protection" and "data acquisition" in long-term monitoring of premature infants.

[0015] Preferably, the signal gain parameter of the millimeter-wave radar sensor can be dynamically adjusted according to the weight of the premature infant: for premature infants weighing <1000g, the gain is set to 0.5-0.8dB; for premature infants weighing 1000-1500g, the gain is set to 0.9-1.2dB, so as to adapt to the differences in chest development of premature infants of different weights and avoid missing or oversaturating subtle chest movement signals.

[0016] Preferably, the multimodal feature model of the AI ​​analysis module is trained in the following way: respiratory movement data, blood oxygen data, and clinical respiratory abnormality annotation data of more than 150 preterm infants with a gestational age of 24-32 weeks are collected (the annotators are two or more neonatologists at the associate chief physician level), and a training dataset is constructed (80% of which is the training set, 10% is the validation set, and 10% is the test set); the initial bidirectional LSTM neural network is iteratively trained using the Adam optimizer, with 50-100 iterations and a learning rate of 0.001-0.003, until the model's sensitivity to the three types of respiratory abnormalities is ≥97.5%, specificity is ≥97%, and false recognition rate is ≤0.3%.

[0017] Preferably, the early warning feedback module further includes an abnormality level determination unit: if the duration of apnea is ≥5 seconds and the blood oxygen saturation drops by >10%, it is determined to be a level one abnormality, and an emergency alarm signal is pushed to the NICU nurse station control console simultaneously; if the apnea lasts for 2-5 seconds or the rapid breathing lasts for ≥30 seconds, it is determined to be a level two abnormality, and an alarm signal is pushed only to the mobile terminal of the medical staff responsible for the premature infant, so as to realize graded early warning and avoid waste of medical resources.

[0018] Preferably, the data association module has a built-in data encryption unit that uses the AES-256 encryption algorithm to encrypt and store the original radar waveform data, blood oxygen data and intervention records during abnormal periods. Only authorized medical staff can access and view the data through the hospital's electronic medical record system after verification through the hospital's unified identity authentication system (supporting fingerprint / facial recognition). At the same time, a data retrieval log is generated (recording the person retrieving the data, time and purpose), ensuring the privacy and security of premature infant medical data.

[0019] Preferably, the system also includes a calibration module, which is electrically connected to the monitoring module. The calibration module can periodically calibrate the millimeter-wave radar sensor using a standard respiratory simulation device (simulating a respiratory rate of 20-80 breaths / min and a chest movement amplitude of 0.5-5mm) at a rate of 7 days per calibration cycle to ensure the accuracy of respiratory motion data acquisition. At the same time, the near-infrared non-contact blood oxygen sensor can be calibrated using a standard blood oxygen simulator (simulating blood oxygen saturation of 80%-100%) at a rate of 14 days per calibration cycle.

[0020] Preferably, the AI ​​analysis module also integrates a signal quality assessment unit, which is used to evaluate the signal-to-noise ratio (SNR) of the respiratory motion signal collected by the millimeter-wave radar sensor in real time. If the SNR < 15dB, the gain adjustment is automatically triggered or a repositioning command is sent to the monitoring module to ensure that the signal quality meets the analysis requirements.

[0021] Preferably, the millimeter-wave radar sensor employs a waveform encoding technique based on chirped sequences. Its transmitted waveform is specifically encoded, which can effectively distinguish between the minute displacement of the chest and abdomen caused by breathing and the interference echoes from other moving objects (such as blankets and infusion tubes) in the incubator, thereby improving the signal-to-noise ratio at the signal source.

[0022] Preferably, the signal processing module further includes a motion artifact suppression unit, which employs an adaptive filtering algorithm to dynamically identify and filter out motion artifacts caused by non-respiratory body movements of premature infants (such as startle reflex and limb twitching), ensuring the extraction accuracy of pure respiratory waveforms.

[0023] An automatic method for identifying abnormal breathing in premature infants includes the following steps:

[0024] S1: Install the monitoring module and embed it into the pre-set mounting slot on the inner wall of the top of the incubator. Adjust the transmission power (-10dBm to -5dBm) and signal gain (0.5-1.2dB) of the millimeter-wave radar sensor according to the premature infant's gestational age and weight, and point the near-infrared non-contact blood oxygen sensor at the premature infant's fingertips or soles of feet. Start the monitoring module to control the millimeter-wave radar sensor (sampling rate 50Hz), the near-infrared non-contact blood oxygen sensor (sampling rate 1Hz), and the temperature and humidity sensor (sampling rate 0.5Hz) to simultaneously collect respiratory motion data, blood oxygen data, and environmental temperature and humidity data.

[0025] S2: The signal processing module receives monitoring data through the Type-C interface, uses the db4 wavelet base wavelet transform algorithm to perform 3-5 level decomposition filtering on the respiratory motion data, removes limb activity interference and temperature and humidity fluctuation interference with amplitude >5mm, extracts pure respiratory waveforms of 0.5-3Hz, and calculates respiratory frequency (updated once every 5 seconds) and chest movement amplitude (updated once every 1 second).

[0026] S3: The AI ​​analysis module receives respiratory rate and chest movement amplitude data output by the signal processing module via Ethernet, as well as blood oxygen data collected by the near-infrared non-contact blood oxygen sensor, and calls the multimodal feature model for real-time analysis.

[0027] If the respiratory displacement signal disappears for ≥2 seconds and the rate of decrease in blood oxygen is >1% / s and the cumulative decrease is >5%, it is determined to be apnea.

[0028] If the respiratory rate fluctuates by more than 20 breaths / min and the difference in chest cavity amplitude is ≥30% and the blood oxygen fluctuation is ≤2% within three consecutive respiratory cycles, it is judged as periodic breathing.

[0029] If a premature infant with a gestational age of <28 weeks has a respiratory rate of >75 breaths / min for ≥10 seconds, or a premature infant with a gestational age of 28-32 weeks has a respiratory rate of >70 breaths / min for ≥10 seconds, it is considered tachypnea.

[0030] S4: Based on the judgment result of the AI ​​analysis module, the early warning feedback module triggers the corresponding level of local audible and visual alarm within 1 second, pushes the abnormal information to the mobile terminals of medical staff through the hospital LAN, and saves the original radar waveform data from 5 seconds before the abnormality to 10 seconds after the abnormality ends in MAT file format; if it is judged to be a level 1 abnormality (breathing apnea ≥ 5 seconds and blood oxygen drop > 10%), an emergency alarm is pushed to the central control console of the nurse station at the same time.

[0031] S5: The data association module receives abnormal information from the early warning feedback module, records the intervention measures and intervention time of medical staff, and records the time of abnormal relief after the premature infant's respiratory rate returns to normal and blood oxygen stabilizes for 5 minutes. The abnormal data and intervention data are associated with the hospital's electronic medical record system through the HL7FHIR protocol to form a clinical data closed loop, and the original data is encrypted and stored using AES-256.

[0032] S6: Periodically start the calibration module, use a standard respiratory simulation device and a standard pulse oximeter to calibrate the accuracy of the millimeter-wave radar sensor and the near-infrared non-contact pulse oximeter, record the calibration data and generate a calibration report to ensure the long-term monitoring accuracy of the system.

[0033] Preferably, in step S3, if the multimodal feature model determines that the respiratory rate is consistently >60 breaths / min and < the respiratory dyspnea threshold corresponding to the gestational age (75 breaths / min for gestational age <28 weeks, 70 breaths / min for gestational age 28-32 weeks), and the blood oxygen saturation is stable (fluctuation ≤2%) and the chest movement amplitude does not decrease suddenly (change ≤10%), then it is determined as "suspected respiratory dyspnea". The warning feedback module only pushes text prompts (excluding sound and light alarms) to the mobile terminals of medical staff, and updates the respiratory rate and blood oxygen data every 30 seconds until the respiratory rate returns to normal or reaches the respiratory dyspnea judgment criteria.

[0034] Preferably, in step S5, the data association module also supports generating periodic analysis reports, which statistically analyze the frequency of respiratory abnormalities in preterm infants, the proportion of abnormality types, and the effectiveness of intervention measures (with an abnormality relief time of <5 minutes as the effective standard) on a weekly / monthly basis, providing data support for optimizing clinical treatment plans.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] 1. This invention employs a 60GHz±2GHz FMCW millimeter-wave radar sensor, embedded in the top of the incubator, maintaining a safe distance of 30-50cm from the premature infant's chest and abdomen. With an ultra-high spatial resolution of 0.1mm±0.02mm, it accurately captures minute respiratory displacements of 0.1-5mm in the chest and abdomen. The entire process requires no skin contact, fundamentally avoiding skin redness and swelling caused by the adhesive pressure of traditional contact electrodes, stratum corneum abrasion due to long-term contact, and skin damage caused by sensor friction. It perfectly adapts to the physiological characteristic of premature infants, whose stratum corneum thickness is only 1 / 3-1 / 2 that of adults, resolving the core contradiction between "skin protection" and "data acquisition" in long-term monitoring of premature infants.

[0037] 2. This invention also utilizes a near-infrared non-contact pulse oximeter sensor. While this sensor needs to be aligned with the fingertips or soles of the premature infant's feet, it eliminates the need for clips or adhesive patches like traditional pulse oximeter probes. Signal acquisition is achieved solely through precise positioning, significantly reducing skin constriction and contact area. Furthermore, the sensor sampling process involves no physical pressure, avoiding impaired local blood circulation caused by contact pressure and reducing the probability of skin damage from external stimuli. This design is well-suited to the physiological characteristics of premature infants, whose skin has extremely low tolerance to external stimuli.

[0038] 3. This invention also ensures that all components of the system that may come into close contact with the premature infant (such as the pulse oximeter sensor probe) are made of medical-grade, low-sensitivity materials. Furthermore, the millimeter-wave radar sensor and temperature / humidity sensor are completely non-contact, requiring no material contact with the skin. Compared to the adhesive layer commonly used in traditional electrode pads, this system avoids skin allergies caused by latex, adhesives, and other components through material selection, effectively reducing the occurrence of allergic symptoms such as itching and rashes, and lowering the health risk of skin barrier damage leading to infection caused by allergies. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the system framework of the present invention;

[0040] Figure 2 This is a flowchart of the method of the present invention;

[0041] Figure 3 This is a flowchart illustrating the workflow of the monitoring module in this invention.

[0042] Figure 4 This is a flowchart of the early warning feedback module in this invention. Detailed Implementation

[0043] To facilitate understanding of the technical solution of the present invention by those skilled in the art, the technical solution of the present invention will now be further described in conjunction with the accompanying drawings.

[0044] Example 1

[0045] like Figures 1 to 4 As shown, this invention provides an automatic identification system for abnormal neonatal breathing, including a monitoring module. The monitoring module is detachably embedded in a pre-set mounting slot on the inner wall of the top of an incubator. It includes an FMCW millimeter-wave radar sensor with a working frequency of 60GHz±2GHz and a spatial resolution of 0.1mm±0.02mm, a near-infrared non-contact blood oxygen sensor, and a temperature and humidity sensor with an accuracy of ±0.5℃ / ±3%RH. The detection surface of the millimeter-wave radar sensor is maintained at a vertical distance of 30-50cm from the chest and abdomen of the premature infant, and its signal transmission power is controlled between -10dBm and -5dBm to avoid electromagnetic radiation affecting the premature infant. It is used to collect minute displacement signals within a range of 0.1-5mm in the chest and abdomen of the premature infant to obtain respiratory motion data. The near-infrared non-contact blood oxygen sensor has detection wavelengths of 660nm±10nm and 940nm±10nm, and is used to collect blood oxygen saturation data from the fingertips or soles of the premature infant (sampling rate 1Hz). The temperature and humidity sensor is used to collect real-time temperature and humidity data within the incubator.

[0046] The FMCW millimeter-wave radar sensor operates at a frequency of 60GHz±2GHz and a spatial resolution of 0.1mm±0.02mm. At a vertical distance of 30-50cm from the chest and abdomen of premature infants, it acquires minute displacements of the chest and abdomen within a range of 0.1-5mm with a low transmit power of -10dBm to -5dBm, converting them into raw respiratory motion data. Simultaneously, it dynamically adjusts the gain according to the weight of premature infants (0.5-0.8dB for <1000g, 0.9-1.2dB for 1000-1500g) to adapt to the differences in chest development among premature infants of different weights, avoiding signal loss or oversaturation.

[0047] Near-infrared non-contact blood oxygen sensor: It uses dual wavelengths of 660nm±10nm and 940nm±10nm to detect blood oxygen saturation in the fingertips / plantar surfaces of premature infants at a sampling rate of 1Hz, ensuring real-time capture of blood oxygen changes.

[0048] Temperature and humidity sensor: Collects environmental data inside the insulation box with an accuracy of ±0.5℃ / ±3%RH, providing environmental parameter references for subsequent interference filtering.

[0049] The millimeter-wave radar sensor uses chirped sequence waveform encoding technology to distinguish between minute displacements caused by breathing and interference echoes from objects such as blankets and IV tubes inside the incubator, thereby improving the signal-to-noise ratio (SNR) of the respiratory signal from the source.

[0050] The signal processing module is electrically connected to the monitoring module via a Type-C interface. It has a built-in signal conditioning circuit and a filtering algorithm unit. The filtering algorithm unit uses a wavelet transform algorithm based on the db4 wavelet base to perform 3-5 layers of decomposition filtering on the respiratory motion data collected by the millimeter-wave radar sensor. This removes interference signals caused by unconscious limb movements of premature infants (amplitude > 5mm), airflow disturbances in the incubator, and temperature and humidity fluctuations, and extracts a pure respiratory waveform with a frequency range of 0.5-3Hz.

[0051] This module receives data from the monitoring module via a Type-C interface. Its core task is to remove interference and extract effective respiratory features. The working logic consists of two steps:

[0052] Basic Filtering: Wavelet Transform for Interference Removal: The built-in db4 wavelet basis wavelet transform algorithm performs 3-5 layers of decomposition filtering on the raw respiratory motion data, accurately filtering out three types of interference: ① unconscious limb movements of premature infants (amplitude > 5mm); ② airflow disturbance in the incubator; ③ temperature and humidity fluctuations, ultimately extracting a pure respiratory waveform with a frequency range of 0.5-3Hz.

[0053] Advanced optimization: Motion artifact suppression: The motion artifact suppression unit, which integrates an adaptive filtering algorithm, dynamically identifies and filters out artifacts caused by non-respiratory body movements such as startle reflex and limb twitching in premature infants, further ensuring the accuracy of respiratory waveform extraction and providing high-quality data input for subsequent AI analysis.

[0054] The AI ​​analysis module is connected to the signal processing module via Ethernet communication. It includes a multimodal feature model based on a bidirectional LSTM neural network. The input feature set of the multimodal feature model includes: respiratory rate extracted from the pure respiratory waveform (calculation accuracy ±1 breath / min), chest movement amplitude (calculation accuracy ±0.05mm), and blood oxygen saturation change rate (unit % / s) collected by a near-infrared non-contact blood oxygen sensor.

[0055] As the system's "decision-making core," the AI ​​analysis module receives processed signal data via Ethernet and uses a well-trained neural network model to identify respiratory anomalies while ensuring the reliability of signal analysis.

[0056] Input feature set: The core input features are the respiratory rate (accuracy ±1 breath / min) and chest movement amplitude (accuracy ±0.05mm) extracted from the pure respiratory waveform, combined with the blood oxygen saturation change rate (% / s) collected by the blood oxygen sensor.

[0057] The multimodal feature model identifies three types of respiratory abnormalities using the following rules:

[0058] Apnea: Disappearance of displacement signal in respiratory movement data for ≥2 seconds, with synchronous oxygen saturation decrease rate >1% / s and cumulative decrease >5%;

[0059] Periodic breathing: The difference in respiratory rate fluctuation within three consecutive respiratory cycles is >20 breaths / min, and the corresponding difference in chest wall movement amplitude is ≥30%, while there is no significant decrease in blood oxygen saturation (fluctuation ≤2%).

[0060] Rapid breathing: For preterm infants with a gestational age of <28 weeks, a respiratory rate of >75 breaths / min is maintained; for preterm infants with a gestational age of 28-32 weeks, a respiratory rate of >70 breaths / min is maintained and the duration is ≥10 seconds.

[0061] Suspected shortness of breath warning: If the respiratory rate is >60 breaths / min but does not reach the shortness of breath threshold, and the blood oxygen is stable (fluctuation ≤2%) and the chest cavity amplitude change is ≤10%, it is judged as "suspected shortness of breath" and a mild warning is triggered.

[0062] Training Basis: Based on respiratory and blood oxygenation data from over 150 preterm infants born at 24-32 weeks gestation, and annotations by associate chief physicians, the training / validation / test sets were divided into an 8:1:1 ratio. The Adam optimizer was used for iterative training 50-100 times (learning rate 0.001-0.003), ultimately achieving a recognition sensitivity ≥97.5%, specificity ≥97%, and false recognition rate ≤0.3%. Real-time Signal Quality Monitoring: An integrated signal quality assessment unit monitors the respiratory motion signal SNR in real time. If the SNR <15dB, gain adjustment is automatically triggered or a repositioning command is sent to the monitoring module to ensure the analyzed data meets quality standards.

[0063] The early warning feedback module is electrically connected to the AI ​​analysis module and includes a local audible and visual alarm unit and a wireless communication unit. The local audible and visual alarm unit is triggered within 1 second after the AI ​​analysis module detects a breathing abnormality. Specifically, apnea corresponds to a flashing red light (frequency 2Hz) + an 80-85dB buzzer, periodic breathing corresponds to a flashing yellow light (frequency 1Hz) + a 70-75dB buzzer, and rapid breathing corresponds to a flashing orange light (frequency 1Hz) + a 70-75dB buzzer. The wireless communication unit pushes the abnormality type, abnormality start time, duration, and real-time blood oxygen data to the mobile terminals of medical staff through the hospital's local area network (supporting the 802.11ax protocol), and automatically saves the original radar waveform data (stored in MAT file format) for the abnormal period (5 seconds before the abnormality ends to 10 seconds after the abnormality ends).

[0064] After the AI ​​module detects an anomaly, the early warning feedback module initiates a multi-dimensional response within one second, achieving a synergy of "local alarm + remote push + data storage":

[0065] Tiered audible and visual alarms: Different alarm forms are matched according to the severity of the abnormality: Apnea (including Level 1 abnormality: apnea ≥ 5 seconds and blood oxygen drop > 10%): flashing red light (2Hz) + 80-85dB buzzer sound; Periodic breathing / rapid breathing (including Level 2 abnormality: apnea 2-5 seconds or rapid breathing ≥ 30 seconds): flashing yellow / orange light (1Hz) + 70-75dB buzzer sound.

[0066] The wireless communication unit pushes anomaly type, start time, duration, and real-time blood oxygen data to medical staff's mobile terminals via the hospital's 802.11ax local area network. For Level 1 anomalies, an emergency alarm is simultaneously pushed to the NICU nurse station's central control console, ensuring priority response to critical situations while avoiding waste of medical resources. Data is saved instantly: the original radar waveform data for the anomaly period (5 seconds before the anomaly to 10 seconds after the anomaly) is automatically saved in MAT file format, providing raw data support for subsequent backtracking.

[0067] The data association module communicates with the early warning feedback module and the hospital's electronic medical record system (HIS / LIS system) via the HL7FHIR protocol. It is used to record medical staff intervention measures (including stimulation methods, oxygen concentration, and medication type), intervention time, and abnormal relief time (based on the criterion that the respiratory rate returns to the normal range and blood oxygen stabilizes for 5 minutes). This forms a clinical data closed loop of "monitoring-identification-early warning-intervention-backtracking" and supports data retrieval by premature infant medical record number, abnormality type, and intervention method.

[0068] This module connects the early warning module to the hospital's HIS / LIS system via the HL7FHIR protocol, constructing a closed loop of clinical data: "monitoring-identification-early warning-intervention-backtracking".

[0069] Data correlation and analysis: Records medical staff intervention measures (stimulation method, oxygen concentration, medication type), intervention time and abnormal relief time (normal respiratory rate and stable blood oxygen for 5 minutes as the standard), and supports data retrieval by medical record number, abnormal type and intervention method; generates statistical reports weekly / monthly to analyze the number of abnormal occurrences, type proportion and intervention effectiveness (relief time < 5 minutes is considered effective), providing a basis for optimizing treatment plans.

[0070] Privacy and security protection: Built-in AES-256 encryption unit encrypts and stores raw radar data, blood oxygen data and intervention records; data can only be accessed by authorized medical staff after authentication by hospital fingerprint / facial recognition, and access logs (person, time, purpose) are generated at the same time, which complies with medical data privacy regulations.

[0071] In embodiments of the present invention, the signal gain parameter of the millimeter-wave radar sensor can be dynamically adjusted according to the weight of the premature infant: for premature infants weighing <1000g, the gain is set to 0.5-0.8dB; for premature infants weighing 1000-1500g, the gain is set to 0.9-1.2dB, in order to adapt to the differences in chest development among premature infants of different weights and avoid missing or oversaturating minute chest movement signals.

[0072] The essence of millimeter-wave radar sensors is to emit high-frequency millimeter waves (typically 60GHz, 77GHz, etc.) and receive the echo signals reflected from the target (the chest of a premature infant), analyzing changes in the echoes to extract physiological motion information. This process relies heavily on the Doppler effect.

[0073] Signal transmission and propagation: The radar transmitter emits millimeter waves of a fixed frequency toward the chest region of the premature infant. The electromagnetic waves penetrate non-metallic obstructions such as clothing and thin blankets and reach the surface of the chest.

[0074] Echo generation and frequency shift: The rib cage produces tiny reciprocating movements (expansion / contraction) with breathing. The moving rib cage will produce a "Doppler frequency shift" on the incident millimeter waves - when the rib cage is closer to the radar, the echo frequency increases; when it is farther away, the echo frequency decreases.

[0075] Signal processing and information extraction: The radar receiver collects echo signals with frequency shift, and after amplification, filtering, demodulation and other processing, the frequency shift signal is converted into corresponding physiological parameters such as chest wall movement amplitude and frequency (i.e. respiratory rate), and finally outputs a signal that can be used for monitoring.

[0076] In an embodiment of the present invention, the multimodal feature model of the AI ​​analysis module is trained in the following manner: respiratory movement data, blood oxygen data, and clinical respiratory abnormality annotation data of more than 150 preterm infants with a gestational age of 24-32 weeks are collected (the annotators are two or more neonatologists at the associate chief physician level), and a training dataset is constructed (80% of which is the training set, 10% is the validation set, and 10% is the test set); the Adam optimizer is used to iteratively train the initial bidirectional LSTM neural network, with 50-100 iterations and a learning rate of 0.001-0.003, until the model's sensitivity to the three types of respiratory abnormalities is ≥97.5%, specificity is ≥97%, and false recognition rate is ≤0.3%.

[0077] Targeted collection of two core physiological data points from preterm infants with a gestational age of 24-32 weeks (covering two key dimensions: "respiratory movement" and "blood oxygenation," achieving complementary multimodal information):

[0078] Respiratory motion data: directly reflects the motion characteristics such as respiratory rate, rhythm, and amplitude (e.g., the motion signal will be significantly interrupted during apnea, and the signal frequency will be significantly increased during rapid breathing).

[0079] Blood oxygen data: indirectly reflects whether respiratory function is normal (for example, abnormal breathing may be accompanied by a decrease in blood oxygen saturation, which is an important auxiliary indicator of respiratory abnormalities).

[0080] The sample size should be ≥150 cases to ensure that the data covers premature infants of different weights and basic health conditions, and to avoid poor model generalization ability due to a single sample.

[0081] In an embodiment of the present invention, the early warning feedback module further includes an abnormality level determination unit: if the duration of apnea is ≥5 seconds and the blood oxygen saturation drops by >10%, it is determined to be a level one abnormality, and an emergency alarm signal is pushed to the NICU nurse station control console simultaneously; if the apnea lasts for 2-5 seconds or the rapid breathing lasts for ≥30 seconds, it is determined to be a level two abnormality, and an alarm signal is pushed only to the mobile terminal of the medical staff responsible for the premature infant, so as to realize graded early warning to avoid waste of medical resources.

[0082] The criteria for determining apnea lasting ≥5 seconds and a decrease in blood oxygen saturation >10% are based on the definition of "high-risk respiratory abnormalities" in the clinical practice guidelines for neonatal intensive care (NICU). Such abnormalities may lead to insufficient oxygen supply to the brain in premature infants, potentially causing short-term neurological damage or long-term developmental risks, thus requiring the highest level of emergency response. In actual monitoring, the system uses a high-precision pulse oximeter (sampling frequency ≥1Hz) to capture pulse oximetry fluctuations in real time. If, within the 5-second window of apnea, the pulse oximetry value drops by more than 10% from the baseline value (typically 92%-95%), a Level 1 abnormality will be immediately identified, preventing missed intervention opportunities due to delayed assessment.

[0083] Regarding the situation of "apnea lasting 2-5 seconds": This duration range is a common "brief apnea" in premature infants, mostly related to the immaturity of the nervous system. It usually does not directly cause severe hypoxia, but continuous monitoring is required to prevent the accumulation of abnormalities. The system records the respiratory rhythm through a chest impedance sensor. If the respiratory signal is interrupted for 2 seconds or more continuously, and the decrease in blood oxygen does not reach the level of a first-degree abnormality, it will be classified as a second-degree abnormality.

[0084] For cases of "rapid breathing lasting ≥30 seconds": Referring to the "Guidelines for the Diagnosis and Treatment of Neonatal Respiratory Distress Syndrome," a respiratory rate exceeding 60 breaths / minute for more than 30 seconds in premature infants may indicate early lung infection, respiratory compensation, or other problems. The system accumulates the duration of rapid breathing in 10-second increments. When three consecutive cycles (30 seconds in total) maintain a respiratory rate >60 breaths / minute, a level-two abnormality assessment is automatically triggered to avoid false alarms due to a single, brief instance of rapid breathing.

[0085] In an embodiment of the present invention, the data association module has a built-in data encryption unit that uses the AES-256 encryption algorithm to encrypt and store the original radar waveform data, blood oxygen data and intervention records of the abnormal period. Only authorized medical staff can access and view the data through the hospital's electronic medical record system after verification through the hospital's unified identity authentication system (supporting fingerprint / facial recognition). At the same time, a data retrieval log is generated (recording the retrieval personnel, time and purpose) to ensure the privacy and security of premature infant medical data.

[0086] Encryption covers all objects: In addition to the original radar waveform data (including core parameters such as waveform amplitude, frequency, and acquisition timestamp) during the abnormal period, blood oxygen data (including related information such as blood oxygen saturation, pulse rate, and measurement site), and intervention records (including intervention type, execution time, operating medical staff, and changes in vital signs after intervention), the encryption scope also extends to related metadata such as data acquisition device number and premature infant unique identification code (an anonymized identifier associated with the electronic medical record system), to avoid data traceability risks caused by metadata leakage.

[0087] The encryption process employs a two-stage protection mechanism: "transmission encryption + storage encryption." When data is transmitted from the acquisition device to the data association module, it is encrypted in real-time using the TLS 1.3 protocol to prevent data interception during transmission. Before data is written to storage media (such as a local hospital server or compliant cloud storage), a built-in encryption unit generates an independent data key based on the AES-256 algorithm to encrypt individual data blocks. The key itself is encrypted using an asymmetric encryption algorithm (RSA-2048) and stored on an independent key management server, avoiding security risks caused by storing the key and data simultaneously.

[0088] Storage media security hardening: The encrypted data is stored in storage devices that meet medical industry security standards. The devices have hardware-level encryption enabled (such as the TPM2.0 trusted platform module) and are configured with a regular data backup mechanism. The backup data is also encrypted with AES-256. The backup media is stored off-site and access permissions are strictly controlled to prevent data loss or leakage due to hardware failure or physical damage.

[0089] In embodiments of the present invention, a calibration module is also included. The calibration module is electrically connected to the monitoring module and can periodically calibrate the millimeter-wave radar sensor using a standard respiratory simulation device (simulating a respiratory rate of 20-80 breaths / min and a chest movement amplitude of 0.5-5mm) at a rate of 7 days per calibration to ensure the accuracy of respiratory motion data acquisition. At the same time, the near-infrared non-contact blood oxygen sensor can be calibrated using a standard blood oxygen simulator (simulating blood oxygen saturation of 80%-100%) at a rate of 14 days per calibration.

[0090] For millimeter-wave radar sensors used to collect respiratory motion data, the calibration module establishes a calibration closed loop by connecting to a standard respiratory simulation device. This simulation device can accurately reproduce common clinical respiratory parameters, with a respiratory rate covering 20-80 breaths / min. It can simulate both low-frequency breathing (20-30 breaths / min) in adults at rest and high-frequency breathing (60-80 breaths / min) after exercise or in special physiological states. The chest movement amplitude is set to 0.5-5mm to match the minute displacement changes of the chest cavity during breathing in different body types (such as infants and adults).

[0091] During calibration, the standard breathing simulation device first outputs a preset standard breathing parameter signal. The millimeter-wave radar sensor simultaneously acquires this simulated signal and transmits it to the calibration module. The module compares and analyzes the sensor's acquired data with the standard signal, calculates the deviation, and automatically generates calibration coefficients to correct the sensor's signal acquisition algorithm. The calibration cycle is set at 7 days, primarily based on the operating characteristics of the millimeter-wave radar sensor—during long-term use, factors such as changes in ambient temperature and the accumulation of electromagnetic interference may cause slight drift in its signal sensitivity. The 7-day cycle allows for timely correction of deviations while avoiding interference with normal monitoring caused by overly frequent calibrations.

[0092] For near-infrared non-contact pulse oximeter sensors, the calibration module uses a standard pulse oximeter simulator as the calibration benchmark. This simulator can simulate the range of human blood oxygen saturation from 80% to 100%, where 80%-93% corresponds to a mild hypoxic state and 94%-100% is a normal blood oxygen level, covering the core needs of daily health monitoring and early warning of potential hypoxic risks.

[0093] The calibration process is consistent with that of radar sensors: a standard pulse oximeter outputs a standard optical signal with a specific saturation level. The near-infrared sensor collects this data and uploads it to the calibration module. The module adjusts the sensor's optical signal analysis model by comparing the deviation between the standard value and the collected value, ensuring the accuracy of pulse oximetry calculations. The calibration cycle is set to 14 days per cycle because the core components of the near-infrared sensor (such as the light source and photodetector) have relatively higher stability and slower performance degradation during short-term use. A 14-day cycle reduces the system's calibration operation costs while maintaining accuracy.

[0094] In an embodiment of the present invention, the AI ​​analysis module also integrates a signal quality assessment unit, which is used to evaluate the signal-to-noise ratio (SNR) of the respiratory motion signal collected by the millimeter-wave radar sensor in real time. If the SNR is less than 15dB, the gain adjustment is automatically triggered or a repositioning command is sent to the monitoring module to ensure that the signal quality meets the analysis requirements.

[0095] When evaluating the signal-to-noise ratio (SNR) of respiratory motion signals acquired by millimeter-wave radar sensors, the signal quality assessment unit does not simply calculate the amplitude ratio of signal to noise. Instead, it employs a dynamic time-domain-frequency domain joint analysis algorithm. In the time domain, a sliding window (the window duration can be dynamically adjusted according to the respiratory rate, typically set to 2-5 seconds) is used to extract the periodic characteristics of the respiratory signal, filtering out transient interference caused by minor limb movements (such as slight finger movements or slight trunk swaying). In the frequency domain, a Fast Fourier Transform (FFT) is used to convert the signal to the frequency domain, locating the characteristic frequency range corresponding to the respiratory signal (0.1-0.3Hz for normal adult respiratory rate and 0.2-0.4Hz for children). The ratio of signal energy to noise energy across the entire frequency band within this range is calculated, ultimately yielding a precise SNR value.

[0096] To achieve "real-time assessment", the unit adopts a parallel computing architecture and is synchronized with the signal acquisition process of the millimeter-wave radar sensor. Every time the sensor completes the acquisition of a frame of signal (the frame interval is usually 50-100ms), the assessment unit immediately starts an SNR calculation. The delay of the whole process is controlled within 10ms, ensuring that fluctuations in signal quality can be captured in time and avoiding deviations in subsequent respiratory parameter analysis due to assessment lag.

[0097] Using SNR < 15dB as the threshold for triggering adjustment commands is based on the clinical accuracy requirements and engineering practice validation of millimeter-wave radar respiratory monitoring. From a clinical application perspective, the measurement errors of core parameters such as respiratory rate and respiratory depth need to be controlled within ±5% to meet the needs of scenarios such as health monitoring and sleep apnea screening. Through extensive experimental verification, when SNR ≥ 15dB, the respiratory features in the signal are clear, and the AI ​​analysis module's recognition error of respiratory rate can be stably controlled within ±3%. However, when SNR < 15dB, noise will mask some respiratory cycle features, causing the respiratory rate misjudgment rate to rise to over 15%, and the respiratory depth calculation deviation to even exceed 20%, failing to meet the analysis requirements.

[0098] In addition, this threshold also takes into account the stability of the system. If the threshold is set too high (such as 20dB), the system will frequently trigger adjustment commands, which will affect the continuity of monitoring. If it is set too low (such as 10dB), the signal quality cannot be effectively guaranteed. Therefore, 15dB is the optimal balance between "accuracy guarantee" and "system stability".

[0099] In an embodiment of the present invention, the millimeter-wave radar sensor employs waveform encoding technology based on chirped sequences. Its transmitted waveform is specifically encoded, which can effectively distinguish between the minute displacement of the chest and abdomen caused by breathing and the interference echoes of other moving objects (such as bedding and infusion tubes) in the incubator, thereby improving the signal-to-noise ratio at the signal source.

[0100] "Chirp sequence" is a commonly used modulation waveform technique in millimeter-wave radar. Its essence is to make the frequency of the radio frequency signal transmitted by the radar change linearly with time (similar to the "frequency change" of birdsong). "Specific coding" adds "sequence rule design" to the basic chirp waveform to form a unique signal "identity".

[0101] Technical features: Compared to traditional single-frequency or simple pulse waveforms, the encoded chirped sequence possesses two key attributes:

[0102] High distance resolution: By measuring the “slope” and “sequence length” of frequency changes, minute distance differences (down to the millimeter level) can be precisely measured, which is the basis for identifying “thoracic and abdominal respiratory displacement” (usually only a few millimeters).

[0103] Distinguishing features: Echoes from different moving targets carry different frequency shifts and phase changes. Encoded waveforms can amplify and record these differences like fingerprints. For example, chest and abdominal movements caused by breathing are periodic, low-frequency, and have regular displacement (usually 10-30 times / minute), while bedding shaking and IV tube swaying are non-periodic, high-frequency, and have random displacement. Encoded waveforms can quickly identify the "characteristic differences" between these two types of echoes through algorithms.

[0104] The core value of this millimeter-wave radar sensor lies in its use of "chirped sequence coding," a "transmitter optimization technology," to specifically address the core problems of "difficulty in identifying minute displacements and numerous interference signals" within incubators. This technology utilizes the high resolution of the coded waveform to capture respiratory movements while reducing interference through "signal source differentiation," ultimately achieving "non-contact, high-precision, and highly reliable" vital sign monitoring. It is particularly suitable for medical monitoring scenarios involving special populations such as newborns and critically ill patients.

[0105] In an embodiment of the present invention, the signal processing module further includes a motion artifact suppression unit. This unit employs an adaptive filtering algorithm that can dynamically identify and filter out motion artifacts caused by non-respiratory body movements of premature infants (such as startle reflex and limb twitching), ensuring the extraction accuracy of pure respiratory waveforms.

[0106] From a technical perspective, the adaptive filtering algorithm employed at the core of this unit has significant advantages. Unlike traditional fixed-parameter filtering algorithms, the adaptive filtering algorithm can analyze the characteristic changes of the input signal in real time and dynamically track and suppress interference signals by continuously adjusting the filtering parameters. Specifically, the algorithm first extracts features from the acquired mixed signal (containing real respiratory signals and motion artifacts), distinguishing between the regular waveform features of the respiratory signal (such as periodic fluctuations and specific frequency ranges) and the random features of motion artifacts (such as sudden signal fluctuations caused by startle reflexes and irregular signal superposition caused by limb twitching). Subsequently, a dynamic filtering model is constructed based on the feature differences between the two, accurately filtering out motion artifact components while preserving the detailed information of the real respiratory signal to the greatest extent, avoiding the respiratory signal distortion problems that may be caused by traditional filtering.

[0107] From the perspective of adapting to the physiological characteristics of premature infants, this unit has been specifically optimized. Because premature infants' nervous systems are not yet fully developed, their non-respiratory body movements are extremely frequent and diverse. Besides common startle reflexes and limb twitches, they may also exhibit trunk twisting, head turning, and other movements. The motion artifacts produced by different movements vary significantly in signal intensity and frequency range. To address this characteristic, the motion artifact suppression unit has pre-established a database of common motion artifact features in premature infants, covering motion signal samples from premature infants of different gestational ages and weights. The algorithm can combine the specific physiological parameters of the monitored subject (such as gestational age and weight) to call matching feature templates, further improving the accuracy of motion artifact recognition. For example, for premature infants with lower gestational age and lower weight, the artifact signal intensity produced by their startle reflex is usually weaker but the frequency is higher. The algorithm automatically adjusts the recognition threshold and filtering frequency band to avoid misjudging weak respiratory signals as artifacts, while accurately capturing and suppressing high-frequency artifacts.

[0108] Adaptive filter structure: It usually adopts the form of "linear combination", that is, the filter output y(n) is the inner product of the input signal x(n) and a set of variable weights wk(n) (k=0,1,...,N−1, N is the filter order): y(n)=∑k=0N−1wk(n)⋅x(n−k).

[0109] This invention provides an automatic identification method for abnormal breathing in premature infants according to Embodiment 1, comprising the following steps:

[0110] S1: Install the monitoring module and embed it into the pre-set mounting slot on the inner wall of the top of the incubator. Adjust the transmission power (-10dBm to -5dBm) and signal gain (0.5-1.2dB) of the millimeter-wave radar sensor according to the premature infant's gestational age and weight, and point the near-infrared non-contact blood oxygen sensor at the premature infant's fingertips or soles of feet. Start the monitoring module to control the millimeter-wave radar sensor (sampling rate 50Hz), the near-infrared non-contact blood oxygen sensor (sampling rate 1Hz), and the temperature and humidity sensor (sampling rate 0.5Hz) to simultaneously collect respiratory motion data, blood oxygen data, and environmental temperature and humidity data.

[0111] Before formally installing the monitoring module, three core preparatory tasks must be completed to ensure equipment compatibility and the safety of premature infants. First, the pre-set mounting slot on the inner wall of the incubator's top must be cleaned and inspected. Use sterile medical alcohol wipes to wipe away dust, residual adhesive, and other debris from the slot. Simultaneously, confirm that the dimensions of the mounting slot match the monitoring module's base (the error must be controlled within ±0.5mm) to avoid loosening or vibration of the module due to installation gaps, which could affect the stability of the radar signal. Second, specialized installation tools must be prepared, including a torque screwdriver (torque range 0.8-1.2 N·m) and an anti-static wrist strap (grounding resistance ≤1MΩ) to prevent electrostatic discharge from damaging the sensor circuitry during installation, or from compromising the module's seal due to overtightening / loosening of the screws. Finally, basic information about premature infants should be recorded in advance. In addition to gestational age and weight, special conditions such as skin edema and limb movement frequency should also be noted to provide a reference for subsequent sensor parameter adjustments. For example, for premature infants with frequent limb movements, the signal gain of the millimeter-wave radar sensor can be appropriately increased (it is recommended to increase it by 0.1-0.2dB) to reduce the interference of motion artifacts on respiratory data.

[0112] The parameter adjustments for millimeter-wave radar sensors and near-infrared pulse oximeter sensors must be strictly based on a grading standard established according to the gestational age and weight of premature infants to avoid inappropriate parameters leading to decreased monitoring accuracy or potential safety risks. The specific grading is as follows:

[0113] For extremely premature infants with a gestational age <28 weeks and a birth weight <1000g: the millimeter-wave radar transmission power should be adjusted to -9.5 to -10 dBm (reducing the power can reduce the potential impact on the fragile nervous system of premature infants), and the signal gain should be set to 0.5-0.7 dB (low gain can reduce the interference of environmental electromagnetic interference on respiratory data); the near-infrared blood oxygen sensor should be aimed at the sole of the foot first (the skin on the sole of the foot is thinner and the blood vessels are relatively denser, which can reduce signal penetration loss). If there is skin damage on the sole of the foot, the fingertip can be selected, but a layer of sterile breathable gauze (thickness ≤0.1mm) should be placed between the sensor and the skin to avoid affecting signal acquisition.

[0114] For premature infants with a gestational age of 28-32 weeks and a weight of 1000-1500g: the millimeter-wave radar transmission power should be adjusted to -7.5 to -9dBm, and the signal gain should be set to 0.8-1.0dB; the near-infrared blood oxygen sensor can be selected for the fingertips or the soles of the feet according to the premature infant's limb movements. If the premature infant has a habit of curling his fingers, it is recommended to select the soles of the feet first to avoid the sensor shifting due to finger movements.

[0115] For premature infants with a gestational age >32 weeks and a weight >1500g: the millimeter-wave radar transmission power should be adjusted to -5 to -7dBm (higher power can improve the sensitivity of respiratory data acquisition and is suitable for premature infants with relatively stable respiratory rhythms), and the signal gain should be set to 1.0-1.2dB; the near-infrared blood oxygen sensor can be flexibly selected for the fingertip or the sole of the foot, but it is necessary to ensure that the sensor probe is in close contact with the skin (contact ≥90% to avoid fluctuations in blood oxygen data caused by air gaps).

[0116] S2: The signal processing module receives monitoring data through the Type-C interface, uses the db4 wavelet base wavelet transform algorithm to perform 3-5 level decomposition filtering on the respiratory motion data, removes limb activity interference and temperature and humidity fluctuation interference with amplitude >5mm, extracts pure respiratory waveforms of 0.5-3Hz, and calculates respiratory frequency (updated once every 5 seconds) and chest movement amplitude (updated once every 1 second).

[0117] Transmission mode and rate: Monitoring data (such as raw respiratory motion signals) is usually real-time time-series data. The Type-C interface needs to work at USB 2.0 Full-Speed ​​(12Mbps) or higher, or adopt the "data + power" integrated mode of USB-CPD (PowerDelivery) protocol - which transmits monitoring data and powers the module at the same time (avoiding separate wiring, suitable for wearable / portable monitoring scenarios).

[0118] Data format adaptation: The interface must support the transmission of "raw analog signal sampled values" or "preprocessed digital signals", usually in conjunction with peripheral protocols such as I2C and SPI (if the module integrates an ADC sampling circuit), or directly transmit quantized 16-bit / 24-bit digital signals (to reduce the processing pressure on the host computer), and must match the output format of the monitoring equipment (such as chest strap sensors, millimeter-wave radar).

[0119] Anti-interference design: Type-C cables must have differential signal transmission characteristics to suppress electromagnetic interference (EMI), especially in complex electromagnetic environments such as medical and industrial fields. They must comply with USB-IF's EMC (electromagnetic compatibility) standards to avoid introducing noise into the interface transmission process and affecting subsequent filtering effects.

[0120] S3: The AI ​​analysis module receives respiratory rate and chest movement amplitude data output by the signal processing module via Ethernet, as well as blood oxygen data collected by the near-infrared non-contact blood oxygen sensor, and calls the multimodal feature model for real-time analysis.

[0121] If the respiratory displacement signal disappears for ≥2 seconds and the rate of decrease in blood oxygen is >1% / s and the cumulative decrease is >5%, it is determined to be apnea.

[0122] If the respiratory rate fluctuates by more than 20 breaths / min and the difference in chest cavity amplitude is ≥30% and the blood oxygen fluctuation is ≤2% within three consecutive respiratory cycles, it is judged as periodic breathing.

[0123] If a premature infant with a gestational age of <28 weeks has a respiratory rate of >75 breaths / min for ≥10 seconds, or a premature infant with a gestational age of 28-32 weeks has a respiratory rate of >70 breaths / min for ≥10 seconds, it is considered tachypnea.

[0124] Signal processing module data: Respiratory rate and chest movement amplitude data must meet the dual requirements of "real-time + continuity". Ethernet transmission latency must be ≤100ms to avoid judgment errors due to data lag; data sampling frequency must be no less than 20Hz to ensure that subtle changes in chest movement (such as amplitude fluctuations during shallow and rapid breathing in premature infants) can be captured. If the data packet loss rate is >5%, the system must automatically trigger the data completion algorithm and issue a "data quality warning" to avoid misjudgment based on incomplete data.

[0125] Near-infrared non-contact blood oxygen sensor data: Blood oxygen data must exclude environmental interference (such as strong light, skin obstruction), and the sensor sampling interval should be ≤0.5s to ensure accurate calculation of the "blood oxygen decline rate" (which needs to be obtained by dividing the difference between two consecutive sampling points by the time interval); when the blood oxygen value is in the range of 80%-100%, the measurement error should be ≤2%, and when it is below 80%, the error should be ≤3%. If the error exceeds the threshold, the judgment logic will temporarily block the blood oxygen index and only provide a preliminary warning based on the respiratory signal. The complete judgment will be restarted after the data returns to normal.

[0126] S4: Based on the judgment result of the AI ​​analysis module, the early warning feedback module triggers the corresponding level of local audible and visual alarm within 1 second, pushes the abnormal information to the mobile terminals of medical staff through the hospital LAN, and saves the original radar waveform data from 5 seconds before the abnormality to 10 seconds after the abnormality ends in MAT file format; if it is judged to be a level 1 abnormality (breathing apnea ≥ 5 seconds and blood oxygen drop > 10%), an emergency alarm is pushed to the central control console of the nurse station at the same time.

[0127] Alarm Level and Sound / Light Correspondence Rules: The module strictly follows the anomaly level output by the AI ​​analysis module (it is recommended to clearly define specific level standards such as Level 1, Level 2, and Level 3, for example, Level 1 is emergency, Level 2 is important, and Level 3 is alert), matching differentiated sound and light signals. Specifically, Level 1 anomaly corresponds to "red high-frequency sound and light" (light flashing at a frequency of 2Hz, alarm sound is a continuous buzzing, volume ≥85dB, ensuring that medical personnel within 3 meters can detect it immediately); Level 2 anomaly corresponds to "yellow mid-frequency sound and light" (light is constantly on, alarm sound is an intermittent buzzing at 1-second intervals, volume ≥75dB); Level 3 anomaly corresponds to "blue low-frequency sound and light" (light flashing slowly at a frequency of 0.5Hz, alarm sound is a single tone, volume ≥65dB), avoiding confusion between different alarm levels.

[0128] Alarm triggering and deactivation mechanism: The triggering time is strictly controlled within 1 second after the AI ​​judgment result is output. The entire process delay from signal reception, level matching to the activation of the audio-visual equipment does not exceed 500ms, with a 500ms redundancy reserved to cope with extreme network or equipment fluctuations. Alarm deactivation requires "double confirmation" conditions - medical staff click "confirm processing" on the mobile terminal or nurse station console, and the AI ​​analysis module determines that the current status has returned to normal (such as apnea is relieved and blood oxygen has returned to the normal range), to avoid misunderstandings that may lead to missed risks.

[0129] S5: The data association module receives abnormal information from the early warning feedback module, records the intervention measures and intervention time of medical staff, and records the time of abnormal relief after the premature infant's respiratory rate returns to normal and blood oxygen stabilizes for 5 minutes. The abnormal data and intervention data are associated with the hospital's electronic medical record system through the HL7FHIR protocol to form a clinical data closed loop, and the original data is encrypted and stored using AES-256.

[0130] When receiving abnormal information transmitted by the early warning feedback module, it does not only receive results such as "abnormal respiratory rate" or "low blood oxygen", but also simultaneously acquires multi-dimensional raw monitoring data, including but not limited to: real-time respiratory rate of premature infants (breaths / minute), blood oxygen saturation (%), abnormal start timestamp, abnormal duration, and warning level (such as "general warning" or "emergency warning"), to ensure the integrity of subsequent related data.

[0131] S6: Periodically start the calibration module, use a standard respiratory simulation device and a standard pulse oximeter to calibrate the accuracy of the millimeter-wave radar sensor and the near-infrared non-contact pulse oximeter, record the calibration data and generate a calibration report to ensure the long-term monitoring accuracy of the system.

[0132] Standard equipment verification: Confirm that the standard respiratory simulator (which must comply with the calibration standards for respiratory monitoring equipment such as ISO80601-2-60) and the standard pulse oximeter (which must comply with the calibration standards for pulse oximeter such as ISO80601-2-61 or ASTM F2149) are within their validity period (a metrological verification certificate must be provided, and the last verification time must not exceed 12 months). At the same time, check whether the power supply and signal output interface (such as RS485, USB) of the equipment are normal to avoid the calibration results being distorted due to the failure of the standard equipment.

[0133] Pre-processing of sensors to be calibrated: Start the system to be calibrated 30 minutes in advance to allow the millimeter-wave radar sensor (which must be in an unobstructed environment free from electromagnetic interference) and the near-infrared blood oxygen sensor (which must avoid direct strong light shining on the probe) to enter a stable working state; if there is dust or stains on the sensor surface, wipe it gently with a lint-free cloth dampened with medical alcohol to prevent contaminants from affecting signal acquisition.

[0134] In an embodiment of the present invention, in step S3, if the multimodal feature model determines that the respiratory rate is consistently >60 breaths / min and < the respiratory dyspnea threshold corresponding to the gestational age (75 breaths / min for gestational age <28 weeks, 70 breaths / min for gestational age 28-32 weeks), and the blood oxygen saturation is stable (fluctuation ≤2%) and the chest movement amplitude does not decrease suddenly (change ≤10%), then it is determined as "suspected respiratory dyspnea". The warning feedback module only pushes text prompt information (excluding sound and light alarms) to the mobile terminal of medical staff, and updates the respiratory rate and blood oxygen data once every 30 seconds until the respiratory rate returns to normal or reaches the respiratory dyspnea judgment standard.

[0135] The "breathing dyspnea threshold corresponding to gestational age" is clearly defined as the core dividing line between "suspected" and "confirmed" dyspnea. Specific segmentation standards must strictly adhere to the following rules and cannot be applied across segments: Gestational age < 28 weeks: Breathing dyspnea threshold is 75 breaths / min, meaning a "suspected" case requires 60 breaths / min < respiratory rate < 75 breaths / min; if the respiratory rate is ≥ 75 breaths / min, a "breathing dyspnea" case is directly triggered (not "suspected"). Gestational age 28-32 weeks (inclusive, exclusive): Breathing dyspnea threshold is 70 breaths / min, meaning a "suspected" case requires 60 breaths / min < respiratory rate < 70 breaths / min; if the respiratory rate is ≥ 70 breaths / min, a "breathing dyspnea" case is directly triggered (not "suspected"). Gestational age ≥ 32 weeks: The current rules do not clearly define a corresponding breathing dyspnea threshold. If a respiratory rate consistently > 60 breaths / min, supplementary judgment based on clinical guidelines is required, and this case is not currently included in the scope of this "suspected dyspnea" rule. The "continuous" time definition of "respiratory rate > 60 breaths / min" requires continuous monitoring for ≥ 10 seconds (i.e., at least 2 continuous monitoring cycles, with the default monitoring cycle being 5 seconds / time) to avoid misjudging as "suspected" due to instantaneous fluctuations (such as 1-2 high-frequency breaths caused by brief crying or changes in body position).

[0136] In an embodiment of the present invention, in step S5, the data association module also supports the generation of periodic analysis reports, which statistically analyze the number of respiratory abnormalities in premature infants, the proportion of abnormality types, and the effectiveness of intervention measures (with the abnormality relief time <5 minutes as the effective standard) on a weekly / monthly basis, providing data support for the optimization of clinical treatment plans.

[0137] The report's data primarily originates from multi-source monitoring equipment and medical records of premature infants in the Neonatal Intensive Care Unit (NICU). Specifically, data on the frequency and type of respiratory abnormalities in premature infants are collected in real-time by bedside monitors, including respiratory rate, blood oxygen saturation, and respiratory waveform. When these indicators exceed normal ranges (e.g., respiratory rate >60 breaths / min or <30 breaths / min, blood oxygen saturation <90%), the system automatically marks them as respiratory abnormalities and categorizes them based on their characteristics (e.g., rapid breathing, apnea, periodic breathing). Data on the effectiveness of interventions is linked to medical staff operation records, including intervention time (e.g., the time of initiation of oxygen therapy, time of using a resuscitation bag for pressurized oxygen delivery), intervention method, and the time it takes for vital signs to return to normal after intervention. The data is strictly judged according to the standard of "abnormality relief time <5 minutes as effective," ensuring the data is authentic, accurate, and clinically relevant.

[0138] The embodiments disclosed in this invention are preferred embodiments, but are not limited thereto. Those skilled in the art can easily understand the spirit of this invention based on the above embodiments and make different extensions and variations, but as long as they do not depart from the spirit of this invention, they are all within the protection scope of this invention.

Claims

1. An automatic neonatal respiratory abnormality identification system, applied to a closed incubator in the neonatal intensive care unit (NICU), characterized in that, include; The monitoring module is detachably embedded in a pre-set mounting slot on the top inner wall of the incubator. It includes an FMCW millimeter-wave radar sensor with a working frequency of 60GHz±2GHz and a spatial resolution of 0.1mm±0.02mm, a near-infrared non-contact blood oxygen sensor, and a temperature and humidity sensor with an accuracy of ±0.5℃ / ±3%RH. The millimeter-wave radar sensor maintains a vertical distance of 30-50cm between its detection surface and the premature infant's chest and abdomen, and its signal transmission power is controlled between -10dBm and -5dBm to avoid electromagnetic radiation affecting the premature infant. It is used to collect minute displacement signals within a 0.1-5mm range in the premature infant's chest and abdomen to obtain respiratory motion data. The near-infrared non-contact blood oxygen sensor has detection wavelengths of 660nm±10nm and 940nm±10nm, and is used to collect blood oxygen saturation data from the premature infant's fingertips or soles (sampling rate 1Hz). The temperature and humidity sensor is used to collect real-time temperature and humidity data within the incubator. The signal processing module is electrically connected to the monitoring module via a Type-C interface. It has a built-in signal conditioning circuit and a filtering algorithm unit. The filtering algorithm unit uses a wavelet transform algorithm based on the db4 wavelet base to perform 3-5 layers of decomposition filtering on the respiratory motion data collected by the millimeter-wave radar sensor. This removes interference signals caused by unconscious limb movements of premature infants (amplitude > 5 mm), airflow disturbances in the incubator, and temperature and humidity fluctuations, and extracts a pure respiratory waveform with a frequency range of 0.5-3 Hz. The AI ​​analysis module, connected to the signal processing module via Ethernet, includes a multimodal feature model built on a bidirectional LSTM neural network. The input feature set of the multimodal feature model includes: respiratory rate (calculation accuracy ±1 breaths / min) and chest movement amplitude extracted from a clean respiratory waveform (calculation accuracy ±0.05 mm), and the rate of change of blood oxygen saturation (unit % / s) collected by a near-infrared non-contact blood oxygen sensor. The multimodal feature model identifies three types of respiratory abnormalities using the following rules: Apnea: Disappearance of displacement signal in respiratory movement data for ≥2 seconds, with synchronous oxygen saturation decrease rate >1% / s and cumulative decrease >5%; Periodic breathing: The difference in respiratory rate fluctuation within three consecutive respiratory cycles is >20 breaths / min, and the corresponding difference in chest wall movement amplitude is ≥30%, while there is no significant decrease in blood oxygen saturation (fluctuation ≤2%). Rapid breathing: For preterm infants with a gestational age of <28 weeks, a respiratory rate of >75 breaths / min is maintained; for preterm infants with a gestational age of 28-32 weeks, a respiratory rate of >70 breaths / min is maintained and the duration is ≥10 seconds. The early warning feedback module, electrically connected to the AI ​​analysis module, includes a local audible and visual alarm unit and a wireless communication unit. The local audible and visual alarm unit triggers within one second after the AI ​​analysis module detects a respiratory abnormality. Specifically, apnea corresponds to a flashing red light (2Hz) + an 80-85dB beep; periodic breathing corresponds to a flashing yellow light (1Hz) + a 70-75dB beep; and rapid breathing corresponds to a flashing orange light (1Hz) + a 70-75dB beep. The wireless communication unit pushes the abnormality type, start time, duration, and real-time blood oxygen data to medical staff's mobile terminals via the hospital's local area network (supporting the 802.11ax protocol), and automatically saves the original radar waveform data (in MAT file format) for the abnormal period (5 seconds before the abnormality ends to 10 seconds after the abnormality ends). The data association module communicates with the early warning feedback module and the hospital electronic medical record system (HIS / LIS system) via the HL7FHIR protocol. It is used to record medical staff intervention measures (including stimulation methods, oxygen concentration, and medication type), intervention time, and abnormal relief time (based on the criterion that the respiratory rate returns to the normal range and blood oxygen stabilizes for 5 minutes). This forms a clinical data closed loop of "monitoring-identification-early warning-intervention-backtracking" and supports data retrieval by premature infant medical record number, abnormality type, and intervention method.

2. The automatic neonatal respiratory abnormality identification system according to claim 1, characterized in that, The signal gain parameter of the millimeter-wave radar sensor can be dynamically adjusted according to the weight of the premature infant: for premature infants weighing <1000g, the gain is set to 0.5-0.8dB; for premature infants weighing 1000-1500g, the gain is set to 0.9-1.2dB, in order to adapt to the differences in chest development of premature infants of different weights and avoid missing or oversaturating subtle chest movement signals.

3. The automatic neonatal respiratory abnormality identification system according to claim 1, characterized in that, The multimodal feature model of the AI ​​analysis module was trained as follows: respiratory movement data, blood oxygen data, and clinical respiratory abnormality annotation data were collected from more than 150 preterm infants with a gestational age of 24-32 weeks (annotated by more than two associate chief physicians-level neonatologists) to construct a training dataset (80% for training, 10% for validation, and 10% for testing); the initial bidirectional LSTM neural network was iteratively trained using the Adam optimizer for 50-100 iterations at a learning rate of 0.001-0.003 until the model achieved a sensitivity ≥97.5%, specificity ≥97%, and false recognition rate ≤0.3% for the three types of respiratory abnormalities.

4. The automatic neonatal respiratory abnormality identification system according to claim 1, characterized in that, The early warning feedback module also includes an abnormality level determination unit: if the duration of apnea is ≥5 seconds and the blood oxygen saturation drops by >10%, it is determined to be a level 1 abnormality, and an emergency alarm signal is pushed to the NICU nurse station control console simultaneously; if the apnea lasts for 2-5 seconds or the rapid breathing lasts for ≥30 seconds, it is determined to be a level 2 abnormality, and an alarm signal is only pushed to the mobile terminal of the medical staff responsible for the premature infant, so as to realize graded early warning and avoid waste of medical resources.

5. The automatic neonatal respiratory abnormality identification system according to claim 1, characterized in that, The data association module has a built-in data encryption unit that uses the AES-256 encryption algorithm to encrypt and store the original radar waveform data, blood oxygen data and intervention records during abnormal periods. Only authorized medical staff can access and view the data through the hospital's electronic medical record system after verification through the hospital's unified identity authentication system (supporting fingerprint / facial recognition). At the same time, a data retrieval log is generated (recording the person who retrieved the data, the time, and the purpose), ensuring the privacy and security of premature infants' medical data.

6. The automatic neonatal respiratory abnormality identification system according to claim 1, characterized in that, It also includes a calibration module, which is electrically connected to the monitoring module. The calibration module can periodically calibrate the millimeter-wave radar sensor using a standard respiratory simulation device (simulating a respiratory rate of 20-80 breaths / min and a chest movement amplitude of 0.5-5mm) at a rate of 7 days per calibration cycle to ensure the accuracy of respiratory motion data acquisition. At the same time, the near-infrared non-contact blood oxygen sensor can be calibrated using a standard blood oxygen simulator (simulating blood oxygen saturation of 80%-100%) at a rate of 14 days per calibration cycle.

7. The automatic neonatal respiratory abnormality identification system according to claim 1, characterized in that, The AI ​​analysis module also integrates a signal quality assessment unit, which is used to evaluate the signal-to-noise ratio (SNR) of the respiratory motion signal collected by the millimeter-wave radar sensor in real time. If the SNR is less than 15dB, the gain adjustment is automatically triggered or a repositioning command is sent to the monitoring module to ensure that the signal quality meets the analysis requirements.

8. The automatic neonatal respiratory abnormality identification system according to claim 1, characterized in that, The millimeter-wave radar sensor employs waveform encoding technology based on chirped sequences. Its transmitted waveform is specifically encoded, which can effectively distinguish between the minute displacement of the chest and abdomen caused by breathing and the interference echoes from other moving objects (such as blankets and infusion tubes) in the incubator, thereby improving the signal-to-noise ratio at the signal source.

9. The automatic neonatal respiratory abnormality identification system according to claim 7, characterized in that, The signal processing module also includes a motion artifact suppression unit, which uses an adaptive filtering algorithm to dynamically identify and filter out motion artifacts caused by non-respiratory body movements of premature infants (such as startle reflex and limb twitching), ensuring the extraction accuracy of pure respiratory waveforms. An automatic method for identifying abnormal breathing in premature infants, applied to the automatic identification system for abnormal breathing in premature infants as described in any one of claims 1-8, characterized by comprising the following steps: S1: Install the monitoring module and embed it into the pre-set mounting slot on the inner wall of the top of the incubator. Adjust the transmission power (-10dBm to -5dBm) and signal gain (0.5-1.2dB) of the millimeter-wave radar sensor according to the premature infant's gestational age and weight, and point the near-infrared non-contact blood oxygen sensor at the premature infant's fingertips or soles of feet. Start the monitoring module to control the millimeter-wave radar sensor (sampling rate 50Hz), the near-infrared non-contact blood oxygen sensor (sampling rate 1Hz), and the temperature and humidity sensor (sampling rate 0.5Hz) to simultaneously collect respiratory motion data, blood oxygen data, and environmental temperature and humidity data. S2: The signal processing module receives monitoring data through the Type-C interface, uses the db4 wavelet base wavelet transform algorithm to perform 3-5 level decomposition filtering on the respiratory motion data, removes limb activity interference and temperature and humidity fluctuation interference with amplitude >5mm, extracts pure respiratory waveforms of 0.5-3Hz, and calculates respiratory frequency (updated once every 5 seconds) and chest movement amplitude (updated once every 1 second). S3: The AI ​​analysis module receives respiratory rate and chest movement amplitude data output by the signal processing module via Ethernet, as well as blood oxygen data collected by the near-infrared non-contact blood oxygen sensor, and calls the multimodal feature model for real-time analysis. If the respiratory displacement signal disappears for ≥2 seconds and the rate of decrease in blood oxygen is >1% / s and the cumulative decrease is >5%, it is determined to be apnea. If the respiratory rate fluctuates by more than 20 breaths / min and the difference in chest cavity amplitude is ≥30% and the blood oxygen fluctuation is ≤2% within three consecutive respiratory cycles, it is judged as periodic breathing. If a premature infant with a gestational age of <28 weeks has a respiratory rate of >75 breaths / min for ≥10 seconds, or a premature infant with a gestational age of 28-32 weeks has a respiratory rate of >70 breaths / min for ≥10 seconds, it is considered tachypnea. S4: Based on the judgment result of the AI ​​analysis module, the early warning feedback module triggers the corresponding level of local audible and visual alarm within 1 second, pushes the abnormal information to the mobile terminals of medical staff through the hospital LAN, and saves the original radar waveform data from 5 seconds before the abnormality to 10 seconds after the abnormality ends in MAT file format; if it is judged to be a level 1 abnormality (breathing apnea ≥ 5 seconds and blood oxygen drop > 10%), an emergency alarm is pushed to the central control console of the nurse station at the same time. S5: The data association module receives abnormal information from the early warning feedback module, records the intervention measures and intervention time of medical staff, and records the time of abnormal relief after the premature infant's respiratory rate returns to normal and blood oxygen stabilizes for 5 minutes. The abnormal data and intervention data are associated with the hospital's electronic medical record system through the HL7FHIR protocol to form a clinical data closed loop, and the original data is encrypted and stored using AES-256. S6: Periodically start the calibration module, use a standard respiratory simulation device and a standard pulse oximeter to calibrate the accuracy of the millimeter-wave radar sensor and the near-infrared non-contact pulse oximeter, record the calibration data and generate a calibration report to ensure the long-term monitoring accuracy of the system.

10. The automatic neonatal respiratory abnormality identification system according to claim 9, characterized in that, In step S3, if the multimodal feature model determines that the respiratory rate is consistently >60 breaths / min and < the respiratory dyspnea threshold for the corresponding gestational age (75 breaths / min for gestational age <28 weeks, 70 breaths / min for gestational age 28-32 weeks), and the blood oxygen saturation is stable (fluctuation ≤2%) and the chest movement amplitude does not decrease suddenly (change ≤10%), then it is determined as "suspected dyspnea". The warning feedback module only pushes text prompts (excluding sound and light alarms) to the mobile terminals of medical staff, and updates the respiratory rate and blood oxygen data every 30 seconds until the respiratory rate returns to normal or reaches the dyspnea judgment criteria.

11. The automatic neonatal respiratory abnormality identification system according to claim 9, characterized in that, In step S5, the data association module also supports the generation of periodic analysis reports, which can statistically analyze the number of respiratory abnormalities in premature infants, the proportion of abnormality types, and the effectiveness of intervention measures (with the abnormality relief time <5 minutes as the effective standard) on a weekly / monthly basis, providing data support for the optimization of clinical treatment plans.

Images