Microfluidic biosensing identification method and system for non-diagnostic purposes

Abstract

The application relates to the technical field of biomedical detection, and discloses a microfluidic biosensing identification method and system for non-diagnostic purposes, wherein the microfluidic biosensing identification method for non-diagnostic purposes comprises the following steps: obtaining a whole blood sample, measuring and analyzing the whole blood sample, constructing a sample state vector, and dynamically adjusting working parameters of a sensor; synchronously collecting multi-modal sensing signals, obtaining the multi-modal sensing signals and environmental parameter signals; filtering, compensating and correcting the environmental parameter signals; performing variational modal decomposition on the multi-modal net signals and implementing physical constraint screening, extracting time domain and frequency domain characteristic parameters; performing evidence fusion and weighting on a multi-modal dynamic characteristic parameter set; training a transfer learning model and performing time sequence correlation analysis to obtain concentration estimation and a diagnosis conclusion; and the application can realize direct detection of a whole blood sample, collaborative analysis of multi-modal signals, real-time monitoring of a dynamic process, environmental self-adaptive compensation and accurate identification under small sample conditions.

Description

Technical Field

[0001] This invention relates to the field of biomedical detection technology, and more specifically, to a microfluidic biosensing identification method and system for non-diagnostic purposes. Background Technology

[0002] Microfluidic biosensing technology combines the miniaturization and integration advantages of microfluidic chips with the high sensitivity and specificity of biosensors, showing broad application prospects in areas such as point-of-care testing, early disease screening, and personalized medicine. Particularly in the diagnosis of time-sensitive diseases such as acute myocardial infarction, rapid and accurate detection of myocardial injury markers is crucial for timely initiation of treatment plans, reducing mortality, and improving prognosis. Traditional laboratory testing methods require samples to be sent to a central laboratory, with testing cycles lasting several hours, which is insufficient to meet the rapid diagnostic needs of pre-hospital emergency care.

[0003] Existing microfluidic biosensing technologies primarily employ methods such as immunochromatography, electrochemical detection, or optical detection to achieve rapid detection of biomarkers on microfluidic chips. However, most of these technologies use endpoint detection modes, only acquiring signal values ​​after reaction equilibrium, failing to capture the dynamic information of biomolecule binding processes and thus losing the diagnostic value inherent in reaction kinetic parameters. Furthermore, existing technologies often rely on a single sensing modality, resulting in insufficient anti-interference capabilities in complex sample matrices such as whole blood and non-standardized pre-hospital emergency environments, making it difficult to guarantee detection accuracy and stability. Simultaneously, clinical sample data from acute myocardial infarction patients at different stages of the disease are scarce and unevenly distributed. The training of existing recognition models relies on large amounts of labeled data, leading to poor generalization ability in areas with insufficient sample coverage, thus limiting the clinical applicability of the system.

[0004] Therefore, there is a need to develop a novel microfluidic biosensing identification method that can achieve direct detection of whole blood samples, collaborative analysis of multimodal signals, real-time monitoring of dynamic processes, adaptive environmental compensation, and accurate identification under small sample conditions, in order to meet the urgent need for rapid pre-hospital diagnosis of acute diseases. Summary of the Invention

[0005] This invention provides a microfluidic biosensing identification method and system for non-diagnostic purposes, which solves the technical problems of insufficient anti-interference ability, difficulty in guaranteeing detection accuracy and stability, and poor generalization ability in areas with insufficient sample coverage in related technologies.

[0006] This invention provides a microfluidic biosensing identification method for non-diagnostic purposes, comprising the following steps:

[0007] Obtain whole blood samples, measure and analyze the whole blood samples, construct sample state vectors, dynamically adjust the sensor's operating parameters based on the sample state vectors, and generate adaptive sensor parameter configurations;

[0008] Based on adaptive sensor parameter configuration, multimodal sensing signals are synchronously acquired to obtain multimodal sensing signals.

[0009] The multimodal sensing signals are filtered, compensated, and corrected to obtain a set of net multimodal signals;

[0010] Variational mode decomposition is performed on the multimodal net signal and physical constraint screening is implemented to extract time-domain and frequency-domain feature parameters, thereby obtaining a set of multimodal dynamic feature parameters;

[0011] Based on the sample state vector, evidence fusion and weighting are performed on the multimodal dynamic feature parameter set to obtain a comprehensive feature vector;

[0012] Based on the comprehensive feature vector, a transfer learning model is trained and a time-series correlation analysis is performed to obtain concentration estimation and diagnostic conclusions.

[0013] In a preferred embodiment, the step of constructing the sample state vector includes:

[0014] An optical transmission detection unit is set up to calculate the relative transmittance of the sample. The reciprocal of the relative transmittance is logarithmically transformed to obtain the optical density value, which is used as the turbidity parameter.

[0015] A pair of parallel plate electrodes are integrated, the AC impedance between the electrodes is measured, and the conductivity of the sample is calculated by combining the geometric parameters of the electrodes, which is then used as the conductivity parameter.

[0016] Add a second wavelength measurement, calculate the optical density values ​​of the first and second wavelengths respectively, and calculate the optical density ratio of the two wavelengths as a hemolysis parameter;

[0017] The turbidity parameter, conductivity parameter, and hemolysis parameter are combined into a sample state vector.

[0018] In a preferred embodiment, the step of synchronously acquiring multimodal sensing signals includes:

[0019] The device is equipped with an electrochemical impedance sensing unit, an optical scattering sensing unit, and a piezoelectric sensing unit. The electrochemical impedance sensing unit includes an interdigitated electrode array, the optical scattering sensing unit includes a side-scattering light detection device, and the piezoelectric sensing unit includes a piezoelectric crystal resonator.

[0020] An integrated flow velocity sensor dynamically adjusts the sampling frequency of the three sets of sensing units based on the measured flow velocity value;

[0021] For the electrochemical impedance sensing unit, a differential measurement mode is adopted, and differential impedance timing data is obtained by calculating the impedance difference between the working electrode and the reference electrode.

[0022] Simultaneous monitoring of environmental parameters such as flow velocity, temperature, and triaxial vibration acceleration.

[0023] In a preferred embodiment, the step of obtaining the multimodal net signal set includes:

[0024] Fast Fourier Transform is performed on the triaxial acceleration time-series signal to obtain the vibration spectrum distribution. The frequency components with amplitudes exceeding a preset threshold in the spectrum are identified by the peak detection algorithm. An adaptive notch filter array is designed for the piezoelectric sensing signal.

[0025] The electrochemical impedance time series signal and the optical scattering time series signal are decomposed into multiple scales using discrete wavelet transform. The sliding window method is used to identify abnormal coefficients and perform soft thresholding. The processed coefficients of each layer are reconstructed by inverse wavelet transform to obtain the signal with vibration interference removed.

[0026] Calculate the rate of temperature change based on the temperature time series signal, calculate the temperature correction factor, and multiply each segment of the signal by the corresponding correction factor.

[0027] The baseline trend is estimated by using a polynomial fitting method, and the baseline-free signal is obtained by subtracting the fitted baseline polynomial from the original signal.

[0028] In a preferred embodiment, the step of obtaining the multimodal dynamic feature parameter set includes:

[0029] The flow field distribution within the channel is calculated by using the finite difference method to obtain the velocity components at each grid point, and the shear stress distribution is calculated based on the velocity field.

[0030] A transport model of biomolecules in a microfluidic channel was established, and the spatiotemporal evolution of biomolecule concentration was described by the convection-diffusion equation to obtain the concentration change curve of the sensor surface over time.

[0031] The theoretical response signal of the sensor is predicted by combining a surface bonding dynamics model;

[0032] Variational mode decomposition is performed on the multimodal net signal, the correlation coefficient between each mode function and the theoretical response curve is calculated, modes with correlation coefficients higher than a preset threshold are retained, and modes with correlation coefficients lower than a preset threshold are removed;

[0033] Extract time-domain feature parameters, including start time, inflection point time, plateau time, rise slope, and plateau value; extract frequency-domain feature parameters, including dominant frequency, spectral centroid, and spectral bandwidth.

[0034] In a preferred embodiment, the step of obtaining the comprehensive feature vector includes:

[0035] The relationship between the reliability of the optical mode and the turbidity parameter was established, and the relationship between the reliability of the electrochemical mode and the conductivity and hemolysis parameters was established to obtain the reliability scores of the three sensing modes.

[0036] For each sensing modality, the degree of support of the modality for each diagnostic result is calculated based on the extracted feature parameters. A basic probability allocation function is constructed, and the evidence weight of each modality is adjusted according to the confidence score.

[0037] For the same biomarker, the evidence combination rule is applied to fuse evidence, and the basic probability allocation after fusion is calculated. The evidence combination rule is applied again for multiple biomarkers to obtain a comprehensive basic probability allocation function for diagnosis.

[0038] The feature parameters of each modality and each marker are weighted and averaged according to the confidence score to calculate the correlation features between multiple markers. The fused single marker feature parameters and the correlation features of multiple markers are combined to form a comprehensive feature vector.

[0039] In a preferred embodiment, the steps of obtaining the concentration estimate and diagnostic conclusion include:

[0040] A multi-physics coupled simulation model of microfluidic biosensing process was established, including a fluid dynamics module, a biomolecule transport module, a surface binding reaction module, and a sensor response module.

[0041] Virtual samples are generated by changing the input parameters of the simulation model. The virtual samples are subjected to the same signal processing and feature extraction as the real samples to obtain a virtual training dataset. The virtual training dataset and the real training dataset are then merged to form a hybrid training dataset.

[0042] The recognition model is trained using a transfer learning strategy. The training process is divided into a pre-training stage and a fine-tuning stage. In the pre-training stage, the model is trained using a virtual training dataset, and in the fine-tuning stage, the pre-trained model is fine-tuned using a real training dataset.

[0043] The comprehensive feature vector is input into the trained recognition model to obtain the concentration estimate of the marker. The temporal correlation features of the marker are analyzed to make disease diagnosis and inference of onset time. The diagnosis conclusion is obtained by matching the diagnostic rule base.

[0044] Dropout layers are added between the fully connected layers of the model. During the prediction phase, dropout activation is maintained for forward propagation. The mean and standard deviation of the prediction results are calculated. The mean is used as the final concentration estimate, and the standard deviation reflects the uncertainty of the prediction.

[0045] In a preferred embodiment, the step of dynamically adjusting the sensor's operating parameters based on the sample state vector includes:

[0046] For electrochemical impedance sensors, when the conductivity parameter is higher than the first threshold, the amplitude of the excitation signal is increased and the measurement frequency range is adjusted to shift to a higher frequency. When the conductivity parameter is lower than the second threshold, the standard excitation amplitude and frequency range are used.

[0047] For optical scattering sensors, when the turbidity parameter is higher than the third threshold, the light source power is reduced and the detector gain is increased; when the turbidity parameter is lower than the fourth threshold, the standard light source power and detector gain are used.

[0048] For piezoelectric sensors, frequency compensation is performed based on the real-time measured temperature value;

[0049] When the hemolysis parameter is higher than the fifth threshold, a correction step for free hemoglobin interference is added to the signal processing.

[0050] In a preferred embodiment, the correlation characteristics between the multiple markers include the ratio of different marker concentrations, the difference in rise time, and the ratio of rise slope.

[0051] Temporal correlation analysis infers the onset time window by comparing the rise rate and rise time of different biomarkers, and calculates the ratio of biomarker concentrations as a quantitative indicator of onset time.

[0052] The diagnostic rule base is built on clinical guidelines and expert knowledge, and it associates biomarker concentration, concentration change rate, concentration ratio with diagnostic conclusions.

[0053] This invention provides a microfluidic biosensing and identification system for non-diagnostic purposes, used to perform the aforementioned microfluidic biosensing and identification method for non-diagnostic purposes, comprising:

[0054] The sample preprocessing module is used to acquire whole blood samples, measure and analyze the whole blood samples, construct sample state vectors, dynamically adjust the sensor's operating parameters based on the sample state vectors, and generate adaptive sensor parameter configurations.

[0055] The signal acquisition module, based on adaptive sensor parameter configuration, performs synchronous acquisition of multimodal sensing signals to obtain multimodal sensing signals;

[0056] The signal processing module is used to filter, compensate, and correct the multimodal sensing signals to obtain a set of net multimodal signals.

[0057] The feature extraction module is used to perform variational mode decomposition on the multimodal net signal and implement physical constraint screening to extract time-domain and frequency-domain feature parameters, thereby obtaining a set of multimodal dynamic feature parameters.

[0058] The feature fusion module, based on the sample state vector, performs evidence fusion and weighting on the multimodal dynamic feature parameter set to obtain a comprehensive feature vector;

[0059] The identification and analysis module, based on comprehensive feature vectors, performs transfer learning model training and time-series correlation analysis to obtain concentration estimates and diagnostic conclusions.

[0060] The beneficial effects of this invention are as follows:

[0061] By establishing a multi-dimensional quantitative evaluation mechanism for sample states and an adaptive synchronous acquisition method for multi-modal sensor signals, adaptive compensation for individual differences in whole blood samples and collaborative acquisition of multi-physics information were achieved. By evaluating the turbidity, conductivity, and hemolysis degree of the samples in the pre-processing region, the operating parameters of each sensor were dynamically adjusted to ensure that the sensors operated within their optimal sensitivity range under different sample states. By integrating three complementary sensing modes—electrochemical impedance spectroscopy, optical scattering, and piezoelectric vibration—and adaptively adjusting the sampling frequency according to flow rate, high-quality multi-modal time-series data reflecting the entire process of biomolecule binding were obtained. This adaptive acquisition strategy improves the system's adaptability to sample differences and environmental changes, laying a solid foundation for subsequent accurate identification.

[0062] This system employs a multimodal feature fusion method guided by physical constraints and using dynamic feature extraction and sample state perception to effectively suppress interference from complex environments and intelligently fuse multi-source information. By establishing fluid dynamics and biomolecular transport models for microfluidic channels, physical laws are used as constraints to guide the feature extraction process, effectively distinguishing genuine biological signals from noise artifacts. Robust signal processing methods such as variational mode decomposition and multi-scale wavelet decomposition, combined with real-time monitoring of vibration acceleration and temperature, effectively suppress interference from non-standard environmental factors such as vibration and temperature fluctuations. By dynamically adjusting the confidence weights of different sensing modes based on sample state and using an evidence theory framework for multimodal fusion, the system fully utilizes the complementarity of multimodal signals, improving the accuracy and robustness of identification and enabling the system to operate stably and reliably in non-standard environments such as pre-hospital emergency care. Attached Figure Description

[0063] Figure 1 This is a flowchart of a microfluidic biosensing identification method for non-diagnostic purposes according to the present invention;

[0064] Figure 2 This is a block diagram of a microfluidic biosensing and identification system for non-diagnostic purposes according to the present invention. Detailed Implementation

[0065] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0066] At least one embodiment of the present invention discloses a microfluidic biosensing identification method for non-diagnostic purposes, such as... Figure 1 As shown, it includes:

[0067] Step 1: Obtain whole blood samples, measure and analyze the whole blood samples, construct sample state vectors, dynamically adjust the sensor's operating parameters based on the sample state vectors, and generate adaptive sensor parameter configurations.

[0068] Specifically, the following steps are included:

[0069] Step 1.1, optical transmission detection of sample turbidity;

[0070] An optical transmission detection unit is set up in the pretreatment area of ​​the microfluidic channel. This unit includes a light-emitting diode (LED) light source and a photodiode detector, located on the upper and lower sides of the microfluidic channel, respectively, with the light path passing perpendicularly through the sample channel. The LED emits red light with a wavelength of 660 nm, which has relatively weak absorption by hemoglobin and mainly reflects the scattering characteristics of the sample. When the whole blood sample flows through the detection area, the photodiode continuously measures the intensity of transmitted light at a sampling frequency of 100 Hz. According to the extended form of Beer-Lambert's law, the ratio of transmitted light intensity to incident light intensity reflects the total attenuation of the sample, which includes both absorption and scattering. The relative transmittance of the sample is calculated by measuring the transmitted light intensity in the blank buffer as a reference value. The reciprocal of the relative transmittance is logarithmically transformed to obtain the optical density value. This optical density value is mainly contributed by the scattering of blood cells and lipid particles in the sample, and therefore can be used as a quantitative indicator of sample turbidity. The continuously collected optical density values ​​are averaged over time to obtain a stable turbidity measurement value, denoted as the turbidity parameter T.

[0071] Step 1.2, Electrochemical detection of sample conductivity;

[0072] A pair of parallel plate electrodes, 500 micrometers apart, are integrated within a microfluidic channel in the pretreatment region. The electrodes are made of platinum. An AC voltage signal with a frequency of 10 kHz and an amplitude of 10 mV is applied to the electrodes using an AC impedance analyzer. At this frequency, the double-layer capacitance at the electrode-solution interface is negligible, and the measured impedance primarily reflects the bulk resistance of the solution. According to Ohm's law, solution resistance is inversely proportional to conductivity. By measuring the AC impedance between the electrodes and combining it with the electrode geometry (electrode area and electrode spacing), the conductivity of the sample is calculated. Conductivity reflects the total concentration and mobility of ions in the sample, primarily contributed by electrolyte ions such as sodium, potassium, and chloride ions, as well as charged protein molecules. Time averaging of continuously measured conductivity values ​​yields a stable conductivity measurement, denoted as the conductivity parameter C.

[0073] Step 1.3, spectral characteristic analysis of the degree of hemolysis in the sample;

[0074] Using the optical transmission detection unit from step 1.1, a 415 nm wavelength measurement is added to the existing 660 nm wavelength measurement. The light-emitting diode (LED) light source is switched between different wavelengths, or a dual-wavelength light source is used to simultaneously emit light at both wavelengths. Photodiodes measure the transmitted light intensity at each wavelength. The 415 nm wavelength is located in the Solette absorption band of hemoglobin, exhibiting strong absorption characteristics for free hemoglobin and relatively weak absorption for intact red blood cells. When hemolysis occurs in the sample, red blood cells rupture, releasing hemoglobin into the plasma, leading to an increase in absorption at the 415 nm wavelength, while the absorption at the 660 nm wavelength changes relatively little. The optical density ratio of the two wavelengths is calculated, which effectively reflects the degree of hemolysis. Specifically, the optical density values ​​at the 415 nm and 660 nm wavelengths are calculated separately and denoted as OD415 and OD660; the ratio H is calculated as OD415 divided by OD660. For normal, non-hemolyzed samples, this ratio is at a low level; when hemolysis occurs, the ratio increases. This ratio is used as a hemolysis indicator and denoted as the hemolysis parameter H.

[0075] Step 1.4, Construction of sample state vector and adaptive adjustment of sensor parameters;

[0076] The turbidity parameter T obtained in step 1.1, the conductivity parameter C obtained in step 1.2, and the hemolysis parameter H obtained in step 1.3 are combined to form a three-dimensional sample state vector, denoted as V, which is equal to the combination of T, C, and H. This vector quantifies the initial state characteristics of the sample. Based on the sample state vector, the operating parameters of each sensor are dynamically adjusted during subsequent detection. The specific adjustment strategies are as follows: For electrochemical impedance sensors, when the conductivity parameter C is higher than the preset threshold C1, it indicates that the sample ion intensity is high and the background conductivity is large. In this case, the amplitude of the excitation signal is increased to improve the signal-to-noise ratio, and the measurement frequency range is adjusted to shift towards higher frequencies. When the conductivity parameter C is lower than the preset threshold C2, the standard excitation amplitude and frequency range are used. For optical scattering sensors, when the turbidity parameter T is higher than the preset threshold T1, it indicates that the sample scattering is strong. In this case, the light source power is reduced to avoid detector saturation, and the detector gain is increased. When the turbidity parameter T is lower than the preset threshold T2, the standard light source power and detector gain are used. For piezoelectric sensors, their operating parameters are mainly affected by temperature, and frequency compensation is performed based on the real-time measured temperature value. When the hemolysis parameter H is higher than the preset threshold H1, it indicates that the sample has significant hemolysis. In this case, a correction step for free hemoglobin interference needs to be added in subsequent signal processing. Through the above adaptive adjustment, it is ensured that each sensor operates in the optimal sensitivity range under different sample conditions, and the adjusted sensor parameter configuration is output.

[0077] In some embodiments, a multi-wavelength spectral scanning method is used instead of dual-wavelength detection to assess the degree of hemolysis. Specifically, a tunable light source or spectrometer is configured in the preprocessing area, and spectral scanning is performed in the wavelength range of 400 nm to 700 nm at 10 nm intervals to obtain the transmission spectrum curve of the sample. By analyzing the morphological characteristics of the spectral curve, especially the position and intensity of the absorption peaks of the Solette band and Q band, the contributions of free hemoglobin, oxyhemoglobin, and deoxyhemoglobin can be more accurately distinguished, thereby obtaining a more accurate assessment of the degree of hemolysis. The purpose of this alternative embodiment is to improve the accuracy and anti-interference ability of hemolysis detection, especially when other interfering factors such as hyperlipidemia are present in the sample. Multi-wavelength spectral analysis can separate the contributions of different interference sources through pattern recognition methods.

[0078] Step 2: Based on adaptive sensor parameter configuration, synchronously acquire multimodal sensing signals to obtain multimodal sensing signals;

[0079] Specifically, the following steps are included:

[0080] Step 2.1, configuration of the multi-physics sensor array in the detection area;

[0081] In the detection area of ​​the microfluidic channel, three sets of sensing units are arranged sequentially according to the fluid flow direction. The first set is an electrochemical impedance sensing unit, containing an array of interdigitated electrodes made of gold. The electrodes are 20 micrometers wide and 20 micrometers apart. The electrode surfaces are modified with capture antibodies using self-assembled monolayer technology, targeting three cardiac markers: troponin I, myoglobin, and creatine kinase isoenzyme. The second set is an optical scattering sensing unit, containing a side-scattering light detection device. A laser diode is incident from the side of the channel at a 45-degree angle with a wavelength of 635 nanometers. The scattered light is received by a photomultiplier tube located vertically. This configuration can detect changes in interfacial scattering characteristics caused by biomolecule binding. The third set is a piezoelectric sensing unit, integrating a piezoelectric quartz crystal resonator at the bottom of the channel with a resonant frequency of 10 MHz. The crystal surface is also modified with capture antibodies. When biomolecules bind to the crystal surface, the increased mass load causes a decrease in the resonant frequency. The detection areas of the three sets of sensing units are spatially close together, with a spacing of less than 1 millimeter, ensuring that the same part of the sample flow process is detected, achieving true synchronous detection.

[0082] Step 2.2, Adaptive synchronous sampling control of flow rate;

[0083] A thermal microflow sensor is integrated at the entrance of the detection area. This sensor includes a pair of thermistors, and the fluid velocity is calculated by measuring the temperature difference between the upstream and downstream thermistors. The flow velocity sensor has a sampling frequency of 1000 Hz, enabling real-time tracking of instantaneous changes in flow velocity. The sampling frequency of the three sensing units is dynamically adjusted based on the measured flow velocity value. Specifically, a baseline flow velocity of 100 μL / min is set, corresponding to a baseline sampling frequency of 100 Hz. When the measured flow velocity is higher than the baseline flow velocity, the sampling frequency is increased linearly according to the ratio of the flow velocity to the baseline flow velocity, ensuring sufficient data density per unit flow distance. When the measured flow velocity is lower than the baseline flow velocity, the sampling frequency is reduced accordingly to minimize data redundancy. The sampling frequency adjustment range is set from 50 Hz to 500 Hz. Sampling by the three sensing units is triggered by a unified clock signal, ensuring strict time synchronization. The sampling process begins when the sample enters the detection area and continues until the detection ends, typically lasting 3 to 5 minutes, obtaining the timing signal data of each sensing unit.

[0084] Step 2.3, differential acquisition of electrochemical impedance signal;

[0085] For the electrochemical impedance sensing unit, a differential measurement mode was employed to suppress common-mode interference. Specifically, each biomarker corresponds to two adjacent interdigitated electrode pairs. One electrode pair is modified with a specific capture antibody as the working electrode, and the other electrode pair is modified with a non-specific protein, such as bovine serum albumin, as the reference electrode. The impedance spectra of the working and reference electrodes were simultaneously measured using an electrochemical workstation, with a measurement frequency range of 0.1 Hz to 100 kHz. A 10 mV AC excitation voltage was applied at each frequency point, and the amplitude and phase of the current response were measured. The impedance change of the working electrode includes both specific binding and non-specific adsorption of the target molecule, while the impedance change of the reference electrode only includes the non-specific adsorption portion. By calculating the impedance difference between the working and reference electrodes, common-mode interference caused by non-specific adsorption and environmental factors was eliminated, yielding an impedance change signal that purely reflects specific binding. Three sets of differential impedance time-series data were obtained for each of the three biomarkers.

[0086] Step 2.4, synchronous monitoring of environmental parameters;

[0087] During the detection process, key environmental parameters affecting the sensor response are monitored simultaneously. A temperature sensor, using a platinum resistance thermometer, is integrated on the microfluidic chip, with a measurement accuracy of 0.1 degrees Celsius and a sampling frequency of 10 Hz, recording real-time temperature changes in the detection area. A triaxial accelerometer is integrated at a fixed position on the chip, with a measurement range of ±2g and a sampling frequency of 1000 Hz, recording the chip's vibration acceleration in three spatial directions. Temperature and vibration data are strictly time-synchronized with the sensor signals, providing a basis for subsequent environmental sensing signal processing. After completing the acquisition process from steps 2.2 to 2.4, the following time-series signals are output: electrochemical impedance spectroscopy, optical scattering, and piezoelectric frequency shift for the three markers, as well as flow velocity, temperature, and vibration acceleration.

[0088] In some embodiments, surface-enhanced Raman scattering (SERS) is used to replace side-scattered light detection to improve the sensitivity of optical detection. Specifically, an array of gold or silver nanoparticles is deposited on the surface of the detection region of a microfluidic channel to form a surface-enhanced Raman scattering active substrate. When target biomolecules bind to the substrate surface, the Raman scattering signal of the molecules is enhanced by several orders of magnitude by the localized surface plasmon resonance effect of the nanoparticles. By acquiring the intensity changes of characteristic Raman peaks using a Raman spectrometer, highly sensitive monitoring of the biomolecule binding process can be achieved. This alternative embodiment aims to further improve detection sensitivity, particularly suitable for cases where biomarker concentrations are extremely low in patients with very early myocardial infarction, reducing the detection limit to the picogram per milliliter level.

[0089] Step 3: Filter, compensate, and correct the multimodal sensing signals to obtain a set of net multimodal signals;

[0090] Specifically, the following steps are included:

[0091] Step 3.1, Frequency domain identification and adaptive suppression of vibration interference;

[0092] The triaxial acceleration time-series signal obtained in step 2 is subjected to a Fast Fourier Transform (FFT) to obtain the vibration spectral distribution. A peak detection algorithm is used to identify frequency components in the spectrum whose amplitude exceeds a preset threshold. These frequency components correspond to the main vibration interference sources, such as vehicle engine vibration and road bumps. Typically, the identified main vibration frequencies are concentrated in the range of 10 Hz to 100 Hz. For the piezoelectric sensing signal, an adaptive notch filter array is designed. The number of notches is the same as the number of identified vibration frequency components. The center frequency of each notch filter is set to the corresponding vibration frequency, the notch bandwidth is set to 10% of the center frequency, and the notch depth is set to 40 dB. The original piezoelectric sensing time-series signal is passed sequentially through the notch filter array to effectively suppress vibration frequency components while retaining the frequency shift signal caused by biomolecule binding. For electrochemical impedance signals and optical scattering signals, vibration mainly causes random fluctuations rather than specific frequency interference; therefore, different processing strategies are adopted.

[0093] Step 3.2, multi-scale wavelet decomposition and vibrational abrupt change removal;

[0094] The electrochemical impedance spectroscopy (EIS) and optical scattering (OSS) time-series signals obtained in step 2 were decomposed using discrete wavelet transform (DWT) at multiple scales. The Daubechies wavelet was selected as the mother wavelet, and the decomposition was set to five levels, resulting in five detail coefficient sequences and one approximation coefficient sequence. The first level of detail coefficients corresponds to the highest frequency components, mainly containing measurement noise; the second and third levels correspond to mid-to-high frequency components, containing abrupt changes caused by vibration; and the fourth and fifth levels, along with the approximation coefficients, correspond to low-frequency components, containing slowly changing signals related to biomolecule binding. Statistical analysis was performed on the detail coefficients at each level, calculating the local variance of each coefficient. When a coefficient value at a certain time point deviates from the statistical distribution of its neighborhood, it is identified as a vibration-induced abrupt change. Specifically, a sliding window method was used, with a window length of 50 sampling points. The mean and standard deviation of the coefficients within the window were calculated. When a coefficient value exceeds the range of the mean plus or minus three times the standard deviation, it is identified as an outlier. The identified outlier coefficients were then subjected to soft thresholding, compressing their amplitude to the threshold level to remove the influence of vibration abrupt changes. The processed coefficients are subjected to inverse wavelet transform to reconstruct the signal with vibration interference removed.

[0095] Step 3.3, segmented detection and compensation of temperature changes;

[0096] Based on the temperature time-series signal obtained in step 2, analyze the temperature change trend. Calculate the time derivative of the temperature, i.e., the rate of temperature change. When the absolute value of the rate of temperature change is less than a preset threshold, such as 0.1 degrees Celsius per minute, the temperature is considered to be in a quasi-steady state, and can be assumed to be constant over a relatively long period. When the absolute value of the rate of temperature change exceeds the threshold, the temperature is considered to be changing rapidly, requiring segmentation. Based on the changes in the sign and amplitude of the rate of temperature change, the entire detection time is divided into several quasi-steady-state segments and rapid change segments. For each temperature segment, calculate the average temperature value within that segment. Based on the Arrhenius equation, the biochemical reaction rate constant has an exponential relationship with temperature. Set the standard temperature to 25 degrees Celsius, and calculate the temperature correction factor for temperature segments deviating from the standard temperature. Specifically, the temperature correction factor is equal to the exponent of the natural constant, where the exponent is the activation energy divided by the gas constant and then by the reciprocal of the temperature difference. For the binding reaction of myocardial markers and antibodies, a typical activation energy of 40 kJ / mol is used. Multiply the sensing signal in each temperature range by the corresponding temperature correction factor to achieve temperature normalization of the signal, and obtain the signal response equivalent to that under standard temperature.

[0097] Step 3.4, polynomial fitting correction for baseline drift;

[0098] Even after processing steps 3.1 to 3.3, the signals may still exhibit slow baseline drift caused by factors such as sensor aging and surface contamination. For each sensing channel, a polynomial fitting method is used to estimate the baseline trend. Specifically, it is assumed that the baseline drift can be described by a low-order polynomial, with the polynomial order set to 2 to 3. The overall trend of the signal is fitted using the least squares method to obtain the baseline polynomial. Subtracting the fitted baseline polynomial from the original signal yields the baseline-de-baseline signal. This signal, centered at zero, reflects the change relative to the initial state. After completing steps 3.1 to 3.4, the output multimodal net signal after environmental interference suppression is obtained, including the net electrochemical impedance signal, net optical scattering signal, and net piezoelectric frequency shift signal for the three markers.

[0099] In some embodiments, an adaptive Kalman filter is used instead of wavelet decomposition to remove vibration interference. Specifically, a state-space model of the sensing signal is established, modeling the biomolecular binding process as a slow evolution of state variables and vibration interference as observation noise. The Kalman filter recursively estimates the true state of the system based on the state model and the observation model, while suppressing observation noise. Unlike a fixed-parameter Kalman filter, the adaptive Kalman filter can dynamically adjust the noise covariance matrix according to the real-time vibration acceleration signal. When the vibration intensity increases, the observation noise variance increases, the weight of the observation value decreases, and it relies more on state prediction; when the vibration intensity decreases, the observation noise variance decreases, and the weight of the observation value increases. The purpose of this alternative embodiment is to provide a unified filtering framework that can simultaneously handle random noise and deterministic interference, and has online adaptive capabilities, making it particularly suitable for complex environments where vibration characteristics change rapidly.

[0100] Step 4: Perform variational mode decomposition on the multimodal net signal and implement physical constraint screening to extract time-domain and frequency-domain feature parameters, thereby obtaining a set of multimodal dynamic feature parameters;

[0101] Specifically, the following steps are included:

[0102] Step 4.1, Calculation of flow field distribution within the microfluidic channel;

[0103] Based on the geometric parameters of the microfluidic channel and the flow velocity measured in step 2, the flow field distribution within the channel was calculated. The microfluidic channel has a rectangular cross-section, with a width of 500 micrometers and a height of 100 micrometers. For laminar flow in a rectangular microchannel, the velocity field distribution can be obtained by solving the Navier-Stokes equations. Under low Reynolds number conditions, the inertial term can be neglected, and the equations simplify to the Stokes equations. A finite difference method was used to establish a grid on the channel cross-section with a grid spacing of 10 micrometers, and the velocity components at each grid point were obtained through iterative solutions. The calculation results show that the flow velocity is the highest at the center of the channel, gradually decreasing towards the wall, and reaching zero at the wall, forming a parabolic velocity profile. Based on the velocity field, the shear stress distribution was further calculated. Shear stress is equal to the fluid dynamic viscosity multiplied by the velocity gradient. At the channel wall, the velocity gradient is the largest, and the shear stress is also the largest. The calculated value of the wall shear stress reflects the intensity of the shearing action of the fluid on the sensor surface, affecting the transport efficiency of biomolecules to the surface.

[0104] Step 4.2: Establishment of a convective-diffusion model for biomolecular transport;

[0105] Based on the flow field distribution calculated in step 4.1, a transport model of biomolecules within the microfluidic channel is established. Biomolecule transport is controlled by both convection and diffusion mechanisms. Convective transport is driven by the fluid velocity field, while diffusion transport is driven by the concentration gradient. The spatiotemporal evolution of biomolecule concentration is described using a convection-diffusion equation. The left side of the equation is the partial derivative of concentration with respect to time, and the right side contains two terms: the first is the diffusion term, which equals the diffusion coefficient multiplied by the Laplace operator of concentration; the second is the convection term, which equals the velocity vector dot product of the negative concentration gradient. The diffusion coefficient is calculated according to the Stokes-Einstein relation, equal to the Boltzmann constant multiplied by temperature, divided by 6 times pi, multiplied by the hydrodynamic viscosity, and then divided by the molecular hydrodynamic radius. For cardiac biomarker proteins, the hydrodynamic radius is approximately 3 to 5 nanometers, and the calculated diffusion coefficient is approximately 10^-11 square meters per second. At the channel wall, i.e., the sensor surface, the boundary condition is set to allow the biomolecules to bind to the captured antibody; the reaction rate is described by the surface binding kinetics equation. By numerically solving the convection-diffusion equation, the spatiotemporal distribution of the biomolecule concentration field was obtained, especially the concentration change curve of the sensor surface over time.

[0106] Step 4.3, theoretical response prediction of surface bonding dynamics;

[0107] Based on the sensor surface concentration change curve obtained in step 4.2, and combined with the surface binding kinetic model, the theoretical response signal of the sensor is predicted. The surface binding process follows Langmuir adsorption kinetics; the binding rate is proportional to the density of unoccupied sites on the surface and the concentration of free molecules in the solution, while the dissociation rate is proportional to the density of bound molecules on the surface. The derivative of the surface binding density with respect to time equals the binding rate minus the dissociation rate. The binding rate constant and dissociation rate constant are key parameters characterizing the strength of antibody-antigen interactions. For high-affinity antibodies, the binding rate constant is approximately 10⁵ per mole per second, and the dissociation rate constant is approximately 10⁻⁴ per second. By solving the surface binding kinetic equation, the evolution curve of the surface binding density over time is obtained. This curve rises rapidly in the initial stage, reflecting the dominance of the binding process; subsequently, the rate of increase gradually slows down, eventually reaching a plateau value, reflecting that binding and dissociation have reached a dynamic equilibrium. For electrochemical impedance sensors, increased surface binding density leads to increased charge transfer resistance, with the impedance change proportional to the binding density. For optical scattering sensors, increased surface binding density alters the interfacial refractive index, with the scattered light intensity change proportional to the binding density. For piezoelectric sensors, increased surface binding density leads to increased mass loading, with the frequency shift proportional to the binding density. Based on the response mechanisms of each sensor, the surface binding density curves are converted into theoretical sensor response curves, serving as physical constraints for subsequent feature extraction.

[0108] Step 4.4, physical constraints for variational mode decomposition are implemented;

[0109] Variational mode decomposition (VMD) is performed on the multimodal net signal output from step 3 to extract signal components at different time scales. VMD decomposes the signal into several intrinsic mode functions (EMFs) with finite bandwidth by solving a variational optimization problem. Compared with empirical mode decomposition (EMD), VMD has a better mathematical foundation and noise resistance. The number of decomposed modes is set to four, corresponding to different physiological process time scales. The first mode corresponds to the fast initial binding process, with a time scale on the order of seconds; the second mode corresponds to the medium-speed transport constraint process, with a time scale on the order of tens of seconds; the third mode corresponds to the slow equilibrium approach process, with a time scale on the order of minutes; and the fourth mode corresponds to the extremely slow sensor drift process, with a time scale on the order of tens of minutes. The variational problem is solved using the alternating direction multiplier method, iteratively updating each mode function and its center frequency until convergence. Each obtained mode function must satisfy physical constraints, meaning its temporal evolution trend should be consistent with the theoretical response curve predicted in step 4.3. Specifically, the correlation coefficient between each modal function and the theoretical response curve is calculated. Modes with high correlation coefficients are considered genuine biological signal components, while modes with low correlation coefficients are considered noise or artifacts. A correlation coefficient threshold of 0.7 is set, retaining modes with correlation coefficients above the threshold and discarding modes with correlation coefficients below the threshold. The retained modal functions are then reconstructed to obtain the physically constrained filtered signal.

[0110] Step 4.5, Extraction of time-domain dynamic characteristic parameters;

[0111] For the signal reconstructed in step 4.4, temporal feature parameters reflecting biomolecule binding dynamics are extracted. Key time points of the signal are identified, including the start time, inflection point time, and plateau time. The start time is defined as the moment when the signal first exceeds three standard deviations of the initial noise level, marking the beginning of biomolecule arrival at the sensor surface. The inflection point time is defined as the moment when the second derivative of the signal is zero, corresponding to the maximum value of the signal rise rate, marking the transition of the binding process from transport-limited to reaction-limited. The plateau time is defined as the moment when the signal change rate drops below 10% of the maximum change rate, marking the approach of binding equilibrium. Feature parameters for each time period are calculated, including the time interval from start to inflection point, reflecting the speed of the transport process; and the time interval from inflection point to plateau, reflecting the speed of the binding reaction. The rise slope of the signal is calculated and defined as the first derivative value at the inflection point, reflecting the maximum binding rate. The plateau value of the signal is calculated, reflecting the binding density at equilibrium, which is related to the target molecule concentration. For the three sensing modes and three biomarkers, the above temporal feature parameters are extracted to form a temporal feature parameter matrix.

[0112] Step 4.6, Extraction of frequency domain feature parameters;

[0113] Frequency domain analysis was performed on the reconstructed signal from step 4.4 to extract frequency domain feature parameters. Using the short-time Fourier transform (SFT) method, the signal was divided into several time windows, each 30 seconds long, with a 50% overlap between windows. A Fourier transform was performed on the signal within each window to obtain the spectrum for that time window. The characteristic parameters of the spectrum were calculated, including the dominant frequency (the frequency component with the largest amplitude, reflecting the main oscillation period of the signal); the centroid of the spectrum (the weighted average of the product of frequency and amplitude, reflecting the center of the spectral energy distribution); and the bandwidth (the frequency range containing 90% of the energy, reflecting the frequency dispersion of the signal). The evolution trend of the dominant frequency over time was analyzed. At different stages of the bonding process, the dominant frequency may drift, reflecting changes in the bonding mechanism. The spectral correlation between adjacent time windows was calculated; high correlation indicates a stable bonding process, while low correlation indicates abrupt changes or interference. For the three sensing modes and three markers, the above frequency domain feature parameters were extracted to form a frequency domain feature parameter matrix. After completing steps 4.1 to 4.6, the set of multimodal dynamic feature parameters after physical constraint filtering is output, including the time-domain feature parameter matrix and the frequency-domain feature parameter matrix.

[0114] In some embodiments, the Hilbert-Huang transform method is used instead of variational mode decomposition to extract time-frequency features. Specifically, the signal is decomposed into several intrinsic mode functions (EMFs) through empirical mode decomposition. A Hilbert transform is performed on each EMF to obtain the instantaneous amplitude, instantaneous frequency, and instantaneous phase. The instantaneous frequencies and amplitudes of all EMFs are expanded on the time-frequency plane to form a Hilbert spectrum. The Hilbert spectrum can accurately describe the time-frequency distribution of a signal, and is particularly suitable for analyzing non-stationary and nonlinear signals. Marginal spectra (i.e., the frequency-energy distribution obtained by integrating over time) and instantaneous energy spectra (i.e., the time-energy distribution obtained by integrating over frequency) are extracted from the Hilbert spectrum. These time-frequency features can more finely characterize the dynamic evolution of biomolecular binding processes. The purpose of this alternative embodiment is to provide higher time-frequency resolution, capable of capturing transient features and nonlinear phenomena in the binding process, and is particularly suitable for complex multi-step binding reactions.

[0115] Step 5: Based on the sample state vector, perform evidence fusion and weighting on the multimodal dynamic feature parameter set to obtain a comprehensive feature vector;

[0116] Specifically, the following steps are included:

[0117] Step 5.1, Evaluation of the impact of sample state on the reliability of sensing modes;

[0118] Based on the sample state vector obtained in step 1, the influence of each sample characteristic parameter on the reliability of different sensing modes is analyzed. For the optical scattering sensing mode, its reliability is mainly affected by the turbidity parameter T. When the turbidity parameter T is high, blood cells and lipid particles in the sample generate strong background scattering, reducing the signal-to-noise ratio of the target signal and decreasing the reliability of the optical mode. The relationship between the reliability of the optical mode and the turbidity parameter is established, described by a sigmoid function. When the turbidity parameter is below the lower threshold T2, the reliability is at its maximum value of 1; when the turbidity parameter is above the upper threshold T1, the reliability drops to its minimum value of 0.2; between the two thresholds, the reliability transitions smoothly. For the electrochemical impedance sensing mode, its reliability is mainly affected by the conductivity parameter C and the hemolysis parameter H. When the conductivity parameter C deviates from the normal range, it indicates that the ionic strength of the sample is abnormal, affecting the kinetics of the electrochemical reaction and the double-layer structure, thus reducing the reliability of the electrochemical mode. When the hemolysis parameter H is high, intracellular substances such as free hemoglobin increase the electrochemical background signal and non-specific adsorption, reducing the reliability of the electrochemical mode. The relationship between the reliability of the electrochemical mode and the conductivity and hemolysis parameters was established using a multivariate sigmoid function to comprehensively consider the influence of both parameters. For the piezoelectric sensing mode, its reliability is mainly affected by the sample viscosity, which is related to turbidity and conductivity. The relationship between the reliability of the piezoelectric mode and the sample state parameters was established. Through the above analysis, reliability scores for the three sensing modes were obtained, denoted as optical reliability, electrochemical reliability, and piezoelectric reliability.

[0119] Step 5.2, a multimodal feature fusion framework based on evidence theory;

[0120] Multimodal feature fusion is performed using the Dempster-Shafer evidence theory framework. Evidence theory can handle uncertainty and incomplete information, making it suitable for multi-sensor fusion scenarios. An identification framework is defined as the set of all possible diagnostic results, including three states: normal, suspected myocardial infarction, and confirmed myocardial infarction. For each sensing modality, based on the feature parameters extracted in step 4, the support level of that modality for each diagnostic result is calculated, and a basic probability allocation function is constructed. The basic probability allocation function assigns unit probability mass to each subset of the identification framework, representing the support of evidence for that subset. Specifically, for a specific sensing modality of a marker, the support level for each diagnostic result is calculated based on the relationship between its feature parameters and a preset diagnostic threshold. For example, when the feature parameters are higher than the normal threshold, the support level for confirmed myocardial infarction is increased; when the feature parameters are slightly higher than the normal threshold, the support level for suspected myocardial infarction is increased; and when the feature parameters are within the normal range, the support level for the normal state is increased. Simultaneously, the evidence weights of each modality are adjusted based on the confidence score obtained in step 5.1. Modalities with high credibility have larger weights in their basic probability assignment functions and thus have a greater impact on the fusion result; modalities with low credibility have smaller weights and thus have a smaller impact.

[0121] Step 5.3, Application of evidence combination rules and calculation of fusion results;

[0122] For the three sensing modalities of the same biomarker, Dempster's combination rule is applied for evidence fusion. Dempster's combination rule calculates the joint basic probability assignment function of two evidence sources by summing the probability products of all possible intersections and then normalizing the result. Specifically, for any subset within the identification framework, the fused basic probability assignment is equal to the sum of the probability products of all evidence pairs whose intersection is that subset, divided by a normalization constant. The normalization constant is equal to 1 minus the sum of the probability products of all evidence pairs whose intersection is an empty set, reflecting the degree of conflict between the evidence. When the evidence is highly consistent, the conflict is small, and the normalization constant is close to 1; when the evidence is contradictory, the conflict is large, and the normalization constant is less than 1. The evidence from the three modalities is combined pairwise sequentially to obtain the fused basic probability assignment function for the biomarker. For each of the three biomarkers, the above fusion process is performed separately to obtain three fused basic probability assignment functions. Furthermore, Dempster's combination rule is applied again to the fusion results of the three biomarkers to obtain a comprehensive diagnostic basic probability assignment function. Decision information is extracted from the comprehensive basic probability assignment function, and the confidence and likelihood of each diagnostic result are calculated. Confidence indicates the degree to which the evidence directly supports the result, and likelihood indicates the degree to which the evidence does not contradict the result.

[0123] Step 5.4, Construction of fused feature vectors;

[0124] In addition to the probability allocation of diagnostic results, a fusion feature vector for subsequent identification models needs to be constructed. The feature parameters of each modality and biomarker extracted in step 4 are weighted and averaged according to the confidence scores obtained in step 5.1. Specifically, for a certain feature parameter, such as the rise slope, a weighted average of the measurements from the three modalities is calculated, with the weights being the normalized values ​​of the confidence scores for each modality. For the three biomarkers, the weighted average of each feature parameter is calculated separately, forming a fused feature parameter set. Furthermore, the correlation features between multiple biomarkers are calculated, including the ratio of different biomarker concentrations, the difference in rise time, and the ratio of rise slopes. These correlation features reflect the temporal correlation of multiple biomarkers and are of significant value for disease staging and onset time inference. The fused single-biomarker feature parameters and multi-biomarker correlation features are combined to form a comprehensive feature vector. After completing steps 5.1 to 5.4, the comprehensive feature vector after sample state perception fusion and the basic probability allocation function for comprehensive diagnosis are output.

[0125] In some embodiments, an attention mechanism of a deep neural network is used to replace evidence theory for multimodal fusion. Specifically, a multimodal fusion neural network is constructed, comprising three parallel feature extraction branches that process feature parameters for electrochemical, optical, and piezoelectric modes, respectively. Each branch consists of a multi-layer fully connected network that extracts mode-specific high-level feature representations. In the fusion layer, an attention mechanism module is introduced, which dynamically generates attention weights for each mode based on the sample state vector. The attention weights are calculated through a small neural network, with the sample state vector as input and the weight coefficients for the three modes as output. The high-level feature representations of each mode are multiplied by their corresponding attention weights and concatenated to form a fused feature vector. The fused feature vector is then input into a subsequent classification network to output a diagnostic result. The purpose of this alternative embodiment is to leverage the end-to-end learning capabilities of deep learning to automatically learn the optimal fusion strategy, avoiding manual design of fusion rules, and is particularly suitable for training scenarios with large-scale datasets.

[0126] Step 6: Based on the comprehensive feature vector, perform transfer learning model training and time series correlation analysis to obtain concentration estimation and diagnostic conclusions;

[0127] Specifically, the following steps are included:

[0128] Step 6.1, Establishment of the multiphysics coupling simulation model;

[0129] A multiphysics coupled simulation model of the microfluidic biosensing process was established to generate virtual training samples. The simulation model consists of four coupled physical modules. The first module is the fluid dynamics module, which calculates the flow field distribution within the microfluidic channel based on the Navier-Stokes equations. Input parameters include channel geometry, inlet velocity, fluid viscosity, and density, and outputs velocity and pressure fields. The second module is the biomolecule transport module, which calculates the concentration field distribution of biomolecules based on convection-diffusion equations. Input parameters include initial concentration, diffusion coefficient, and flow field distribution, and outputs the spatiotemporal evolution of the concentration field. The third module is the surface binding reaction module, which calculates the binding density on the sensor surface based on the Langmuir kinetic equations. Input parameters include surface concentration, binding rate constant, and dissociation rate constant, and outputs the time evolution of the surface binding density. The fourth module is the sensor response module, which calculates the sensing signal according to the physical mechanisms of different sensors. For electrochemical impedance sensors, it calculates the change in charge transfer resistance based on surface binding density; for optical scattering sensors, it calculates the change in refractive index and scattered light intensity based on surface binding density; and for piezoelectric sensors, it calculates the mass load and frequency shift based on surface binding density. The four modules are coupled through parameter passing: the output of the fluid dynamics module serves as the input of the transport module, the output of the transport module serves as the input of the reaction module, and the output of the reaction module serves as the input of the sensor response module. The finite element method is used to numerically solve the simulation model, establishing a three-dimensional mesh, setting boundary and initial conditions, and obtaining the spatiotemporal distribution of each physical quantity through time-stepping.

[0130] Step 6.2, parameterization of virtual samples;

[0131] By systematically modifying the input parameters of the simulation model, virtual samples covering a broad parameter space are generated. For biomarker concentration parameters, sampling is performed within clinically relevant ranges: troponin I concentration is set to 0.01–100 ng / mL, myoglobin concentration to 10–1000 ng / mL, and creatine kinase isoenzyme concentration to 1–100 ng / mL. A Latin hypercube sampling method is used to generate uniformly distributed sampling points in a three-dimensional concentration space, with each sampling point corresponding to a concentration combination. For flow rate parameters, sampling is performed within the range of 50–200 μL / min; for temperature parameters, sampling is performed within the range of 15–35 degrees Celsius; for sample matrix parameters, different hematocrit and lipemia levels are simulated by adjusting parameters such as background scattering coefficient and background conductivity in the simulation model. For each parameter combination, the simulation model is run to generate time-series signals for three sensing modalities. Noise consistent with actual measurement characteristics is superimposed on the simulation signals, including Gaussian white noise and 1 / f noise, with the noise level set according to the signal-to-noise ratio of the actual sensor. Through the above process, thousands of virtual samples are generated, each containing multimodal time-series signals and corresponding biomarker concentration labels.

[0132] Step 6.3, Feature extraction of virtual samples and dataset construction;

[0133] The virtual samples generated in step 6.2 undergo the same signal processing and feature extraction procedures as the real samples. Specifically, for the virtual multimodal time-series signals, the environmentally aware robust preprocessing of step 3, the physical constraint-guided dynamic feature extraction of step 4, and the sample state-aware multimodal feature fusion of step 5 are executed sequentially to obtain a comprehensive feature vector for each virtual sample. The comprehensive feature vector is paired with the corresponding biomarker concentration label to form a virtual training dataset. Simultaneously, a small number of real clinical samples are collected from diagnosed acute myocardial infarction patients and healthy controls. These samples undergo the same detection and feature extraction procedures to form a real training dataset. The number of real samples is typically tens to hundreds, far fewer than the number of virtual samples. The virtual and real training datasets are merged to form a hybrid training dataset. To balance the contributions of virtual and real samples, real samples are assigned a higher weight, with a weight ratio of 10:1, meaning the weight of one real sample is equal to the weight of 10 virtual samples.

[0134] Step 6.4, training the transfer learning recognition model;

[0135] A transfer learning strategy is employed to train the recognition model, fully leveraging the numerical advantage of virtual samples and the distribution characteristics of real samples. The recognition model utilizes a deep neural network architecture, comprising a feature encoder and a concentration regressor. The feature encoder, composed of multiple fully connected layers, maps the input comprehensive feature vector to a high-dimensional latent space, extracting abstract feature representations. The concentration regressor consists of multiple parallel regression branches, each corresponding to a specific biomarker, outputting an estimated concentration value for that biomarker. The training process is divided into two phases. The first phase is the pre-training phase, using a virtual training dataset to train the model. The loss function is defined as the mean squared error between the predicted and actual concentrations, and the stochastic gradient descent algorithm is used to optimize the model parameters. Due to the sufficient number of virtual samples, the model can fully learn the general laws of biosensing processes, including signal response patterns at different concentrations, the synergistic relationship of multimodal signals, and the influence of environmental factors. The pre-training phase involves a sufficient number of iterations until the loss function converges. The second phase is the fine-tuning phase, using a real training dataset to fine-tune the pre-trained model. In the fine-tuning phase, the low-level parameters of the feature encoder are frozen, and only the top-level parameters and the parameters of the concentration regressor are updated. This approach preserves the general feature representations learned from virtual samples while adapting to the distribution characteristics of real data. A smaller learning rate is used during the fine-tuning phase to avoid overfitting. Through transfer learning, the model achieves good generalization performance even with a small number of real samples.

[0136] Step 6.5, Multi-marker temporal association analysis and diagnostic inference;

[0137] The comprehensive feature vector output from step 5 is input into the recognition model trained in step 6.4 to obtain the concentration estimates of the three biomarkers. In addition to concentration values, the temporal correlation characteristics of the biomarkers also need to be analyzed for disease diagnosis and onset time estimation. The rate of change of biomarker concentration is calculated by numerical differentiation of the concentration time-series data. Myoglobin begins to rise rapidly 1 to 2 hours after myocardial infarction, with a large rate of increase; troponin I rises only 3 to 4 hours later, with a relatively small rate of increase. By comparing the rate of increase and the time of increase of different biomarkers, the onset time window can be inferred. Specifically, when the myoglobin concentration has increased while the troponin I concentration is still within the normal range, the onset time is inferred to be within 1 to 3 hours; when both are increased, the onset time is inferred to be more than 3 hours. The ratio of biomarker concentrations is calculated. The ratio of myoglobin to troponin I concentration is higher in the early stages of the disease and gradually decreases over time. The dynamic change of this ratio provides a quantitative indicator of the onset time. A diagnostic rule base is established, with rules based on clinical guidelines and expert knowledge, linking parameters such as biomarker concentration, rate of change of concentration, and concentration ratio with the diagnostic conclusion. For example, rule 1 states that if the troponin I concentration is greater than 0.04 ng / mL and the myoglobin concentration is greater than 70 ng / mL, the diagnosis is acute myocardial infarction; rule 2 states that if the troponin I concentration is between 0.01 and 0.04 ng / mL, the diagnosis is suspected myocardial infarction, requiring further investigation. Based on the concentration values ​​output by the recognition model and the calculated association features, a diagnostic rule base is matched to obtain the diagnostic conclusion.

[0138] Step 6.6, Uncertainty estimation and result output;

[0139] To assess the reliability of the diagnostic results, the uncertainty of the predictions was calculated. A Bayesian deep learning approach was employed, introducing a probability distribution into the parameters of the identification model and estimating the uncertainty through Monte Carlo sampling. Specifically, dropout layers were added between the fully connected layers of the model, maintaining dropout activation during the prediction phase and performing multiple forward propagations. Each propagation yielded different predictions due to the randomness of the dropout. The mean and standard deviation of the multiple predictions were calculated. The mean was used as the final concentration estimate, while the standard deviation reflected the uncertainty of the predictions. A large standard deviation indicated high uncertainty in the model's prediction of that sample, potentially due to deviations in sample features from the training data distribution or the presence of unidentified confounding factors. An uncertainty threshold was set; when the standard deviation exceeded the threshold, a low-confidence warning was displayed in the diagnostic report, prompting medical staff to consider other clinical information for comprehensive judgment or to conduct a retest. The final output was a comprehensive diagnostic report containing the following: concentration estimates and uncertainties of the three biomarkers, the rate of change of biomarker concentrations, the concentration ratios of the multiple biomarkers, the diagnostic conclusion (normal, suspected myocardial infarction, or confirmed myocardial infarction), the inference of the onset time window, and a diagnostic confidence assessment. After completing steps 6.1 to 6.6, small sample recognition based on physical model enhancement is achieved, and accurate and reliable diagnostic results are output.

[0140] In some embodiments, a generative adversarial network (GAN) approach is used to replace a physical simulation model to generate virtual training samples. Specifically, a GAN is constructed comprising two neural networks: a generator and a discriminator. The generator takes a random noise vector and a marker concentration label as input and outputs a virtual multimodal sensing signal. The discriminator takes the sensing signal as input and outputs the probability that the signal is a real or virtual sample. Through adversarial training, the generator learns to generate virtual samples with a distribution consistent with real samples, while the discriminator learns to distinguish between real and virtual samples. After training convergence, the generator can generate realistic virtual samples based on a given concentration label. Compared to a physical simulation model, a GAN does not require explicit establishment of physical equations but learns the sample distribution in a data-driven manner, enabling it to capture complex nonlinear relationships and randomness that are difficult for physical models to describe. The purpose of this alternative embodiment is to provide a more flexible method for generating virtual samples, particularly suitable for situations where the physical mechanisms are not fully understood or difficult to model accurately.

[0141] A non-diagnostic microfluidic biosensing and identification system is provided for performing the aforementioned non-diagnostic microfluidic biosensing and identification method, such as... Figure 2 As shown, it includes:

[0142] The sample preprocessing module is used to acquire whole blood samples, measure and analyze the whole blood samples, construct sample state vectors, dynamically adjust the sensor's operating parameters based on the sample state vectors, and generate adaptive sensor parameter configurations.

[0143] The signal acquisition module, based on adaptive sensor parameter configuration, performs synchronous acquisition of multimodal sensing signals to obtain multimodal sensing signals;

[0144] The signal processing module is used to filter, compensate, and correct the multimodal sensing signals to obtain a set of net multimodal signals.

[0145] The feature extraction module is used to perform variational mode decomposition on the multimodal net signal and implement physical constraint screening to extract time-domain and frequency-domain feature parameters, thereby obtaining a set of multimodal dynamic feature parameters.

[0146] The feature fusion module, based on the sample state vector, performs evidence fusion and weighting on the multimodal dynamic feature parameter set to obtain a comprehensive feature vector;

[0147] The identification and analysis module, based on comprehensive feature vectors, performs transfer learning model training and time-series correlation analysis to obtain concentration estimates and diagnostic conclusions.

[0148] In one embodiment of the present invention, a specific example is provided:

[0149] This invention focuses on the application of rapid pre-hospital emergency diagnosis for patients with acute myocardial infarction. In practical application, after an ambulance picks up a patient with chest pain, medical personnel collect approximately 50 microliters of blood from the patient's fingertip and inject the sample into the inlet of a microfluidic chip. The chip automatically draws in the sample through capillary action, without the need for an external pump. The sample flows through a pretreatment area, where the system automatically assesses the sample status and adjusts sensor parameters before entering the detection area, where multimodal sensors begin synchronous signal acquisition. The entire detection process lasts 5 minutes, during which the system processes signals and extracts features in real time. After the detection is completed, the system outputs a diagnostic report, which is displayed on a portable terminal device.

[0150] A 14-day field test was conducted within the jurisdiction of a municipal emergency medical center. During the test, two portable microfluidic biosensor systems were deployed, with data collection points covering ambulances and stations, collecting a total of 83 raw data sets. Examples of the two types of raw data acquisition are shown below:

[0151] Examples of data collected during the preprocessing stage are shown in Table 1:

[0152] Table 1: Example of data collected during the preprocessing stage;

[0153]

[0154] Table 2 shows examples of multimodal and environmental data collected in the detection area.

[0155] Table 2: Examples of multimodal and environmental data collected in the detection area;

[0156]

[0157] The microfluidic biosensing identification method of this invention enables rapid and accurate diagnosis of acute myocardial infarction in a pre-hospital emergency setting, saving valuable treatment time for patients and improving clinical prognosis.

[0158] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A microfluidic biosensing identification method for non-diagnostic purposes, characterized in that, Includes the following steps: Obtain whole blood samples, measure and analyze the whole blood samples, construct sample state vectors, dynamically adjust the sensor's operating parameters based on the sample state vectors, and generate adaptive sensor parameter configurations; Based on adaptive sensor parameter configuration, multimodal sensing signals are synchronously acquired to obtain multimodal sensing signals. The multimodal sensing signals are filtered, compensated, and corrected to obtain a set of net multimodal signals; Variational mode decomposition is performed on the multimodal net signal and physical constraint screening is implemented to extract time-domain and frequency-domain feature parameters, thereby obtaining a set of multimodal dynamic feature parameters; Based on the sample state vector, evidence fusion and weighting are performed on the multimodal dynamic feature parameter set to obtain a comprehensive feature vector; Based on the comprehensive feature vector, transfer learning model training and time-series correlation analysis are performed to obtain concentration estimation and diagnostic conclusions; The steps for obtaining the multimodal dynamic feature parameter set include: The flow field distribution within the channel is calculated by using the finite difference method to obtain the velocity components at each grid point, and the shear stress distribution is calculated based on the velocity field. A transport model of biomolecules in a microfluidic channel was established, and the spatiotemporal evolution of biomolecule concentration was described by the convection-diffusion equation to obtain the concentration change curve of the sensor surface over time. The theoretical response signal of the sensor is predicted by combining a surface bonding dynamics model; Variational mode decomposition is performed on the multimodal net signal, the correlation coefficient between each mode function and the theoretical response curve is calculated, modes with correlation coefficients higher than a preset threshold are retained, and modes with correlation coefficients lower than a preset threshold are removed; Extract time-domain feature parameters, including start time, inflection point time, plateau time, rise slope and plateau value; extract frequency-domain feature parameters, including main frequency, spectral centroid and spectral bandwidth. The steps for obtaining the comprehensive feature vector include: The relationship between the reliability of the optical mode and the turbidity parameter was established, and the relationship between the reliability of the electrochemical mode and the conductivity and hemolysis parameters was established to obtain the reliability scores of the three sensing modes. For each sensing modality, the degree of support of the modality for each diagnostic result is calculated based on the extracted feature parameters. A basic probability allocation function is constructed, and the evidence weight of each modality is adjusted according to the confidence score. For the same biomarker, the evidence combination rule is applied to fuse evidence, and the basic probability allocation after fusion is calculated. The evidence combination rule is applied again for multiple biomarkers to obtain a comprehensive basic probability allocation function for diagnosis. The feature parameters of each modality and each marker are weighted and averaged according to the confidence score to calculate the correlation features between multiple markers. The fused single marker feature parameters and the multi-marker correlation features are combined to form a comprehensive feature vector. The steps for obtaining concentration estimates and diagnostic conclusions include: A multi-physics coupled simulation model of microfluidic biosensing process was established, including a fluid dynamics module, a biomolecule transport module, a surface binding reaction module, and a sensor response module. Virtual samples are generated by changing the input parameters of the simulation model. The virtual samples are subjected to the same signal processing and feature extraction as the real samples to obtain a virtual training dataset. The virtual training dataset and the real training dataset are then merged to form a hybrid training dataset. The recognition model is trained using a transfer learning strategy. The training process is divided into a pre-training stage and a fine-tuning stage. In the pre-training stage, the model is trained using a virtual training dataset, and in the fine-tuning stage, the pre-trained model is fine-tuned using a real training dataset. The comprehensive feature vector is input into the trained recognition model to obtain the concentration estimate of the marker. The temporal correlation features of the marker are analyzed to make disease diagnosis and inference of onset time. The diagnosis conclusion is obtained by matching the diagnostic rule base. Dropout layers are added between the fully connected layers of the model. During the prediction phase, dropout activation is maintained for forward propagation. The mean and standard deviation of the prediction results are calculated. The mean is used as the final concentration estimate, and the standard deviation reflects the uncertainty of the prediction.

2. The microfluidic biosensing identification method for non-diagnostic purposes according to claim 1, characterized in that, The steps for constructing the sample state vector include: An optical transmission detection unit is set up to calculate the relative transmittance of the sample. The reciprocal of the relative transmittance is logarithmically transformed to obtain the optical density value, which is used as the turbidity parameter. A pair of parallel plate electrodes are integrated, the AC impedance between the electrodes is measured, and the conductivity of the sample is calculated by combining the geometric parameters of the electrodes, which is then used as the conductivity parameter. Add a second wavelength measurement, calculate the optical density values ​​of the first and second wavelengths respectively, and calculate the optical density ratio of the two wavelengths as a hemolysis parameter; The turbidity parameter, conductivity parameter, and hemolysis parameter are combined into a sample state vector.

3. The microfluidic biosensing and identification method for non-diagnostic purposes according to claim 1, characterized in that, The steps for synchronously acquiring multimodal sensing signals include: The device is equipped with an electrochemical impedance sensing unit, an optical scattering sensing unit, and a piezoelectric sensing unit. The electrochemical impedance sensing unit includes an interdigitated electrode array, the optical scattering sensing unit includes a side-scattering light detection device, and the piezoelectric sensing unit includes a piezoelectric crystal resonator. An integrated flow velocity sensor dynamically adjusts the sampling frequency of the three sets of sensing units based on the measured flow velocity value; For the electrochemical impedance sensing unit, a differential measurement mode is adopted, and differential impedance timing data is obtained by calculating the impedance difference between the working electrode and the reference electrode. Simultaneous monitoring of environmental parameters such as flow velocity, temperature, and triaxial vibration acceleration.

4. The microfluidic biosensing identification method for non-diagnostic purposes according to claim 1, characterized in that, The steps for obtaining the multimodal net signal set include: Fast Fourier Transform is performed on the triaxial acceleration time-series signal to obtain the vibration spectrum distribution. The frequency components with amplitudes exceeding a preset threshold in the spectrum are identified by the peak detection algorithm. An adaptive notch filter array is designed for the piezoelectric sensing signal. The electrochemical impedance time series signal and the optical scattering time series signal are decomposed into multiple scales using discrete wavelet transform. The sliding window method is used to identify abnormal coefficients and perform soft thresholding. The processed coefficients of each layer are reconstructed by inverse wavelet transform to obtain the signal with vibration interference removed. Calculate the rate of temperature change based on the temperature time series signal, calculate the temperature correction factor, and multiply each segment of the signal by the corresponding correction factor. The baseline trend is estimated by using a polynomial fitting method, and the baseline-free signal is obtained by subtracting the fitted baseline polynomial from the original signal.

5. A microfluidic biosensing identification method for non-diagnostic purposes according to claim 1, characterized in that, The step of dynamically adjusting the sensor's operating parameters based on the sample state vector includes: For electrochemical impedance sensors, when the conductivity parameter is higher than the first threshold, the amplitude of the excitation signal is increased and the measurement frequency range is adjusted to shift to a higher frequency. When the conductivity parameter is lower than the second threshold, the standard excitation amplitude and frequency range are used. For optical scattering sensors, when the turbidity parameter is higher than the third threshold, the light source power is reduced and the detector gain is increased; when the turbidity parameter is lower than the fourth threshold, the standard light source power and detector gain are used. For piezoelectric sensors, frequency compensation is performed based on the real-time measured temperature value; When the hemolysis parameter is higher than the fifth threshold, a correction step for free hemoglobin interference is added to the signal processing.

6. A microfluidic biosensing and identification system for non-diagnostic purposes, characterized in that, A microfluidic biosensing identification method for non-diagnostic purposes as described in any one of claims 1-5, comprising: The sample preprocessing module is used to acquire whole blood samples, measure and analyze the whole blood samples, construct sample state vectors, dynamically adjust the sensor's operating parameters based on the sample state vectors, and generate adaptive sensor parameter configurations. The signal acquisition module, based on adaptive sensor parameter configuration, performs synchronous acquisition of multimodal sensing signals to obtain multimodal sensing signals; The signal processing module is used to filter, compensate, and correct the multimodal sensing signals to obtain a set of net multimodal signals. The feature extraction module is used to perform variational mode decomposition on the multimodal net signal and implement physical constraint screening to extract time-domain and frequency-domain feature parameters, thereby obtaining a set of multimodal dynamic feature parameters. The feature fusion module, based on the sample state vector, performs evidence fusion and weighting on the multimodal dynamic feature parameter set to obtain a comprehensive feature vector; The identification and analysis module, based on comprehensive feature vectors, performs transfer learning model training and time-series correlation analysis to obtain concentration estimates and diagnostic conclusions.

Images