Exhaled nitric oxide detection method, system and device based on multi-source data fusion

Abstract

The application discloses a kind of based on multi-source data fusion expiratory nitric oxide detection method, system and equipment.Belong to medical detection technical field.It includes: using phased multitask learning framework, utilize standard gas data training model to master the basic nonlinear response of sensor and environmental cross sensitivity;Then, the clinical expiratory real sample data are fused, and the model is consolidated physical law while learning the ability to suppress complex expiratory matrix effect and individual difference interference by multitask learning.Training completed lightweight artificial intelligence correction model is deployed in embedded microprocessor, and can carry out integrated intelligent correction to real-time acquisition sensor signal, environmental temperature and humidity and expiratory flow.The application breaks through the performance limit of commercial sensor, realizes the sub-ppb level ultra-sensitive detection of expiratory nitric oxide and high clinical consistency output, and has strong environmental robustness, provides low-cost, high-precision and reliable solution for primary medical care and family health monitoring.

Description

Technical Field

[0001] This invention relates to the field of medical testing devices and sensing technology, and in particular to a method, system and device for detecting exhaled nitric oxide based on multi-source data fusion. Background Technology

[0002] Exhaled nitric oxide (FeNO) is an important biomarker for airway inflammatory diseases. Currently, the gold standard method for high-precision FeNO detection (such as chemiluminescence) involves expensive and bulky equipment, making it difficult to popularize. Although low-cost commercial electrochemical sensors are portable and economical, they face three major technical bottlenecks in practical applications: (1) limited detection sensitivity, making it difficult to meet the needs of accurate measurement at low concentrations; (2) nonlinearity and drift in the response signal; and (3) susceptibility to interference from changes in environmental temperature and humidity and the matrix effect of complex components in human exhaled air, resulting in poor consistency between clinical measurement results and the gold standard.

[0003] Existing calibration and improvement schemes are mostly limited to using standard gases in ideal environments, or simply addressing physical interferences by guiding a fixed expiratory flow rate. These methods cannot effectively compensate for the complex chemical matrix effects caused by individual differences in expiratory breath composition. Therefore, there is an urgent need in this field for an innovative technology to overcome these shortcomings and enable low-cost sensors to achieve accurate and stable detection in real clinical settings. Summary of the Invention

[0004] Therefore, the purpose of this invention is to provide a method, system, and device for detecting exhaled nitric oxide based on multi-source data fusion to solve the technical problems of insufficient detection sensitivity and susceptibility to interference from the environment and biological matrix of the aforementioned low-cost sensors.

[0005] To achieve the above objectives, the present invention provides a method for detecting exhaled nitric oxide based on multi-source data fusion, comprising the following steps: S1. Acquire the raw electrical signals collected by the nitric oxide sensor from exhaled gas, acquire ambient temperature and humidity parameters, and acquire expiratory flow rate; construct a fusion training dataset containing standard gas datasets and clinical exhalation datasets. S2. An AI calibration model is trained using a phased training strategy. In the first phase, the AI ​​calibration model learns the basic response of the sensor and environmental sensitivity. In the second phase, it learns to identify and extract nitric oxide-specific concentration information from mixed signals in the presence of complex exhaled gas matrix. A composite loss function is used to optimize the learning loss of the two phases, resulting in the trained AI calibration model. S3. The trained artificial intelligence correction model is lightweighted and deployed in an embedded device to perform integrated correction of real-time sensor signals, environmental parameters and expiratory flow, and directly output a high-precision expiratory nitric oxide concentration value.

[0006] Furthermore, in S1, the construction of the fused training dataset comprising the standard gas dataset and the clinical exhalation dataset includes: Within a temperature- and humidity-controlled environment, standard nitric oxide gases of different known concentrations are introduced into a nitric oxide sensor. Simultaneously, the sensor's original electrical signal, ambient temperature, and ambient humidity are collected to form a first-type sample dataset with the standard gas concentration as the true value. A prototype device containing the nitric oxide sensor and a chemiluminescence gold standard device are used to simultaneously collect breath samples from the same subject. The prototype device collects the sensor's original electrical signal, ambient temperature, ambient humidity, and expiratory flow rate parameters. Using the concentration measured by the gold standard device as the reference true value, a second-type sample dataset is formed, serving as the clinical breath dataset.

[0007] Furthermore, in S2, the composite loss function is a weighted sum of the first-stage learning loss and the second-stage learning loss, L_total = L_task1 + λ * L_task2, where λ is a hyperparameter for adjusting the learning weights of the two tasks.

[0008] Furthermore, the lightweight processing described in S3 includes structured pruning and low-bit integer quantization of the model, resulting in a final deployed model size of less than 200KB and a single inference time of less than 150 milliseconds on the embedded microprocessor.

[0009] Furthermore, during the creation of the standard gas dataset, in a temperature- and humidity-controlled environment, standard nitric oxide gas of different known concentrations is introduced into the nitric oxide sensor. In the low concentration range of 0-10 ppb, concentration points are set at intervals no greater than 1 ppb; in the range above 10 ppb, points are set at wider intervals. At each concentration point, the high-frequency raw electrical signal waveform of the sensor, ambient temperature, and humidity are recorded simultaneously, with the known standard gas concentration as the true value. Multiple repeated measurements are performed at each point to obtain statistical characteristics.

[0010] Furthermore, it also includes comparing the output value with the gold standard reference value, and evaluating the overall accuracy by calculating the root mean square error, mean absolute error, and coefficient of determination.

[0011] This invention also provides a breath nitric oxide detection system based on multi-source data fusion, used to perform the steps of the above-described breath nitric oxide detection method based on multi-source data fusion, comprising: a nitric oxide sensor module for acquiring raw electrical signals; an environmental sensor module for monitoring environmental temperature and humidity; a flow sensor module for monitoring exhaled breath flow; a microprocessor internally storing and running the pre-trained lightweight artificial intelligence correction model; and an output device for outputting nitric oxide concentration values.

[0012] Furthermore, the lightweight AI correction model is stored in the non-volatile memory of the microprocessor.

[0013] The present invention also provides a breath nitric oxide detection device, the device comprising the system described above.

[0014] The exhaled nitric oxide detection method, system, and device disclosed in this application, based on multi-source data fusion, have at least the following advantages compared to existing technologies: Detection sensitivity is improved by orders of magnitude: By enhancing the analysis of weak signal features and suppressing noise through AI models, the effective detection limit of the system is significantly improved to the sub-ppb level, breaking through the original performance limit of commercial sensors.

[0015] High consistency with clinical measurements: By learning from a large number of clinical samples, the model inherently eliminates the interference of individual physiological differences and complex expiratory matrix, making its test results in the real population highly consistent with the gold standard method.

[0016] Strong environmental and biological robustness: The integrated calibration scheme can simultaneously respond to fluctuations in environmental parameters and variations in biological samples, ensuring the long-term stability and reliability of the equipment under different usage scenarios.

[0017] It boasts excellent cost-effectiveness: with almost no increase in hardware costs, the low-cost sensor system achieves detection performance close to that of high-end chemiluminescence analyzers through algorithm-enabled technology, making it highly valuable for widespread application. Attached Figure Description

[0018] Figure 1 This is a block diagram of the hardware structure of the detection system according to an embodiment of the present invention.

[0019] Figure 2 This is a flowchart of the training method for the artificial intelligence correction model of this invention.

[0020] Figure 3 This is a flowchart of the real-time detection method of the present invention. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] like Figure 1 As shown, this invention includes core components such as a nitric oxide sensor, a temperature and humidity sensor, a flow sensor, a microprocessor (including an AI calibration model), and an output device. Specifically, one embodiment of this invention provides a breath nitric oxide detection device based on multi-source data fusion, with a microcontroller (MCU) as the core processing unit. It connects to and controls the following modules: The sensing module includes an electrochemical nitric oxide sensor, a temperature and humidity composite sensor, and a differential pressure flow / pressure sensor (for monitoring expiratory flow and oral pressure).

[0023] Human-computer interaction and output module: includes a display screen for showing concentration results and operation guidance.

[0024] Auxiliary modules: necessary power management circuits and communication interfaces (such as USB, Bluetooth).

[0025] The microcontroller's non-volatile memory stores a lightweight artificial intelligence correction model and driver firmware. The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0026] This invention also provides a breath nitric oxide detection system based on multi-source data fusion, used to perform the steps of the above-described breath nitric oxide detection method based on multi-source data fusion, comprising: a nitric oxide sensor module for acquiring raw electrical signals; an environmental sensor module for monitoring environmental temperature and humidity; a flow sensor module for monitoring exhaled breath flow; a microprocessor internally storing and running the pre-trained lightweight artificial intelligence correction model; and an output device for outputting nitric oxide concentration values.

[0027] Furthermore, the lightweight AI correction model is stored in the non-volatile memory of the microprocessor.

[0028] like Figure 2 As shown, the training process specifically includes two stages: basic model pre-training and multi-task joint fine-tuning. That is, the present invention also provides a method for detecting exhaled nitric oxide based on multi-source data fusion, including the following steps: S1. Acquire the raw electrical signal collected by the nitric oxide sensor from the exhaled gas, acquire the ambient temperature and humidity parameters, and acquire the expiratory flow rate; construct a fusion training dataset containing the standard gas dataset and the clinical expiratory breath dataset; Furthermore, in S1, the construction of the fused training dataset comprising the standard gas dataset and the clinical exhalation dataset includes: Within a temperature- and humidity-controlled environment, standard nitric oxide gases of different known concentrations are introduced into a nitric oxide sensor. Simultaneously, the sensor's original electrical signal, ambient temperature, and ambient humidity are collected to form a first-type sample dataset with the standard gas concentration as the true value. A prototype device containing the nitric oxide sensor and a chemiluminescence gold standard device are used to simultaneously collect breath samples from the same subject. The prototype device collects the sensor's original electrical signal, ambient temperature, ambient humidity, and expiratory flow rate parameters. Using the concentration measured by the gold standard device as the reference true value, a second-type sample dataset is formed, serving as the clinical breath dataset.

[0029] Furthermore, during the creation of the standard gas dataset, in a temperature- and humidity-controlled environment, standard nitric oxide gas of different known concentrations is introduced into the nitric oxide sensor. In the low concentration range of 0-10 ppb, concentration points are set at intervals no greater than 1 ppb; in the range above 10 ppb, points are set at wider intervals. At each concentration point, the high-frequency raw electrical signal waveform of the sensor, ambient temperature, and humidity are recorded simultaneously, with the known standard gas concentration as the true value. Multiple repeated measurements are performed at each point to obtain statistical characteristics.

[0030] Sample Collection: Recruitment included healthy volunteers and patients with respiratory diseases exhibiting typical variations in exhaled nitric oxide levels (e.g., bronchial asthma patients, ranging from mild to severe). Exhaled breath samples were collected simultaneously using a prototype device and a chemiluminescence gold standard device via a three-way valve. Prototype Device Recording: Raw electrical signal waveforms, ambient temperature and humidity, expiratory flow-time curves, and pressure-time curves. Gold Standard Device provided a true reference value for FeNO concentration.

[0031] Standard procedure and quality control: Subjects are required to exhale steadily at a constant target flow rate (50 mL / s) for at least 10 seconds. Flow curves are analyzed offline, and only samples with a stable flow rate within ±10% of the target flow rate and a stable plateau period lasting ≥6 seconds are retained as qualified samples.

[0032] From each qualified sample (including standard gas and clinical exhalation), a uniform feature vector is extracted, including: Sensor response characteristics: extracted from steady-state plateau signals, such as mean current, rise / fall slope, and signal integral. Interaction terms between temperature / humidity and current can be constructed to enhance model capabilities.

[0033] Synchronized environmental parameters: temperature (T) and relative humidity (RH).

[0034] Expiratory dynamics characteristics: mean expiratory flow rate during plateau phase (essential characteristic), and mean oral pressure during plateau phase (optional).

[0035] S2. The artificial intelligence correction model is trained using a phased training strategy. In the first phase, the artificial intelligence correction model learns the basic response of the sensor and the sensitivity of the environment. In the second phase, it learns to identify and extract the specific concentration information of nitric oxide from the mixed signal in the presence of a complex exhalation matrix. The learning loss of the two phases is optimized using a composite loss function to obtain the trained artificial intelligence correction model.

[0036] Phase 1: Modeling fundamental physical laws (pre-training) Using only a standard gas dataset, a lightweight neural network (e.g., a fully connected network with 2-3 hidden layers) is trained to learn the sensor's fundamental nonlinear response and its cross-sensitivity to ambient temperature and humidity. After this stage, the model becomes a "physical law simulator."

[0037] Phase Two: Integration of Clinical Anti-interference Capabilities (Joint Fine-tuning) This stage employs a multi-task learning framework, aiming to learn clinical resilience while retaining the physical laws.

[0038] Task Setup: Task 1 (Maintaining Physical Laws): Continue using standard gas data, and calculate the error between the predicted value and the standard concentration using a loss function. Task 2 (Clinical Anti-interference Learning): Introduce clinical expiratory breath data, and calculate the error between the predicted value and the gold standard reference value using a loss function.

[0039] Training Mechanism: The model trained in the first stage is used as initialization, and its underlying parameters can be partially frozen. During training, two types of data are input into the model alternately or in batches. The model is optimized through a composite loss function (e.g., L_total = L_task1 + λ* L_task2), automatically learning to integrate two types of knowledge: that is, based on accurate physical inversion, it learns to separate biological matrix interference from clinical mixed signals.

[0040] Final output: This framework yields a unified AI correction model that inherently integrates three key capabilities: nonlinear correction, environmental compensation, and matrix effect suppression. Furthermore, in S2, the composite loss function is a weighted sum of the first-stage learning loss and the second-stage learning loss, L_total = L_task1 + λ * L_task2, where λ is a hyperparameter for adjusting the learning weights of the two tasks.

[0041] In a preferred embodiment, the degree of matrix interference is defined as a quantification target, thereby guiding the model's internal representation to actively separate interfering components unrelated to the nitric oxide signal. For a clinical breath sample, the original sensor signal can be conceptually decomposed as: S_clinical = S_NO + S_matrix + Noise, where S_NO is the pure nitric oxide signal and S_matrix is ​​the matrix interference signal.

[0042] The ambient temperature and humidity (T, RH) and expiratory flow rate measured from clinical samples were input into the "basic physical model" that had completed the first phase of pre-training (based on standard gases only).

[0043] This fundamental physical model, based on patterns learned from pure standard gases, predicts the magnitude of the signal that should be generated under the current environmental conditions (T, RH) and flow rate (Flow) if only pure nitric oxide is available. This predicted signal is denoted as S_pred_pure.

[0044] Calculate and normalize the signal residual: Residual = |S_clinical - S_pred_pure|. This residual includes the matrix interference signal and the error caused by the underlying model's failure to fit perfectly.

[0045] The normalized residual value is used as a surrogate label (pseudo-label) for the "matrix interference level" of this clinical sample. Assuming the basic model is well trained, the larger the residual, the greater the deviation of the current signal from the pure NO response pattern, i.e., the stronger the matrix interference.

[0046] In the second stage of the multi-task learning network, a parallel, lightweight output branch (head) is added, specifically for "matrix interference estimation".

[0047] The main task branch outputs the corrected FeNO concentration (core task). The auxiliary task branch outputs a scalar, namely the estimated "matrix interference level". Shared feature extraction layers: Both branches share most of the network layers at the bottom, which are responsible for extracting high-dimensional features from the original inputs (signal, T, RH, flow, etc.). This forces the shared layers to learn features that can simultaneously explain both concentration and interference.

[0048] S3. The trained artificial intelligence correction model is lightweighted and deployed in an embedded device to perform integrated correction of real-time sensor signals, environmental parameters and expiratory flow, and directly output a high-precision expiratory nitric oxide concentration value.

[0049] To enable the AI ​​correction model to perform real-time inference on resource-constrained embedded microprocessors, it needs to be lightweighted and converted to a new format, mainly including the following steps: Model compression: The trained model undergoes structured pruning and quantization. For example, a weight-based pruning strategy can be used to remove redundant connections, and the network can be quantized with low-bit integers (e.g., converting weights and activation values ​​from 32-bit floating-point numbers to 8-bit integers) to significantly reduce model size and computational cost.

[0050] Format conversion and integration: Convert the compressed model into a format supported by the target embedded processor or inference framework. For example, a dedicated toolchain such as TensorFlow Lite for Microcontrollers can be used to convert the model into C language source code and integrate it into the device firmware.

[0051] Implementation Results: After the lightweight process described above, the AI ​​correction model can run efficiently on a typical low-cost microcontroller (e.g., an MCU based on an ARM Cortex-M series core). In a preferred embodiment, the final deployed model size can be compressed to less than 150KB, and the single forward inference time can be less than 100 milliseconds, thereby meeting the real-time and cost requirements of breath detection devices.

[0052] As attached Figure 3 The steps are shown to simultaneously collect parameters such as raw electrical signals, ambient temperature, ambient humidity, and expiratory flow rate, and input them into the AI ​​model for real-time correction.

[0053] The device operates as follows: Signal acquisition: Guide the user to exhale, and the microprocessor reads the original sensor signals, ambient temperature and humidity, and exhalation flow rate in real time.

[0054] Feature extraction and inference: Once the system determines that it has entered the stable plateau phase of exhalation, it immediately extracts the input feature vector that is completely isomorphic to the training phase and inputs it into the embedded AI correction model.

[0055] Output results: The model performs real-time inference, directly outputting the FeNO concentration value after comprehensive correction, and displaying it on the screen. The entire process is completed within hundreds of milliseconds, achieving real-time detection.

[0056] Furthermore, the lightweight processing described in S3 includes structured pruning and low-bit integer quantization of the model, resulting in a final deployed model size of less than 200KB and a single inference time of less than 150 milliseconds on the embedded microprocessor.

[0057] Furthermore, it also includes comparing the output value with the gold standard reference value, and evaluating the overall accuracy by calculating the root mean square error, mean absolute error, and coefficient of determination.

[0058] The aforementioned data fusion and phased training framework are expected to deliver the following significant technical benefits: 1) Breakthrough in detection sensitivity: By suppressing noise and matrix interference, the effective detection limit of the system is expected to be improved to the sub-ppb level; 2) High clinical consistency: The detection results are expected to show a high correlation coefficient (R²) with the chemiluminescence gold standard on a broad range of clinical samples. 2) It will be better than 0.95; 3) Strong robustness: It is expected that there will be no significant difference in performance in subgroup analyses of different ages, genders and physiological states.

[0059] To verify the above effects, the following evaluation method can be used: An independent test set is constructed, containing a series of low-concentration standard gases and clinical breath samples covering different populations. The output values ​​of this system are compared with the gold standard reference values, and the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) are calculated. 2 The overall accuracy was evaluated using subgroup analysis; the model's generalization ability was verified using subgroup analysis; and environmental robustness was verified by testing standard gases under different temperature and humidity conditions.

[0060] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for detecting exhaled nitric oxide based on multi-source data fusion, characterized in that, Includes the following steps: S1. Acquire the raw electrical signals collected by the nitric oxide sensor from exhaled gas, acquire ambient temperature and humidity parameters, and acquire expiratory flow rate; construct a fusion training dataset containing standard gas datasets and clinical exhalation datasets. S2. An AI calibration model is trained using a phased training strategy. In the first phase, the AI ​​calibration model learns the basic response of the sensor and environmental sensitivity. In the second phase, it learns to identify and extract nitric oxide-specific concentration information from mixed signals in the presence of complex exhaled gas matrix. A composite loss function is used to optimize the learning loss of the two phases, resulting in the trained AI calibration model. S3. The trained artificial intelligence correction model is lightweighted and deployed in an embedded device to perform integrated correction of real-time sensor signals, environmental parameters and expiratory flow, and directly output a high-precision expiratory nitric oxide concentration value.

2. The method for detecting exhaled nitric oxide based on multi-source data fusion according to claim 1, characterized in that, In S1, the construction of the fused training dataset containing the standard gas dataset and the clinical exhalation dataset includes: Within a temperature- and humidity-controlled environment, standard nitric oxide gases of different known concentrations are introduced into a nitric oxide sensor. Simultaneously, the sensor's original electrical signal, ambient temperature, and ambient humidity are collected to form a first-type sample dataset with the standard gas concentration as the true value. A prototype device containing the nitric oxide sensor and a chemiluminescence gold standard device are used to simultaneously collect breath samples from the same subject. The prototype device collects the sensor's original electrical signal, ambient temperature, ambient humidity, and expiratory flow rate parameters. Using the concentration measured by the gold standard device as the reference true value, a second-type sample dataset is formed, serving as the clinical breath dataset.

3. The method for detecting exhaled nitric oxide based on multi-source data fusion according to claim 1, characterized in that, In S2, the composite loss function is the weighted sum of the first-stage learning loss and the second-stage learning loss, L_total = L_task1 + λ * L_task2, where λ is a hyperparameter for adjusting the learning weights of the two tasks.

4. The method for detecting exhaled nitric oxide based on multi-source data fusion according to claim 1, characterized in that, The lightweighting process described in S3 includes structured pruning and low-bit integer quantization of the model, resulting in a final deployed model size of less than 200KB and a single inference time of less than 150 milliseconds on the embedded microprocessor.

5. The method for detecting exhaled nitric oxide based on multi-source data fusion according to claim 1, characterized in that, When creating the standard gas dataset, in a temperature- and humidity-controlled environment, standard nitric oxide gas of different known concentrations is introduced into the nitric oxide sensor. In the low concentration range of 0-10 ppb, concentration points are set at intervals no greater than 1 ppb; in the range above 10 ppb, points are set at wider intervals. At each concentration point, the sensor's high-frequency raw electrical signal waveform, ambient temperature, and humidity are recorded synchronously, using the known standard gas concentration as the true value. Each point is measured repeatedly to obtain statistical characteristics.

6. The method for detecting exhaled nitric oxide based on multi-source data fusion according to claim 1, characterized in that, It also includes comparing the output value with the gold standard reference value and evaluating the overall accuracy by calculating the root mean square error, mean absolute error and coefficient of determination.

7. A breath nitric oxide detection system based on multi-source data fusion, characterized in that, The steps for performing the exhaled nitric oxide detection method based on multi-source data fusion as described in any one of claims 1-5 include: a nitric oxide sensor module for acquiring raw electrical signals; an environmental sensor module for monitoring environmental temperature and humidity; a flow sensor module for monitoring exhaled flow; a microprocessor internally storing and running the pre-trained lightweight artificial intelligence correction model; and an output device for outputting nitric oxide concentration values.

8. The exhaled nitric oxide detection system based on multi-source data fusion according to claim 6, characterized in that, The lightweight AI correction model is stored in the microprocessor's non-volatile memory.

9. A breath nitric oxide detection device, characterized in that, The device includes the system as described in claims 6-7.

Images