A method and system for monitoring drainage fluid using multi-frequency impedance tomography

Abstract

The application provides a drainage liquid multi-frequency impedance tomography monitoring method and system. The method applies electric excitation at multiple preset frequencies through an electrode array surrounding the outer wall of a drainage container and collects boundary voltage responses, obtains the multi-frequency conductivity spatial distribution after regularization constraint inversion, extracts impedance values at each frequency for the region of interest and fits a frequency dispersion model to obtain frequency dispersion characteristic parameters, constructs an impedance fingerprint vector based on the frequency dispersion characteristic parameters and inputs a pre-trained classification model for pattern classification, and outputs the drainage liquid classification. The scheme uses the relaxation characteristic differences of different component drainage liquids in the frequency dispersion model, uses the frequency dispersion dimension information density to compensate for the insufficient spatial resolution under the condition of low electrode density, and realizes sensitive, automatic and continuous monitoring of the component change of the drainage liquid.

Description

Technical Field

[0001] This application relates to the field of medical monitoring technology, specifically to a method and system for automatically monitoring the composition of drainage fluid using multi-frequency electrical impedance tomography and signal pattern recognition technology. Background Technology

[0002] In postoperative drainage monitoring, changes in the composition of drainage fluid are a key indicator for assessing a patient's recovery. Current techniques primarily rely on visual observation of the fluid's color and turbidity by healthcare professionals, or on periodic sample collection for laboratory analysis. Visual interpretation is limited by observer subjectivity and intermittent sampling, failing to provide continuous perception of gradual changes in composition. While single-frequency electrical impedance tomography (SETT) can acquire cross-sectional impedance distribution, the impedance differences between different drainage fluid components are limited at a single frequency, and its inherent pathological nature makes it difficult to provide reliable component differentiation under low electrode density conditions. Therefore, a technical solution is needed that can utilize the differences in impedance dispersion characteristics of drainage fluid at different frequencies to achieve automatic, continuous, and sensitive monitoring of compositional changes. Summary of the Invention

[0003] To address the shortcomings of existing single-frequency impedance tomography (ITT) techniques, which cannot effectively distinguish between different components of drainage fluid and lack automated component discrimination capabilities, this application provides a multi-frequency impedance tomography monitoring method and system for drainage fluid. This method obtains the spatial distribution of multi-frequency conductivity by applying multi-frequency electrical excitation to the drainage fluid, extracts dispersion feature parameters to construct an impedance fingerprint vector, and achieves automatic component discrimination through pattern classification. It can sensitively identify component changes by utilizing the differences in relaxation characteristics exhibited by different drainage fluid components in the dispersion model. Specifically, this application provides the following technical solutions: Firstly, this application provides a method for multi-frequency impedance tomography monitoring of drainage fluid, including: By using an electrode array surrounding the outer wall of the drainage container, electrical excitation is applied to the drainage fluid at multiple preset frequencies and the boundary voltage response is collected to obtain multi-frequency measurement data. Based on the difference between the multi-frequency measurement data and the reference data, the spatial distribution of multi-frequency conductivity of the drainage fluid cross section is obtained by regularized constraint inversion; For the region of interest in the multi-frequency conductivity spatial distribution, the impedance value corresponding to each frequency is extracted and a dispersion model is fitted to obtain the dispersion characteristic parameters. An impedance fingerprint vector is constructed based on the dispersion feature parameters. The impedance fingerprint vector is then input into a pre-trained classification model for pattern classification, and the category of drainage fluid composition is output.

[0004] Optionally, the dispersion model is a Cole-Cole impedance dispersion model, and the dispersion characteristic parameters include low-frequency limiting resistance, high-frequency limiting resistance, relaxation time constant, and dispersion coefficient.

[0005] Optionally, the impedance fingerprint vector includes the mean and standard deviation of the dispersion characteristic parameters in the region of interest, and a frequency ratio characteristic component constructed based on multiple preset impedance ratios.

[0006] Optionally, the electrode array includes multiple electrodes evenly distributed along the circumference of the drainage container, and multiple rounds of acquisition are completed sequentially using an adjacent excitation-adjacent measurement mode to obtain the multi-frequency measurement data.

[0007] Optionally, the regularization constraint inversion adopts a hybrid constraint of smooth regularization term and sparse gradient regularization term, and solves the spatial distribution of conductivity corresponding to each frequency by iteratively solving in alternating directions.

[0008] Optionally, before inputting the impedance fingerprint vector into the pre-trained classification model, the impedance fingerprint vector is further subjected to standardization processing; the pre-trained classification model is a support vector machine classifier based on a kernel function.

[0009] Optionally, the method further includes performing sliding window time-series smoothing on the classification results, calculating the cumulative deviation statistic on the time series of the impedance fingerprint vector to perform change point detection, and outputting a state change flag when the cumulative deviation statistic exceeds a preset decision threshold.

[0010] Optionally, before performing the acquisition boundary voltage response, a system calibration step is also included: performing full-band reference measurement on the drainage container filled with reference medium to obtain the reference data, and pre-calculating the sensitivity matrix corresponding to each frequency.

[0011] Secondly, this application provides a multi-frequency impedance tomography monitoring system for drainage fluid, comprising: The multi-frequency excitation acquisition module is used to apply electrical excitation to the drainage fluid at multiple preset frequencies through an electrode array surrounding the outer wall of the drainage container and acquire the boundary voltage response to obtain multi-frequency measurement data. The fault reconstruction module is used to obtain the spatial distribution of multi-frequency conductivity of the drainage fluid cross section through regularized constraint inversion based on the difference between the multi-frequency measurement data and the reference data. The dispersion analysis and fingerprint construction module is used to extract the impedance values ​​corresponding to each frequency in the region of interest in the multi-frequency conductivity spatial distribution and fit the dispersion model to obtain dispersion feature parameters, and construct an impedance fingerprint vector based on the dispersion feature parameters. The pattern classification module is used to input the impedance fingerprint vector into a pre-trained classification model for pattern classification, output the drainage fluid component category and classification confidence, and perform change point detection on the time series of the impedance fingerprint vector to output state change flags.

[0012] Optionally, it also includes an early warning output module, which is used to output multi-level early warning signals based on the component category, classification confidence level and state change indicator, combined with trend analysis results.

[0013] The beneficial effects of this application are as follows: By applying electrical excitation at multiple preset frequencies to the drainage fluid and collecting the boundary voltage response, the spatial distribution of multi-frequency conductivity is obtained through regularized constraint inversion. Then, a dispersion model is fitted to the region of interest to extract dispersion feature parameters and construct an impedance fingerprint vector. Finally, a pre-trained classification model is used to automatically output the component category. This scheme utilizes the differences in relaxation time constants and dispersion coefficients of different drainage fluid components in the dispersion model to compensate for the insufficient spatial resolution under low electrode density conditions with the information density of the dispersion dimension. It achieves sensitive differentiation of drainage fluid component changes without increasing the number of electrodes. Simultaneously, the frequency ratio feature contained in the impedance fingerprint vector has an intrinsic ability to suppress electrode contact impedance drift, ensuring that the monitoring results remain stable for electrode state changes during long-term use. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall process of the multi-frequency impedance tomography monitoring method for drainage fluid provided in the embodiments of this application.

[0015] Figure 2 This is a schematic diagram of the multi-frequency excitation and boundary voltage acquisition subprocess provided in the embodiments of this application.

[0016] Figure 3 This is a schematic diagram of the multi-frequency impedance tomography reconstruction subprocess provided in an embodiment of this application.

[0017] Figure 4 This is a schematic diagram of the sub-process for extracting Cole-Cole dispersion parameters and constructing fingerprint vectors provided in an embodiment of this application.

[0018] Figure 5 This is a schematic diagram of the sub-process for classifying and determining the composition pattern of the drainage fluid in an embodiment of this application.

[0019] Figure 6 A schematic diagram comparing the Cole-Cole dispersion curves of various drainage fluids provided in the embodiments of this application.

[0020] Figure 7 This is a schematic diagram of the timing response for CUSUM variable point detection provided in an embodiment of this application.

[0021] Figure 8This is a schematic diagram of the architecture of the drainage fluid multi-frequency impedance tomography monitoring system provided in the embodiments of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0024] This application provides a method and system for multi-frequency impedance tomography monitoring of drainage fluid. In a specific implementation, this method applies electrical excitation at multiple preset frequencies to an electrode array surrounding the outer wall of the drainage container, acquires boundary voltage responses to obtain multi-frequency measurement data, and then obtains the spatial distribution of multi-frequency conductivity through regularized constraint inversion. Subsequently, impedance values ​​at each frequency are extracted from the region of interest, and a dispersion model is fitted to obtain dispersion feature parameters. Based on the dispersion feature parameters, an impedance fingerprint vector is constructed, and pattern classification is performed using a pre-trained classification model to output the drainage fluid component category. This method solves the technical problem that single-frequency impedance tomography cannot effectively distinguish different components of drainage fluid and lacks automated discrimination capability in existing technologies. It achieves the technical effect of using the information density of the dispersion dimension to compensate for insufficient spatial resolution, thereby enabling sensitive identification of changes in drainage fluid composition.

[0025] like Figure 1 As shown, this method includes the following steps: S000 performs full-band reference measurements on the drainage container filled with reference medium to obtain reference data and pre-calculates the sensitivity matrix corresponding to each frequency.

[0026] like Figure 8 As shown, the multi-frequency impedance tomography monitoring system for drainage fluid in this embodiment includes a main control processor and a flexible electrode array arranged around the outer wall of the drainage container. The electrode array contains 16 Ag / AgCl electrodes, arranged circumferentially around the drainage container. The electrodes are evenly distributed at equal intervals, and the effective contact area of ​​each electrode is... The electrode substrate uses a flexible polyimide film to adapt to the curved outer wall of the drainage container. Ag / AgCl material was chosen because of its stable contact potential, low polarization resistance, and superior electrochemical performance in a physiological saline environment compared to stainless steel or platinum electrodes. The main control processor uses an ARM Cortex-M7 architecture with a main frequency of... Built-in hardware floating-point unit and DSP instruction set extension, configuration In-chip SRAM and On-chip Flash storage, connected via FMC bus External SDRAM is used to store large-scale matrix data.

[0027] After the system is powered on, it first performs electrode contact impedance detection to assess the physical adhesion quality between the electrodes and the outer wall of the drainage container. The main control processor controls the DDS signal source output frequency to be... The sinusoidal constant current excitation signal, the excitation current amplitude Select The basis for the detection frequency is the double-layer capacitance in the Randle equivalent circuit at the electrode-electrolyte interface at that frequency. The impedance has been sufficiently reduced, and the measurement results mainly reflect the solution resistance. and charge transfer resistance The system exhibits good sensitivity to electrode adhesion. Excitation is applied to each of the 16 electrodes individually, and the impedance amplitude between each electrode and the reference electrode is measured. The impedance amplitude of a normally bonded electrode should be within ~ Within the range.

[0028] After the test was completed, the coefficient of variation of the contact impedance of the 16 electrodes was calculated. ,in The standard deviation of the 16 impedance values, Its arithmetic mean. When This indicates insufficient electrode array consistency, suggesting poor adhesion or electrode damage. Further individual assessment is performed on each electrode: if... If the electrode is in a fault state, then the corresponding electrode state vector component is set to 0. Faulty electrodes will be automatically skipped in subsequent S100 acquisition phases. For example, within the cross-sectional radius of the drainage container... ,filling NaCl salt water ( , Under these conditions, the impedance values ​​from E1 to E15 in the 16-electrode measurement results are... ~ Fluctuation within a range, mean Standard deviation (After removing outliers) Overall consistency is acceptable. E16 electrode impedance. Deviation rate It was identified as a fault and marked. .

[0029] After the electrodes pass the inspection, full-band reference voltage acquisition is performed. The drainage container is filled with a reference medium with a conductivity of [missing information]. of NaCl physiological saline, at this concentration The conductivity has been calibrated and confirmed using a four-electrode conductivity meter. The system injects current sequentially using 16 adjacent electrode pairs as excitation pairs according to the Sheffield adjacent excitation protocol, at 8 preset frequencies. Complete the full cycle of data acquisition one by one. During each excitation cycle, current flows from the electrodes. Injection, from The current flows out, and the voltage difference between the remaining 13 adjacent electrode pairs is measured simultaneously. This is repeated for 16 rounds. Each independent measurement value. A reference voltage matrix is ​​obtained after acquiring all 8 frequencies. This matrix will serve as the subtraction reference for differential imaging in the subsequent S200 stage.

[0030] To verify the signal quality of the reference measurement, the same excitation mode was repeatedly sampled 10 times and the signal-to-noise ratio was calculated. Requires all channels .Should The threshold ensures that the measurement uncertainty of the reference voltage is less than [a certain value]. This will not introduce significant systematic errors into subsequent difference calculations. For example, for the first excitation pair in... The voltage difference between electrodes E3 and E4 was measured, and the average value was collected after 10 repeated measurements. Standard deviation , The quality is acceptable. If the SNR of a channel does not meet the standard, the system marks that channel as low quality and reduces its weight in subsequent inversion.

[0031] Perform Jacobian sensitivity matrix pre-calculation. Based on the circular cross-sectional geometry model (radius) of the drainage container. The cross-section is discretized using a first-order triangular finite element mesh. The mesh contains... One triangular unit and approximately Each node, unit area In uniform conductivity Under the condition of solving the Poisson equation based on the complete electrode model Boundary conditions include current injection conditions on the electrode surface. Zero flux condition at the non-electrode boundary The global stiffness matrix of the assembly after finite element discretization. The nodal potentials are then solved to extract the boundary electrode voltages. Sensitivity matrix. The The elements are calculated using the adjoint field method:

[0032] in For the first The excitation field distribution corresponding to the measurement number This represents the corresponding measurement field distribution. Each frequency corresponds to a Jacobian matrix. Size of raw data for a single-frequency matrix Bytes (single-precision floating-point format), 8 complete frequency matrix groups Bytes It resides in external SDRAM and is accessed on demand via the FMC bus. Since the geometry of the funnel container remains unchanged during use, the Jacobian matrix can be pre-calculated offline once during the manufacturing stage for long-term use, and is loaded directly from Flash to SDRAM upon power-up.

[0033] Finally, the system loads the pre-trained classifier model parameters from Flash storage. These model parameters include 80 support vectors from the RBF-SVM classifier. The corresponding Lagrange multipliers 10 bias terms for binary classifiers and the training set statistics required for standardization. and CUSUM baseline parameters for each category The total storage size of all model parameters is approximately [amount missing]. The data is loaded into the inference buffer of the on-chip SRAM. At this point, the S000 calibration phase is complete, and the system outputs a calibration completion flag and a fault electrode mask, entering a periodic monitoring loop.

[0034] The S100 uses an electrode array surrounding the outer wall of the drainage container to apply electrical excitation to the drainage fluid at multiple preset frequencies and collect boundary voltage responses to obtain multi-frequency measurement data.

[0035] like Figure 2 As shown, this stage involves the generation and injection of multi-frequency excitation signals and the acquisition and demodulation of boundary voltage responses.

[0036] The multiple preset frequencies are frequency values ​​that are equally spaced by eight logarithmic intervals. Frequency coverage range from arrive Step ratio The selection of this frequency range is based on biological tissues and suspensions. Dispersion characteristics: Within this frequency band, the cell membrane capacitance gradually transitions from a high-impedance barrier at low frequencies to a short-circuit path at high frequencies. The interfacial polarization effect (Maxwell-Wagner effect) of suspended particles (red blood cells, white blood cells, fat droplets) produces characteristic dispersion behavior. Characteristic relaxation frequencies of bloody drainage fluid. ~ Serous drainage fluid ~ Purulent drainage fluid ~ All frequencies fall within the frequency coverage area of ​​this scheme, ensuring that the dispersion transition zone can be fully sampled. The logarithmically evenly spaced distribution ensures that each frequency point is uniformly distributed on the logarithmic frequency axis, and the dispersion information increment between adjacent frequencies is approximately equal, avoiding the problems of over-dense high-frequency bands and under-sampling low-frequency bands caused by linear distribution.

[0037] The excitation current amplitude is set to The determination of this value takes into account both safety and signal-to-noise ratio constraints: the IEC 60601-1 medical device safety standard stipulates that the auxiliary current should not exceed [a certain value]. RMS, It leaves a safety margin of 2 times; on the other hand, in The measurement pair with the furthest distance between adjacent excitation modes at the lowest frequency is approximately [voltage value missing]. , The signal generated by the excitation is still higher than the quantization noise floor of the 16-bit ADC. This ensures an effective signal-to-noise ratio across all frequency bands and channels.

[0038] The data acquisition process employs time-division multiplexing to perform frequency scanning and electrode rotation. The DDS signal source (AD9833 level, frequency resolution) is used. The DDS outputs sinusoidal signals of various frequencies sequentially, which are then converted into a precision constant current output by a Howland constant current source circuit. Each time a new frequency is switched, the circuit waits for the DDS output to stabilize and for the downstream PLL to lock, which takes approximately [time period missing]. The 16:1 analog switch matrix controls the switching timing of the current injection electrode pairs according to the Sheffield protocol. wheel( Current from Injection, from The current flows out, while another set of switch matrices selects the measurement electrode pairs. Each round measures the voltage difference between 13 unexcited adjacent electrode pairs, which is then transmitted through an 8-channel instrumentation amplifier (INA128, gain...). After differential amplification, the signal is fed into a 16-bit SAR ADC (ADS8688). (Simultaneous sampling) is used for digitization.

[0039] The acquired raw voltage signal is subjected to lock-in amplification and synchronous demodulation to accurately extract the amplitude and phase information of each frequency component. The lock-in amplification utilizes a reference signal with the same frequency as the excitation signal to achieve extremely narrowband filtering, thus optimizing the measured signal. Respectively with reference sine and reference cosine After multiplication, the components are low-pass filtered to extract the in-phase components. and orthogonal components Restore signal amplitude and phase The low-pass filter employs a 4th-order Butterworth IIR structure with a cutoff frequency of... Ensure that adjacent frequency components (nearest frequency ratio is 1) are treated correctly. The inhibition level of ) exceeded ADC sampling rate (for the highest frequency) Achieving 10x oversampling), per frequency per round of sampling Sample points. ARM Cortex-M7 processor with The main frequency calls the IIR filtering and vector multiplication functions of the CMSIS-DSP library to complete the real-time demodulation operation.

[0040] For example, at the excitation frequency The voltage of electrode pairs E3-E4 was measured under the first excitation condition (E1-E2). The ADC was used... collection Point (coverage) Duration, including approximately 20 episodes The signal period is used to obtain the in-phase component through phase-locked demodulation. Orthogonal components Calculate the amplitude. phase The negative phase indicates that the drainage fluid impedance contains a weak capacitive component. Compared with the reference measurement obtained from S000, (Pure saline solution state), difference value A negative value indicates that the conductivity of the tested drainage fluid is lower than that of the reference saline in the direction of the channel (the effective conductivity of bloody drainage fluid is reduced due to the suspension of red blood cells). This difference value will be used as the input of the S200 stage inversion algorithm.

[0041] When an electrode is marked as faulty in the electrode state vector ( When an E16 fault occurs, the system automatically skips all excitation-measurement cycles that use that electrode as the injection or measurement electrode. Taking the E16 fault as an example, the excitation pairs involved in E16... and Skipped, involving measurements of E16 and Skipped in other excitation rounds, the actual number of effective measurements is reduced to One. Corresponding to. The skipped positions are marked with NaN, and in the subsequent S200 stage, the weights of these positions are set to zero in the noise weighting matrix.

[0042] Time budget verification of complete 8-frequency 16-round acquisition: Total time consumption of single-frequency single-round excitation including signal setup and stabilization and 16-bit ADC acquisition (Among them, the collection window) Point Occupation The signal setup time has been stabilized by the preceding frequency switching. (Coverage), single frequency full 16 rounds 8-frequency scanning Plus frequency switching DDS lock-in stabilization time Total collection time All demodulation results are assembled into a measurement voltage matrix. (Complex form includes amplitude and phase), data volume (Polar coordinates, double float32 format, each measurement point stores 8 bytes of amplitude and 8 bytes of phase, including a 64-byte frame header) (bytes / frame), plus acquisition timestamp It is then written to the shared SRAM buffer and passed to the S200 tomographic reconstruction module.

[0043] The selection criteria for the Sheffield adjacent excitation protocol and its impact on subsequent processing. Compared with other alternatives (such as diagonal excitation and adaptive excitation), the adjacent excitation mode has the following advantages: simplest hardware implementation, requiring only two adjacent switches to be turned on simultaneously, and the switch matrix control logic is a simple shift counting; shortest excitation current path, with minimal impact of electrode contact impedance on the injected current amplitude; and moderate spatial information redundancy of 208 independent measurements, ensuring sufficient inversion constraints without introducing excessive acquisition time overhead. Its limitation is that the sensitivity in the central region is approximately lower than that of the diagonal excitation mode. However, this scheme has already achieved additional information gain in the dispersion dimension through a multi-frequency strategy, and the moderate reduction in spatial resolution can be compensated for by the discriminative power of the dispersion dimension. Furthermore, the actual amount of independent information in the 208 measurements is limited by… The inherent degrees of freedom of the electrode system (approximately) (each independent parameter), through regularization constraints Compressing the dimensional solution space to an effective size The use of subspaces ensures the well-defined nature of the problem.

[0044] The conductivity of the drainage fluid is temperature-dependent. The temperature coefficient of NaCl solution is approximately In practical use, the drainage fluid is drawn from inside the body ( After flowing into the external drainage container, the temperature gradually decreases. Without compensation, this will introduce a systematic conductivity deviation. This method naturally eliminates the effect of slow temperature drift through a differential imaging strategy: [reference measurement...] The first execution after system installation, subsequent differential operations. It only reflects the change of the drainage fluid itself relative to the reference state; for the extraction of dispersion parameters in the S300 stage, the frequency ratio feature The effect of temperature on impedance at various frequencies is approximately equal to the common mode (because...) Temperature coefficient at ~ (Approximate frequency independent within the frequency band), ratio calculations automatically eliminate temperature drift. For absolute parameters... and Slow drift, relaxation band of CUSUM algorithm Sufficient tolerance space has been reserved. The norm change caused by temperature fluctuations is approximately much smaller The width of the relaxation band.

[0045] To prevent bit-flipping errors or DMA transfer errors during ADC sampling, an integrity check is performed on the ADC acquisition data for each frequency and each round after it is written to the SRAM buffer. If the check fails (probably...), ( / frame), the system discards the data from that round and in the next... The frequency is re-acquired for that round in a cycle. A single round of re-acquisition takes only [time missing]. This will not affect the overall time budget.

[0046] S200, based on the difference between the multi-frequency measurement data and the reference data, the spatial distribution of multi-frequency conductivity of the drainage fluid cross section is obtained through regularized constraint inversion.

[0047] like Figure 3 As shown, this stage inverts the conductivity variation distribution inside the drainage fluid cross section from the boundary voltage differential data.

[0048] The difference is calculated element-wise. Obtain the differential voltage matrix The physical significance of the differential imaging strategy lies in the fact that, after subtracting the reference voltage, the differential value only reflects the change in conductivity of the drainage fluid relative to the reference saline solution, eliminating the influence of common-mode system errors such as electrode contact impedance, lead resistance, and container wall thickness. For the measurement channels associated with electrodes whose electrode state vectors are marked as faulty in the S000 stage (28 channels are involved in the E16 fault), their corresponding... Positions are preserved with NaN labels. A diagonal noise-weighted matrix is ​​constructed based on statistical data from 10 repeated measurements of the S000 reference. , ,in For the first The noise standard deviation of measurement channel 1 (estimated from the standard deviation of 10 repeated measurements). The faulty channel weight is set to... The weighting strategy automatically masks the influence of low signal-to-noise ratio (SNR) channels during inversion without reducing matrix dimensions. This strategy ensures that channels with high SNR contribute more to the inversion results, suppressing the noise amplification effect introduced by low SNR channels.

[0049] For each of the 8 frequencies Solve the following hybrid regularization optimization problems independently:

[0050] in For the first The differential voltage column vector of frequencies, The Jacobian sensitivity matrix pre-calculated for S000. Let be the vector of cross-sectional conductivity variation to be solved, where each component represents a triangular mesh element relative to the reference medium. The conductivity deviation.

[0051] The objective function consists of three terms: the first term is a weighted data fidelity term, which minimizes the change in the boundary voltage predicted by the model. The measured differential voltage The weighted residuals between the two terms ensure consistency between the inversion results and the measured data. The second term is the Tikhonov smoothing regularization term, a matrix... This is a Laplacian smoothing matrix constructed based on the adjacency relationship of a triangular mesh. diagonal elements equals unit The number of adjacent units, off-diagonal elements If unit and Shared edge, otherwise zero, penalizing abrupt changes in conductivity between adjacent cells to suppress high-frequency spatial oscillations caused by noise. The third term is the total variation sparse gradient regularization term, the discrete gradient operator. ( The number of sides inside the triangular mesh is approximately For each inner edge, calculate the difference between the elements on both sides, i.e. The Components , No. Components ( , For the first (The unit numbers on both sides of the inner edge of the strip), and the remaining components are zero. Norms encourage piecewise constant solutions, which means that a few sharp spatial abrupt boundary conditions exist in the solution. This is suitable for real-world scenarios where there are component stratification interfaces or blood clots in the drainage fluid.

[0052] The two complementary regularization mechanisms are: the Tikhonov term provides smooth constraints in uniform regions to prevent noise amplification, and the TV term allows discontinuous transitions at the boundaries to preserve spatial features. The combination of these two regularization mechanisms simultaneously addresses both noise reduction and boundary preservation.

[0053] Regularization parameters and Offline calibration was performed using the L-curve method. On a set of physical phantoms containing five known conductivity distributions (single-circular inclusion body, double-circular inclusion body, semi-circular layered body, ring-shaped distribution body, and random dispersion body), calibration was performed. and A total of 30 combinations are solved one by one to reconstruct the relative error. and structural similarity As a dual indicator, in - Select the parameter combination corresponding to the point of maximum curvature on the double logarithmic curve. Offline calibration results are displayed. , Average on 5 types of phantoms and The purer Tikhonov scheme ( hour , ) and pure TV solution ( hour , However, improvements have been made in all aspects (including the presence of staircase artifacts).

[0054] Due to TV items Since the norm is not differentiable, the split Bregman iterative algorithm is used to decouple the original problem into two alternately solved subproblems. An auxiliary variable is introduced. ( The number of edges inside the grid is approximately (strips) and Bregman parameters Augmented Lagrange parameters Each iteration contains: Subproblem 1 (pure) Minimize): Solve the normal equation Key implementation optimization: Left-side system matrix The system is a symmetric positive definite matrix that remains invariant during iterations. Before the first iteration, the system... Perform a Cholesky decomposition (computational workload) FLOP), each subsequent iteration only requires updating the right-hand vector and performing two trigonometric calculations (previous generation - back generation) (approximately 1000 FLOPs in total). (FLOP), improving computational efficiency by nearly 100 times. Cholesky factor use and The sparsity requires only about [amount] storage. Approximately 1 non-zero elements (approximately 100 elements) ), residing in the on-chip SRAM.

[0055] Subproblem 2 (element-by-element soft thresholding): where the soft threshold function Perform the operation independently for each edge, with a threshold. This operation forces edges with gradient magnitudes below a threshold to zero (facilitating the piecewise constant), retains edges with gradient magnitudes above the threshold and reduces their magnitudes (preserving the true boundaries).

[0056] Bregman parameter update: This update step accumulates constraint violations, driving subsequent iterations to satisfy them more strictly. Constraints. The convergence criterion is the relative change of parameters. The maximum number of iterations is 15. Experience shows that convergence is usually achieved in 8 to 12 iterations.

[0057] After the inversion is completed, physical constraints are applied to the results for post-processing: if the total conductivity of a certain element after reconstruction is... (Conductivity cannot be negative), then the cutoff is... , It represents a numerically stable, minimum positive value. The cutoff rate is typically lower than [a certain value]. The artifact appears in the edge artifact region directly below the electrode. Multi-frequency conductivity distribution is obtained after independent frequency inversion. That is, 8 frames of tomographic images, each with a spatial resolution approximately equal to the diameter of the container. (about This satisfies the need to distinguish different component regions within the cross-section.

[0058] For example, with bloody drainage fluid (conductivity) Approximately 1 / 3 of the cross section The scene is defined as an area of ​​128 grid cells located in the lower fan-shaped region of the cross-section. At certain frequencies, the differential voltage of the measurement channel near the bloody region is larger, especially with the E8-E9 electrode pair (directly facing the bloody region). The channel difference is smaller in areas far from the bloody region, E12-E13 electrode pair (facing away from the bloody region) The spatial difference is approximately 10 times. It converges after 12 splitting Bregman iterations (the 12th relative change). The mean change in reconstructed conductivity of 128 units in the hemorrhagic region (Total conductivity) Consistent with blood conductivity), 384 units in the non-hemorrhagic region. (Correct change is zero), the boundary transition zone is approximately 12 units wide, and the reconstruction relative error is... .

[0059] A dispersion effect can be observed by comparing the results at 8 frequencies: hour (At low frequencies, the red blood cell membrane blocks intracellular conductivity, resulting in the lowest effective conductivity.) hour (At high frequencies, the membrane short-circuits, and the intracellular fluid participates in conduction, increasing the effective conductivity), resulting in changes in conductivity deviation across frequency bands. This is blood. The dispersion phenomenon forms the physical basis for the subsequent S300 stage dispersion model fitting.

[0060] To further improve the reconstruction quality, this application introduces multi-frequency collaborative consistency verification based on independent inversion at 8 frequencies. Since the conductivity distribution of the same drainage fluid cross-section has spatial structural consistency at 8 frequencies, that is, the spatial position of the ROI remains unchanged at each frequency (the same blood clot will not change its position due to different excitation frequencies), only the conductivity amplitude changes with frequency. This prior knowledge is used to perform post-processing verification of the inversion results.

[0061] Specifically, the cross-frequency correlation coefficient of each pixel in the 8 frames of images is calculated. If the conductivity of a pixel at frequency 8 shows a monotonic trend (positively or negatively correlated with frequency), If the change is irregular, it is determined to be a true component region; If the pixel is identified as a noise artifact, then all 8 frame values ​​of that pixel will be attenuated. This is to suppress its interference with subsequent dispersion fitting. This verification does not change the ROI segmentation result; it only applies a confidence weight to the cell-level conductivity values. For example, the center pixel of the blood region... At 8 frequencies The sequence is , with frequency index correlation coefficient (Strong positive correlation, absolute value of conductivity deviation monotonically decreases with frequency), identified as a high-confidence real region. Pixels in edge noise regions. The sequence is , (Irregular fluctuations) are identified as noise artifacts and are subject to attenuation.

[0062] The convergence and numerical stability of the inversion algorithm are guaranteed by the following mechanism: the prerequisite for Cholesky decomposition is the system matrix. Zhengding, due to Only positive semidefinite (because) lead to (Insufficient column rank), regularization term The addition ensured Positive definiteness, The smallest eigenvalue is approximately , The smallest eigenvalue is approximately ,make The condition number is approximately It is far superior to the case without regularization. condition number (Severely pathological). Convergence rate of split Bregman iteration versus augmented Lagrangian parameters. Selection related: Overly strong regularization in subproblem 1 leads to slow convergence. When the auxiliary variable is too large The relaxation of constraints leads to a decrease in TV effect. Experiments show that... A balance was struck between convergence speed and TV retention capability, with an average number of convergence iterations on the five phantom types. Second-rate.

[0063] Computational cost analysis of single-frequency inversion: Cholesky decomposes it into a one-time preprocessing step (approximately...) FLOP, ARM Cortex-M7 Peak demand ), the previous generation back generation in each iteration FLOP plus soft threshold FLOP total 15 iterations Including Cholesky decomposition and matrix multiplication, the total time for a single frequency is approximately... However, because the 8 frequencies share the same... and Matrix, system matrix China only As the frequency changes, Cholesky decomposition can be performed on the eight system matrices in advance, and the factors can be cached. (allocated in SDRAM), requiring only previous generation back-generation during runtime. Optimized total time for 8-frequency serial inversion: Add matrix-vector multiplication Approximately Considering FMC bus access latency and DMA transfer overhead, the actual time consumption is approximately... This satisfies the design constraints.

[0064] Frequency dependence analysis of inversion accuracy: due to the Jacobian matrix The elemental amplitude is related to the excitation frequency (at high frequencies, the increased current penetration depth improves the sensitivity in the central region), and the reconstruction quality varies at different frequencies. In the central region of a circular cross-section, low frequencies... The relative error of the reconstruction is approximately ,high frequency The error was reduced to approximately At high frequencies, the central region exhibits higher sensitivity and better inversion accuracy. This frequency-dependent accuracy distribution is naturally utilized by the LM fitting in the S300 stage, with the fitting residuals automatically assigning greater weight to high-precision frequency points, making the dispersion parameter estimation more reliable. The situation is reversed in the edge region (close to the electrode): the accuracy at each frequency is high and the differences are small. ~ Therefore, if the ROI is located at the edge of the cross section, the reconstruction accuracy of each frequency is almost uniform.

[0065] S300, for the region of interest in the multi-frequency conductivity spatial distribution, extract the impedance value corresponding to each frequency and fit the dispersion model to obtain the dispersion characteristic parameters.

[0066] like Figure 4 As shown, this stage extracts dispersion parameters from the multi-frequency tomographic reconstruction results and constructs an impedance fingerprint vector for subsequent classification.

[0067] Automatic region-of-interest segmentation is performed on 8 frames of tomographic images. The mean plot of the absolute values ​​of conductivity changes at 8 frequencies is calculated. , The grid cell index is used. The mean plot integrates the variation amplitude information of all frequencies, and has an averaging and suppression effect on single-frequency noise. The Otsu adaptive thresholding method is used to determine the segmentation threshold for the mean plot. This method treats the gray-level histogram of the mean map as a superposition of two types of distributions (variable and invariant regions), and iterates through all possible binarization thresholds. Choose to make the inter-class variance The maximized value is taken as the optimal split point, where , Let be the prior probabilities of the two classes. , The mean is the sum of the two classes. The mean plot shows values ​​higher than... The cells are marked as regions of interest: like Otherwise, it is 0. The reason for choosing the Otsu method instead of a fixed threshold is that the conductivity deviation of different drainage fluid components varies greatly (the deviation of blood is approximately...). The slurry properties deviation is approximately A fixed threshold cannot adapt to different scenarios.

[0068] Two morphological post-processing steps are performed after segmentation to ensure ROI quality: area verification requirements. Each unit, meaning the area of ​​the region of interest is not less than the total area of ​​the cross-section. If the value is below this, it indicates that the component variation is too weak or noise dominates, and the system marks the frame as low confidence rather than forcibly extracting potentially unreliable parameters; connectivity verification requires that the ROI must contain at least one area. Connected regions of a unit, isolated scattered units (area) (Noise is excluded.) For example, the mean plot in the bloody drainage scenario shows a bimodal distribution, with the mean value of the bloody area being approximately... (The peak value is obvious), and the area of ​​no change is approximately (Noise level), Otsu algorithm output threshold The region was segmented into 128 ROI units, forming a connected region, with an area percentage of [missing information]. It passed two checks.

[0069] For each pixel within the region of interest ( ), extracting its impedance value sequence across 8 frequencies. Since the output of electrical impedance tomography reconstruction is volume conductivity. (unit This needs to be converted to impedance dimensions compatible with the Cole-Cole dispersion model. This method uses normalized impedance transformation: ,in The reference dielectric conductivity is used. The physical meaning of this conversion is: each unit is considered as a virtual cylinder with a cross-sectional area of ​​1 and a length of 1; its impedance is the resistivity (the reciprocal of conductivity), with dimensions of... This normalization transformation ensures that the impedance values ​​of all pixels are on the same order of magnitude, avoiding interference from absolute voltage amplitude differences caused by electrode distance variations in the dispersion curve shape. The dispersion curve for each ROI pixel is obtained after the transformation. excitation frequency The horizontal axis represents the normalized impedance value (logarithmic scale). This is an 8-point data sequence on the vertical axis.

[0070] A Cole-Cole impedance dispersion model is fitted to the dispersion curve of each ROI pixel. The mathematical expression of the Cole-Cole model is:

[0071] The model contains four parameters to be fitted, each with a clear physical correspondence: The low-frequency limiting resistance (the high impedance value corresponding to the conduction of only extracellular fluid when the frequency approaches zero, reflecting the volume concentration of suspended particles and the extracellular ion intensity). It is the high-frequency limiting resistance (the low impedance value corresponding to the conduction of both intracellular and extracellular fluids when the frequency approaches infinity, reflecting the total ion content). The characteristic relaxation time constant (corresponding to the RC charge-discharge time of the cell membrane, reflecting the particle size and membrane capacitance of suspended particles) ), The dispersion coefficient ( , For an ideal single-relaxation Debye model corresponding to a uniform particle size suspension, The broadening of the relaxation time distribution corresponds to multi-particle-size or non-uniform suspensions. The differences in these four parameters observed in drainage fluids of different compositions are as follows: For bloody drainage fluids, the difference is due to the diameter of red blood cells... ~ And its unique biconcave disc-shaped film structure produces ~ , ~ Serous drainage fluid is produced due to the extremely weak polarization of protein molecules (nanoscale). ~ , ~ Purulent drainage fluid has a high white blood cell diameter. ~ Mixed with cell debris to produce ~ , ~ .

[0072] Since the EIT reconstruction output in this method is a real conductivity (phase-locked demodulation only takes the amplitude), and the converted impedance is also a real value, the fitting is performed in the real domain. The model's real part predictions for 8 frequencies are compared with the measured values ​​to construct a residual vector. parameter vector The Levenberg-Marquardt nonlinear least squares algorithm is used to minimize... The algorithm's iterative update rule is as follows: Jacobian matrix By calculating the analytical partial derivatives of the Cole-Cole model, , , , Since this method fits in the real domain, the partial derivatives mentioned above... Item with and These correspond to the real and imaginary parts, respectively; however, only the real part is considered in actual calculations. Damping factor Adaptive adjustment: When the residual decreases Accelerated convergence, when residuals increase Enhance stability, initial value Parameter initialization strategy: (Lowest frequency impedance) (Highest frequency impedance) in The frequency corresponding to the point of fastest impedance change (by calculation) (Maximum value location) Parameter constraints , , This is achieved through parameterization using logarithmic transformation, letting , , This allows optimization to be performed in an unconstrained space. The convergence criterion is... or residual change Maximum of 50 iterations.

[0073] The quality of the fit is evaluated using the coefficient of determination. , The average impedance at 8 frequencies. If the average of all pixels in the ROI... The current frame is then marked as low confidence, and the S400 classifier subsequently reduces the weight of that frame (confidence decay factor). This strategy ensures that classification decisions are not misled by the degradation of fit quality.

[0074] For example, the center pixel of the ROI in a bloody drainage fluid scene Its 8-frequency impedance value is (Exhibiting a typical dispersion decreasing pattern from high to low). Initial value , , The maximum cross-frequency band difference corresponds to Therefore , After 12 LM iterations, the result converged, and the residual... From the beginning Down to (Reduced by 3 orders of magnitude). Fitting results: , , (characteristic frequency) , falling Within the dispersion interval) (Characterizing the broadening of erythrocyte particle size distribution). Fit quality Far exceeding Threshold.

[0075] After fitting all 128 ROI pixels, spatial statistical aggregation was performed. The ROI spatial mean and standard deviation for each parameter are as follows: ( ), ( ), ( ), ( The standard deviation reflects the spatial homogeneity of parameters within the ROI. Bloody drainage fluid has a smaller standard deviation (relatively uniform red blood cell distribution), while mixed drainage fluid has a larger standard deviation (spatial separation of multiple components). Four additional frequency ratio characteristics were calculated, using the average impedance of the ROI as the basis. Based on: Low frequency to high frequency ratio (Reflects total dispersion amplitude, blood volume approximately) And slurry properties ), intermediate frequency span ratio (Slope of the curve near the relaxation frequency), normalized dispersion amplitude (Normalized dispersion to eliminate differences in absolute magnitude), maximum dispersion gradient normalization (Relative gradient of the region with the fastest frequency dispersion). These four frequency ratio characteristics are calculated in the form of ratios, so that the drift of electrode contact impedance is automatically eliminated as a common-mode factor by the synchronous scaling of the numerator and denominator. Even if the electrode contact state deteriorates during long-term use, causing a systematic increase in the impedance of the entire channel, the ratio characteristics remain stable.

[0076] Finally, the above 12-dimensional features are assembled into an impedance fingerprint vector. .like Figure 6 As shown, the fingerprint vectors of different components of drainage fluid form distinguishable clusters in 12-dimensional space: bloody... , , ; slurry , , Purulent , , ; chylous , , The relaxation time constant showed a maximum cross-category difference of up to 12 times (serous vs. purulent), and the dispersion coefficient showed a significant difference. The difference between low and high frequency ratios reached All of them are more than three times the standard deviation within their respective classes, which lays the physical foundation for the high accuracy of subsequent SVM classifiers.

[0077] The convergence and robustness of the Cole-Cole model fitting are guaranteed through the following strategy. The sensitivity of the LM algorithm to initial values ​​is an inherent challenge in nonlinear fitting; if the initial value deviates too far from the true value, it may converge to a local minimum. Our initial value strategy minimizes this risk: The lowest frequency measured impedance is directly used as the low-frequency limit estimate, and the relative error of this estimate usually does not exceed [a certain value]. (because (It is already close to the low-frequency limit region). The measured impedance at the highest frequency has a relative error of approximately [missing information]. ( Although it has not fully reached the high-frequency limit, it is close to the plateau region. By locating the point where the impedance decreases the fastest, i.e. The maximum value corresponding to and geometric mean frequency This yields an estimate of the relaxation time, based on the physical principle that the inflection point (maximum slope point) of the Cole-Cole model occurs precisely at... To verify that the fit did not get trapped in a local minimum, each pixel was initialized with three different initial values ​​(standard initial value, ...). Double the initial offset value Perform LM fitting independently using the initial values ​​(multiplied by 1), and take the result with the smallest residual as the final output. If the difference between the optimal residuals of the three initial values ​​does not exceed 1 / 3... If the difference exceeds a certain threshold, it is considered globally optimal and reliable; otherwise, it is considered unreliable. If so, then a low confidence flag is set for that pixel. In actual operation, The three initial values ​​for the above pixels converge to the same result (residual difference). Only noise pixels located at the edge of the ROI occasionally exhibit initial value sensitivity.

[0078] Physical validity verification of dispersion parameters. After fitting, physical range verification is performed on the four parameters to eliminate ill-fitting results: Verification is (corresponding conductivity) ~ (within the reasonable range of biological parameters) Verification is (High-frequency impedance must be less than low-frequency impedance). Verification is (corresponding characteristic frequency) ~ (Covering all biological suspension relaxation frequencies) Verification is ( Physically, this corresponds to an extremely broadened relaxation distribution, which is almost impossible to occur in drainage fluid. Pixels that fail any of the validations are excluded from the ROI statistical aggregation. For example, in a bloody scene, 126 out of 128 ROI pixels passed all validations, while 2 edge pixels failed due to noise interference. (lower than) (Lower limit), excluded, exclusion rate It has no substantial impact on statistical aggregation.

[0079] Information complementarity analysis of 12-dimensional fingerprint vector features. Mean features of the first 4 dimensions. This reflects the group-level dispersion characteristics of the ROI, i.e., the macroscopic electrical characteristics of the drainage fluid as a whole suspension. Intermediate 4-dimensional standard deviation characteristics. This information carries details about the spatial heterogeneity within the ROI. Homogeneous components (e.g., pure blood) have a small standard deviation, while mixed components (e.g., coexistence of blood clots and serous fluid) have a large standard deviation. This information is particularly crucial for distinguishing between mixed drainage fluid (Category 5) and pure drainage fluid. (Later 4-dimensional frequency ratio characteristics) It provides a relative measure of intrinsic robustness to systematic measurement drift when electrode adhesion degradation leads to a systematic increase in the total channel impedance. hour, .because For frequency-independent contact resistance, the change in ratio For direct use only Time drift effect of For example, bloody drainage fluid. The drift effect suppressed .

[0080] Regarding the impact of ROI segmentation on the accuracy of subsequent parameter extraction: If the Otsu threshold is too low, causing the ROI to include some pixels in the invariant region, these pixels... This will lead to Since the frequency is constant, the dispersion curve is flat, and the Cole-Cole fitting will yield... (Near-zero dispersion amplitude) and uncertainties These outlier fitting results are obtained through the aforementioned The verification was automatically rejected because A zero dispersion amplitude indicates that the feature is not a valid characteristic of any type of drainage fluid. Conversely, if the Otsu threshold is too high, causing some edge pixels of the ROI to be missed, the missed pixels, which are precisely the pixels with lower precision, may actually improve the accuracy of statistical aggregation because the reconstruction accuracy of the central region of the ROI is higher than that of the edge region. Neither of these two deviation directions will cause the classifier to fail systematically.

[0081] Computational complexity of fingerprint vectors: Otsu threshold scan about Comparison operations; Normalization transformation of pixels about Division by division; The LM fitting of each pixel is the computational bottleneck, with an average of [number] pixels per pixel. The iteration, each iteration contains Calculation of secondary partial derivatives and Matrix inversion (approximately) FLOP), total FLOP; Statistical aggregation and ratio calculation FLOP. ARM Cortex-M7 with Peak execution, S300 total time approximately In practice, considering memory access latency and pipeline efficiency (approximately...), Peak utilization), actual measured time approximately .

[0082] S400, construct an impedance fingerprint vector based on the dispersion feature parameters, input the impedance fingerprint vector into a pre-trained classification model for pattern classification, and output the drainage fluid component category.

[0083] like Figure 5 As shown, this stage converts the impedance fingerprint vector into drainage fluid component category labels and detects state abrupt changes.

[0084] The impedance fingerprint vector z-score normalization is performed before inputting the data into the classification model. The transformation is performed using statistics from the offline training set loaded from Flash in the S000 stage. ,in and These are the mean vector and standard deviation vector for each dimension of the training set. The training set contains 200 fingerprints of 5 types of drainage fluid samples (40 samples of each type, collected from drainage fluid samples of different patients at different stages). This represents element-wise division. Standardization eliminates the differences in physical dimensions among the 12-dimensional features: (unstandardized) Quantity is , Quantity is , Dimensionless , Dimensionless If directly fed into the classifier, the Euclidean distance of the RBF kernel function will be... and The magnitude difference dominates, leading to The discriminative information of the ratio feature is obscured. After standardization, the mean of each dimension is 0 and the variance is 1, and the 12 features contribute evenly to the classifier.

[0085] The pre-trained classification model is a multi-class support vector machine classifier based on radial basis function kernels. The five drainage fluid component categories are: serous (Category 1, normal clear postoperative drainage), bloody (Category 2, containing red blood cells), purulent (Category 3, containing white blood cells and necrotic tissue), chylous (Category 4, containing lymphatic fluid and fat particles), and mixed (Category 5, a mixture of two or more components). A one-to-one multi-class strategy is used, with each of the five categories combined in pairs to generate... There are 1 binary classifier, each binary classifier Only categories are used in the training set. and Training on the samples. For the standardized fingerprint vectors... Each binary classifier calculates a decision value:

[0086] Kernel function Radial basis functions, kernel width parameter (Based on feature number) The heuristic setting of the total variance of the training set (the total variance after standardization is 1), and the regularization parameter Candidate values ​​were searched using a 5-fold hierarchical cross-validation grid search. , A total of 20 combinations were determined, and the average accuracy of the optimal combination using 5-fold cross-validation was [value missing]. The decision results of the 10 classifiers are aggregated through a voting mechanism: for each binary classifier... ,like Then category Get 1 vote, otherwise category Received 1 vote, final original category The model contains approximately 80 support vectors (distributed across 10 sub-classifiers), with storage requirements... Bytes (float32 format), plus and each bytes, totaling approximately The computational cost per inference run is approximately 80 kernel function calculations, each involving 12-dimensional Euclidean distance and 1 exponential operation. FLOP, plus 10 cumulative votes and approximately Total FLOPs, ARM Cortex-M7 single-precision floating-point peak The time for inference is much lower than .

[0087] Classification confidence scores are mapped to calibrated posterior probabilities using Platt scaling. For the class receiving the highest number of votes... Select the category to participate The mean decision value of each binary classifier Substitute into the sigmoid function ,parameter and These are the sigmoid parameters fitted offline using 5-fold cross-validation on the training set. The larger the absolute value, the steeper the decision boundary for that category. Final confidence level Range of values (Because the winning class has already received the highest number of votes, its posterior probability must be...) ).

[0088] To suppress class jumps caused by single-frame classification noise, the most recent The classification results of the frames are subjected to sliding window majority voting: When multiple categories have the same number of votes, the category with the highest confidence weight is selected. The state transition confirmation mechanism introduces anti-jitter logic: when... Compared with the current confirmed category At different times, start the switching wait counter; only when 3 consecutive frames ( The switch is confirmed only when all voting results in the window belong to the same new category. This design filters out occasional single-frame or double-frame misclassifications (such as transient interference when a bubble passes through an ROI), effectively preventing false alarms.

[0089] In parallel, the L2 norm time series of the standardized fingerprint vectors are analyzed. Perform CUSUM cumulative sum and change point detection. The L2 norm of the standardized fingerprint vector is chosen as the basis for one-dimensional monitoring statistics: z-score standardization maps all dimensions to the same order of magnitude, giving the L2 norm a mathematically balanced metric meaning; the L2 norm integrates the joint shift of 12-dimensional features, and any simultaneous shift in one or more dimensions will lead to an increase in the norm, making it more sensitive than monitoring a single dimension; furthermore, the L2 norm is a scalar, avoiding the computational complexity of multi-dimensional CUSUM. Upward cumulative sum. Detecting the increasing trend of fingerprint norm, downlink cumulative sum The detection norm shows a decreasing trend. When or Output status change flag Parameter meanings: For the currently confirmed steady-state category The mean of the standardized fingerprint norm in the training set (5 classes corresponding to 5 sets of parameters). To allow for bias, a relaxation band of half a standard deviation is used to filter out normal statistical fluctuations and avoid oversensitivity. The decision threshold is determined according to the Siegmund approximation formula. (where the normalized threshold) Normalized relaxation band ), calculated Frame (approximately) (One false alarm per minute), acceptable in a clinical setting; regarding detection sensitivity, when the true mean deviates... Average detection delay frame( This meets the need for timely early warning of changes in composition.

[0090] when When triggered, the system checks the voting results of the sliding window: if If the switching confirmation conditions are met, the baseline parameter switching will be performed, updating the CUSUM reference parameter to the new category statistic. , , , Reset cumulative amount Update confirmation record .like However, if the sliding window has not yet confirmed the switch (e.g., only 2 frames of offset), the CUSUM accumulation will continue to accumulate without being zeroed. This design captures the weak but continuous offset signal in the early stage of the gradual component change.

[0091] For example, such as Figure 7 As shown, this simulates a scenario where the drainage fluid abruptly changes from serous to bloody. The initial steady state is serous (Category 1), and the CUSUM parameter is loaded. , , , Frame 1 shows normal slurry properties. , (No accumulation within the relaxation band). Frame 2 is also normal. , Frame 3 marks the start of the transition (a sudden change in composition occurs). , Frame 4 shows enhanced blood characteristics. , Frame 5 deviates further. , Frame 6 confirms the mutation. , ,trigger At this point, the classification results of the last 5 frames in the sliding window are... majority vote (Bloodlust), a new category for 4 consecutive frames. Confirm the switch. The system outputs the final result: ingredient category. Confidence level , Detection delay of 4 frames ( The CUSUM baseline was then switched to blood-type parameters. , And clear it to zero.

[0092] The generalization performance of the classifier was confirmed through the following offline validation experiments. The training set consisted of 200 samples (40 samples from each of the 5 classes) collected from drainage fluid samples at different time points from 10 postoperative patients, covering serous (clear drainage fluid 48-72 hours postoperatively), bloody (active bleeding period 6-24 hours postoperatively), purulent (drainage fluid from postoperative infected cases), chylous (chylous leakage cases after thoracic surgery), and mixed (transitional samples with two or more components). Leave-One-Patient-Out (LOPO) cross-validation results: Overall classification accuracy... The most easily confused pair in the confusion matrix is ​​bloodthirstiness and mixedness. The cross-classification error rate was the highest, followed by serous and chylous types. (Mutual classification error rate). The highest accuracy rate in distinguishing between blood and serous fluid was achieved. The relaxation time constants of these two types differ by a factor of 6 ( vs In the 12-dimensional feature space, the inter-class distance is much greater than the intra-class variance.

[0093] In a 12-dimensional feature space, RBF-SVM constructs nonlinear decision boundaries for five classes. Visualization by performing PCA on the training set to reduce the dimension to 2 reveals: the first principal component (interpretation) Variance is mainly composed of and Linear combination constitutes the second principal component (explanation) Variance is mainly composed of and Composition. Serous and purulent types showed the greatest separation along the PC1 axis (PC1 coordinate difference approximately...). The separation between hemorrhagic and chylous lesions is greatest on the PC2 axis (PC2 coordinate difference approximately...). The mixed class is located in the middle region of the other four classes in the PCA space, and it has the largest number of support vectors (approximately 30, accounting for the majority of the total). This reflects the complexity of the decision boundary for this category. This analysis verifies the rationality of the physical design of the fingerprint vector. and The two features carry the most important class distinguishing information, which is consistent with the physical meaning of the Cole-Cole model (the difference in relaxation time constant of different suspended particles is the physical basis for distinguishing signals).

[0094] Unlike the acute mutations in the S400 example, gradual changes are more common in clinical practice, such as postoperative infections causing drainage fluid to gradually transition from serous to purulent. It increases slowly over several hours. (Simulation) by A gradually changing scene with a linearly increasing rate: Initial (Typical values ​​of slurry properties), after minutes reached Entering the vicinity of the category boundary. The CUSUM accumulation process is as follows, before... Single frame offset within minutes (300 frames) Below the relaxation zone , Almost no accumulation; the first ~ Minute-normalized norm frame-by-frame offset Beyond the relaxation zone, Start linear accumulation; in the first... minutes achieve trigger Compared to the fixed threshold method (when...) (Time alarm), single-frame offset in a gradual scene always does not exceed Therefore, the fixed threshold method can never detect it, while CUSUM achieves sensitive detection of gradual changes by accumulating small deviations, which is the core advantage of choosing CUSUM instead of a simple threshold.

[0095] Because the mean and variance of the fingerprint norms of the five types of drainage fluids are different (serous fluids) , Blood and guts , Purulent , ; chylous , Mixed , During baseline switching, all CUSUM parameters must be updated synchronously to avoid false alarms immediately after the switch. The sequence of operations executed at the moment of switching is: pause CUSUM accumulation, load the new category parameter group, and reset to zero. and Waiting for 3 frames ( Establish a new steady-state reference and restore CUSUM accumulation. The waiting window ensures that the transient signal attenuates during switching, and that the EWMA has not yet fully tracked the new mean at the moment of switching, thus allowing for the waiting period. The residual between the EWMA value and the new mean decreased to The initial residual level is then accumulated for 3 frames to confirm the new steady state.

[0096] When an outlier appears in the fingerprint vector output by the S300 (e.g.) Caused by numerical overflow, or (Exceeding physical constraints), the S400 module performs input validation before standardization: checking each dimension of the 12-dimensional vector to see if it falls within the observation range of the training set. Within the interval. Dimensions outside this interval are truncated to the boundary value and marked as an anomaly. When the number of abnormal dimensions At that time, the frame is marked as unreliable as a whole, the classification result is not included in the sliding window voting, and CUSUM does not accumulate the frame's score. The value and warning level remain unchanged. This protection mechanism is executed every frame. The computational cost of the comparison operation is negligible.

[0097] S500 outputs multi-level early warning signals based on the component category, classification confidence level, and state change indicators, combined with trend analysis results.

[0098] This phase integrates the classification results into clinically actionable multi-level early warning signals and performs long-term trend tracking.

[0099] The numerical encoding of the classification results and the core dispersion parameters are subjected to exponentially weighted moving average trend tracking. The EWMA recursive form is as follows: Smoothing factor Corresponding equivalent time constant It can effectively remove filters with a shorter cycle than Random impulse noise and single-frame misclassification fluctuations, while retaining a period longer than Minute-by-minute trend changes. The tracked objects include three dimensions: low-to-high frequency impedance ratio. (The single-dimensional feature most sensitive to changes in composition), relaxation time constant (Physical quantities directly reflecting the characteristics of suspended particles), and numerical codes for classification categories (serous = 1, hemorrhagic = 2, purulent = 3, chylous = 4, mixed = 5, with increasing values ​​corresponding to increasing clinical severity). The direction of the three-dimensional trend is determined based on the difference between 5 frames: if... Then the trend is judged to be upward (changing in a more serious direction). If the trend is downward (moving towards recovery), it is determined; otherwise, it is determined to be stable. Threshold , and For the current confirmed category This parameter in the training set The quantile extreme value range uses intra-class extreme values ​​instead of global extreme values ​​to ensure that the threshold is adaptive to different categories.

[0100] The warning level is determined based on a combined decision table of three dimensions: component category, rate of change, and duration, outputting a 4-level warning signal: Level 0 (normal state, LED always green) corresponds to: classification result as slurry (category 1) and confidence level. The EWMA trend is stable. Clinical meaning: Postoperative drainage is recovering normally.

[0101] Level 1 (Attention status, LED solid yellow) corresponds to: the classification result shows bloodiness (Category 2) but the duration does not exceed frame( This actually refers to a situation where blood is detected in all 7 consecutive frames, or in serous drainage fluid. A rising trend that has not yet triggered the CUSUM turning point. Clinical implications: Small amounts of postoperative hemorrhage are within the physiological range and require monitoring by nursing staff but no immediate intervention is necessary. Duration The threshold is determined based on clinical experience. Transient hemorrhage is usually caused by changes in body position or transient capillary leakage and should not last longer than [a certain duration]. .

[0102] Level 2 (Warning status, LED flashing orange) +buzzer Intermittent) corresponds to: bloody drainage fluid lasting longer than frame( In cases where purulent drainage (Category 3) gradually appears (EWMA trend confirms an upward trend and...), 5 consecutive frames This situation ( ) indicates possible postoperative bleeding or infection, and requires notification of a doctor for evaluation.

[0103] Level 3 (Emergency State, LED constantly red and buzzer continuously sounding) corresponds to one of the following three situations: An acute component mutation is detected ( )and acute rate of change of characteristics (Right now Inside Changes exceeded Acute appearance of purulent drainage (a direct jump from non-purulent to purulent with a certain confidence level). ); Chylic drainage fluid appears (chylous leakage is a serious complication requiring emergency treatment). Acute rate threshold. Physical meaning: In normal physiological fluctuations The change per second does not exceed ( ), correspond Normal fluctuations, false positive probability .

[0104] Confidence protection mechanism: When The system maintains the previously confirmed warning level without any changes (including neither downgrading nor escalating), avoiding level oscillations caused by low-confidence frames. This mechanism is particularly critical in the following scenario: when a bubble briefly interferes with the dispersion curve when passing through the ROI, affecting the fitting quality. A sharp drop in confidence, without protection, could lead to false oscillations from level 3 to level 0 and back to level 3.

[0105] The warning signal is output synchronously through two channels: local indication and remote communication. The local indication module includes: an RGB three-color LED driver circuit (common anode, PWM dimming), outputting green / yellow / orange / red respectively according to levels 0 / 1 / 2 / 3; and an active buzzer (drive frequency...). ), at level 2 A square wave enable generates an intermittent audible alarm, which is continuously enabled at level 3. Remote communication uses the BLE 4.2 low-power Bluetooth protocol, transmitting monitoring data frames in broadcast mode. The frame structure includes a frame identifier, timestamp, component category, confidence level, alert level, status flag, and core dispersion parameters. , , ), total length The maximum number of bytes satisfies BLE 4.2 broadcast frame. Byte payload constraints. Broadcast interval is set to [value] at level 0 / 1. To save power, it automatically switches to [mode / function] at level 2 / 3. To ensure real-time response from the nurse station terminal. Communication range. Covering the ICU single-room nurse station.

[0106] Historical data storage uses an external SPI Flash chip (W25Q64, capacity...). The structure uses a circular buffer. Each frame stores a record containing a fingerprint vector, classification result, CUSUM state, and timestamp. The size of a single frame record is... Bytes. By Sampling interval, Hourly storage frame, , Flash capacity Storeable approximately Hourly data. Write pointers are managed in a circular fashion; when writing to the end of the Flash memory, it wraps back to the starting address to overwrite the oldest data. Read pointers are controlled by the nurse station terminal when connected via BLE. Batch backtracking reading of the most recent data is supported after the nurse station connects via BLE. Historical data frames are used for post-event analysis and electronic medical record archiving.

[0107] For example, the bloody mutation scenario from S400 is continued. Frame 6 output. (Bloodlust) , Acute rate calculation The level 3 trigger condition is met. The LED immediately switches to a solid red light, the buzzer sounds continuously, and the BLE broadcast interval changes from [previous setting]. Switch to And send an emergency frame. First detected from CUSUM The latency to the warning hardware output is determined by the software logic execution time: the total time for decision table lookup and GPIO / BLE driver execution does not exceed [a certain value]. EWMA tracking results: initial value (Slurry steady state), frames 3-6 The inputs are respectively , , , When frame 6 is obtained through EWMA recursion... 5-frame trend difference The trend indicator is rising.

[0108] like Figure 8 As shown, this application also provides a multi-frequency impedance tomography monitoring system for drainage fluid. The functional modules of the system share the same ARM Cortex-M7 main control processor platform (STM32H743 level) in physical implementation, and the functional modules are divided through software time-sharing scheduling.

[0109] The multi-frequency excitation acquisition module includes: a DDS signal source chip (AD9833, MCLK, frequency resolution ), generating 8 preset frequency sine waves; improved Howland constant current source circuit (output impedance) ), converting the DDS voltage signal into Precision constant current output, current amplitude stability is better than ( ); Analog switch matrix (ADG1606, on-resistance) ), respectively controlling the switching timing of the excitation electrode pair and the measurement electrode pair; 8-channel differential instrumentation amplifier array (INA128, CMRR The voltage difference between the measuring electrodes is amplified; a 16-bit SAR ADC (ADS8688, 8-channel synchronous sampling) is used. The module completes the analog-to-digital conversion. It also includes an overcurrent protection circuit that detects when the excitation current exceeds... (safety threshold) (times) The internal excitation output is automatically cut off.

[0110] The tomographic reconstruction module resides in software within the ARM processor. It utilizes the FMC bus (32-bit data width). Clock, effective bandwidth Accessing external SDRAM The Jacobian matrix is ​​formed. The core inversion operations include the front / back substitution of the Cholesky trigonometric factors and soft thresholding. The Cholesky factor is utilized... and Sparsity of matrices, non-zero elements are reduced by a factor of 1. Each element is stored in CSR compressed format in on-chip SRAM, occupying approximately [amount missing]. 8-frequency serial inversion, working vectors with 15 iterations per frequency ( Bytes / Vectors (number of vectors) occupy SRAM buffer.

[0111] The dispersion analysis and fingerprint construction modules are also implemented in software. The execution sequence is as follows: Otsu threshold calculation (scanning the 512-element histogram at 256 gray levels) → ROI mask generation → pixel-by-pixel Cole-Cole LM fitting (128 pixels) Maximum of 50 iterations, each iteration contains Jacobian calculations and Matrix inversion → ROI statistical aggregation → Frequency ratio calculation → Fingerprint vector assembly. Peak SRAM usage per processing run is approximately... (Including ROI temporary impedance array) (Byte and LM workspace).

[0112] The pattern classification module retrieves model parameters from the on-chip SRAM (...). The pipeline reads SVM support vectors and Platt parameters, performs standardization → 80 RBF kernel function calculations → voting for 10 sub-classifiers → confidence mapping → sliding window voting → CUSUM cumulative update. The CUSUM state variable ( , (Frame counter) and the resident SRAM usage of the classification history of the last 5 frames byte.

[0113] The warning output module includes: GPIO driving RGB LEDs (3 PWM channels). Carrier frequency), GPIO enable active buzzer ( (Natural frequency), BLE 4.2 wireless communication SoC (nRF52832 level, with built-in ARM Cortex-M4F and) The BLE SoC communicates with the main controller processor via the SPI interface, and autonomously manages the broadcast timing after receiving 25-byte frame data. It also includes an RF front-end and a W25Q64 SPI Flash memory chip.

[0114] Data flow between modules is transmitted through a pre-allocated buffer in shared SRAM: the multi-frequency excitation acquisition module will... ( Bytes) written to SRAM address The starting region, from which the fault reconstruction module reads and... ( Bytes) written The initial output buffer is from which the dispersion analysis module reads and generates the fingerprint vector. ( Bytes) written The initial classification input area, the pattern classification module will output the result tuple ( Bytes) written The initial warning module input register. The entire processing pipeline is controlled by the main loop timer (TIM2, (Triggered by a periodic interrupt), each module executes sequentially in a fixed order: S000→S100→S200→S300→S400→S500. The effective processing time for a single round is approximately... The remaining Enter Stop2 low-power mode.

[0115] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit described above can be implemented in hardware.

[0116] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for multi-frequency impedance tomography monitoring of drainage fluid, characterized in that, include: By using an electrode array surrounding the outer wall of the drainage container, electrical excitation is applied to the drainage fluid at multiple preset frequencies and the boundary voltage response is collected to obtain multi-frequency measurement data. Based on the difference between the multi-frequency measurement data and the reference data, the spatial distribution of multi-frequency conductivity of the drainage fluid cross section is obtained by regularized constraint inversion; For the region of interest in the multi-frequency conductivity spatial distribution, the impedance value corresponding to each frequency is extracted and a dispersion model is fitted to obtain the dispersion characteristic parameters. An impedance fingerprint vector is constructed based on the dispersion feature parameters. The impedance fingerprint vector is then input into a pre-trained classification model for pattern classification, and the category of drainage fluid composition is output.

2. The method according to claim 1, characterized in that, The dispersion model is the Cole-Cole impedance dispersion model, and the dispersion characteristic parameters include low-frequency limiting resistance, high-frequency limiting resistance, relaxation time constant, and dispersion coefficient.

3. The method according to claim 2, characterized in that, The impedance fingerprint vector includes the mean and standard deviation of the dispersion characteristic parameters within the region of interest, as well as the frequency ratio characteristic component constructed based on multiple preset impedance ratios.

4. The method according to claim 1, characterized in that, The electrode array includes multiple electrodes evenly distributed along the circumference of the drainage container. Multiple rounds of data acquisition are completed sequentially using an adjacent excitation-adjacent measurement mode to obtain the multi-frequency measurement data.

5. The method according to claim 1, characterized in that, The regularization constraint inversion adopts a hybrid constraint of smooth regularization term and sparse gradient regularization term, and solves the spatial distribution of conductivity corresponding to each frequency through alternating direction iteration.

6. The method according to claim 1, characterized in that, Before inputting the impedance fingerprint vector into the pre-trained classification model, the impedance fingerprint vector is further subjected to standardization processing; the pre-trained classification model is a support vector machine classifier based on a kernel function.

7. The method according to claim 6, characterized in that, It also includes performing sliding window time-series smoothing on the classification results, calculating the cumulative deviation statistic on the time series of the impedance fingerprint vector to perform change point detection, and outputting a state change flag when the cumulative deviation statistic exceeds a preset decision threshold.

8. The method according to claim 1, characterized in that, Before performing the acquisition boundary voltage response, a system calibration step is also included: performing full-band reference measurement on the drainage container filled with reference medium to obtain the reference data, and pre-calculating the sensitivity matrix corresponding to each frequency.

9. A multi-frequency impedance tomography monitoring system for drainage fluid, characterized in that, include: The multi-frequency excitation acquisition module is used to apply electrical excitation to the drainage fluid at multiple preset frequencies through an electrode array surrounding the outer wall of the drainage container and acquire the boundary voltage response to obtain multi-frequency measurement data. The fault reconstruction module is used to obtain the spatial distribution of multi-frequency conductivity of the drainage fluid cross section through regularized constraint inversion based on the difference between the multi-frequency measurement data and the reference data. The dispersion analysis and fingerprint construction module is used to extract the impedance values ​​corresponding to each frequency in the region of interest in the multi-frequency conductivity spatial distribution and fit the dispersion model to obtain dispersion feature parameters, and construct an impedance fingerprint vector based on the dispersion feature parameters. The pattern classification module is used to input the impedance fingerprint vector into a pre-trained classification model for pattern classification, output the drainage fluid component category and classification confidence, and perform change point detection on the time series of the impedance fingerprint vector to output state change flags.

10. The system according to claim 9, characterized in that, It also includes an early warning output module, which is used to output multi-level early warning signals based on the component category, classification confidence level and state change indicators, combined with trend analysis results.

Images