Method for establishing tumor risk prediction model based on bioelectrical impedance technology

Abstract

The invention discloses a method for establishing a tumor risk prediction model based on a bioelectrical impedance technology, and particularly relates to the field of tumor risk prediction, which comprises the following steps: on the basis of the traditional Chinese medicine meridian theory, collecting body surface bioelectrical impedance signals, performing extraction and fusion analysis of multi-scale characteristics, and dynamically optimizing model selection; and non-invasive tumor risk diagnosis is realized. According to the establishment method of the tumor risk prediction model based on the bioelectrical impedance technology, by adopting a multi-channel bioelectrical impedance acquisition circuit, the pertinence and stability of signal acquisition are optimized, and the problem that the data throughput is too large under the high sampling rate is relieved; the influence of physique typing and dynamic physiological change on basic impedance is compensated, so that the signal quality and the real-time performance of subsequent processing are improved; through the reinforcement learning strategy, the tumor risk prediction model matched with the physiological state of the patient is dynamically selected, and the real-time response speed and the diagnosis decision accuracy are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tumor risk prediction, more particularly, the present application relates to a method for establishing a tumor risk prediction model based on bioelectrical impedance technology. BACKGROUND

[0002] With the development of precision medicine and modernization of traditional Chinese medicine, the non-invasive bioelectrical impedance detection technology combined with traditional meridian theory provides a new idea for early screening of lung tumors. In traditional Chinese medicine diagnosis, the lung state is often assisted by judging the resistance changes of the acupoints of the lung meridian of Hand Taiyin. The existing technology mainly uses single-point electrodes to measure the impedance values of specific acupoints, and realizes simple virtual and real syndrome differentiation through experience threshold comparison. However, due to the limitation of limited measurement dimensions, the accuracy of the diagnosis result and the sensitivity to complex lesions are often affected.

[0003] In order to improve the diagnosis accuracy and information richness, the existing technology develops an array electrode, combines time-frequency joint analysis, obtains multi-acupoint bioelectric signals from multiple spatial sites, and introduces time-frequency analysis to extract deeper physiological and pathological characteristics, so as to increase the dimension of data acquisition and the complexity of data analysis.

[0004] However, in actual use, there are still some shortcomings, such as high sampling rate, which increases the data throughput and makes it difficult to realize real-time feature locking; the multi-source heterogeneous data of twelve meridians lack a fusion framework consistent with traditional Chinese medicine theory, resulting in dynamic characteristics being submerged by noise, especially when facing traditional constitution typing and dynamic physiological changes, which easily affects the basic impedance, thereby causing the problem of key features being submerged, ultimately limiting the real-time response speed in complex diagnosis scenarios and the accuracy and robustness of the final diagnosis decision. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the present application provides a method for establishing a tumor risk prediction model based on bioelectrical impedance technology, which solves the problems raised in the background art through the following scheme.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical scheme: The method for establishing a tumor risk prediction model based on bioelectrical impedance technology comprises: S1: selecting a non-invasive skin measurement point on the body surface of a target patient, and acquiring a first bioelectric signal through a multi-channel bioelectrical impedance acquisition circuit; S2: using an impedance dynamic compensation mechanism to adapt the first bioelectric signal to individual skin impedance, and acquiring a first key signal; S3: based on the first key signal, using a deep learning algorithm for multi-scale decomposition to extract a first signal feature corresponding to the skin measurement point; S4: calculating the degree of pair-wise interdependence between the first signal features through a correlation matrix; S5: integrating the first signal features corresponding to the skin measurement points using an attention-guided cross-modal fusion network, and regulating the information flow through a differentiable gating mechanism to obtain second signal features that distinguish between benign and malignant tumors; S6: calculating the absolute value of correlation based on historical diagnosis data for the second signal features, and selecting third signal features (number of feature attribute values); S7: based on the third signal features, dynamically obtaining ROC-AUC values and weighting through reinforcement learning strategies, and selecting a tumor risk prediction model that matches the physiological state of the target patient.

[0007] Preferably, the S1, the selection of the non-invasive skin measurement points needs to meet the following conditions: The main measurement points are located at at least 3 core acupoints of the middle-fu, chi ze, kong zui, tai yuan, and shao shang of the lung meridian, supplemented by the lung meridian and other acupoints; The number of measurement points is ≤12 and the distribution conforms to the topological direction of the lung meridian; According to the individual constitution type of the target patient, additional acupoints are selected.

[0008] Preferably, the S2, the implementation of the impedance dynamic compensation mechanism, specifically includes: Complete the static correction of the patient's constitution type through the skin impedance pre-compensation coefficient; Adopt adaptive weighted filtering, based on the correlation between the skin impedance pre-compensation coefficient and the constitution type, dynamically obtain the adaptive weight factor reflecting the constitution sensitivity and the impedance change rate, and calculate the compensation value of the first bioelectric signal , specifically expressed as: Wherein, represents the first bioelectric signal pre-compensated by individual differences, represents the constitution coefficient, which is used to quantify the sensitivity of different constitutions to impedance changes, represents the change amount of real-time skin impedance relative to the reference impedance, represents the dynamic response factor, which is used to adjust the response speed to the impedance change rate, represents the change rate of real-time skin impedance over time.

[0009] Preferably, the S2, before implementing the impedance dynamic compensation mechanism, calculates the skin impedance pre-compensation coefficient, specifically including: Based on the real-time skin impedance , the reference impedance and the real-time skin temperature Real-time skin impedance pre-compensation coefficient , specifically expressed as: , wherein, is expressed as the rate of change of skin impedance with temperature, is expressed as the reference temperature.

[0010] Preferably, S3, the first signal feature includes minimum value type features and maximum value type features corresponding to all non-invasive skin measurement points; The minimum value type features extract the minimum values of impedance maximum value, impedance minimum value, A curve area, B curve area, total curve area, zero-crossing point sampling number, and rising edge slope by summarizing the measurement values; The maximum value type features summarize the maximum values of falling edge slope and negative mutation amplitude of all measurement points.

[0011] Preferably, S4, the degree of mutual dependence between the first signal features is calculated using a weighted correlation coefficient , specifically expressed as: , wherein, and respectively represent different first signal features, and respectively represent the weighted average values corresponding to different first signal features, represents the number of data of the first signal features, represents the index of the first signal features, represents the meridian topology weight.

[0012] Preferably, S5, the generation of the second signal feature, specifically includes: Assigning weights based on the similarity between the first signal features corresponding to each measurement point and tumor characteristics; Integrating a differentiable gating mechanism using a gating unit structure, the gating unit combines the patient's physiological state and TCM constitution typing.

[0013] Preferably, S7, the selection of the tumor risk prediction model, specifically includes: According to the patient's state, use a reinforcement learning strategy to select a tumor risk prediction model; The environment of the reinforcement learning strategy is defined as the patient's physiological state space, which includes constitution typing and syndrome evolution; The state of the reinforcement learning strategy is composed of the third signal features; The action of the reinforcement learning strategy is to select a prediction model from a pre-set model library; In the preset model library, an tumor risk prediction model corresponding to the physiological state space is acquired, and the preset model library is used to save the corresponding relationship between the physiological state space and the tumor risk prediction model.

[0014] Technical effects and advantages of the present application: 1、The present application selects non-invasive skin measurement points in line with the topological direction of the lung meridian through S1, and adopts a multi-channel bioelectrical impedance acquisition circuit, thereby optimizing the pertinence and stability of signal acquisition, reducing the initial noise interference of multi-source heterogeneous data, relieving the problem of excessive data throughput under high sampling rate, and laying a data foundation for subsequent real-time feature locking. 2、The present application compensates and suppresses noise through the impedance dynamic compensation mechanism of S2 in view of the influence of physique typing and dynamic physiological changes on the basic impedance, generates a continuous and smooth measurement curve, solves the problem of key feature submersion caused by physique difference and environmental interference of multi-source heterogeneous data, and improves signal quality and real-time performance of subsequent processing. 3、The present application dynamically selects a tumor risk prediction model matching the physiological state of the patient through the reinforcement learning strategy of S7, solves the problem of insufficient model adaptability in complex diagnosis scenarios, and improves the real-time response speed and accuracy of diagnosis decision. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 A step block diagram of the method for establishing a tumor risk prediction model based on bioelectrical impedance technology according to the embodiments of the present application is provided.

[0016] Figure 2 A flowchart of acquiring a second signal feature in the method for establishing a tumor risk prediction model based on bioelectrical impedance technology according to the embodiments of the present application is provided. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0018] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.

[0019] Hereinafter, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first," "second," and "third" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0020] As attached Figure 1 The method for establishing a tumor risk prediction model based on bioelectrical impedance technology, as shown, includes, based on the theory of meridians in traditional Chinese medicine, collecting bioelectrical impedance signals from the body surface, performing impedance compensation, deep learning multi-scale feature extraction and fusion, feature screening, and dynamically selecting a model based on a reinforcement learning strategy to achieve non-invasive tumor risk diagnosis. Specifically, it includes: S1: Select non-invasive skin measurement points on the target patient's body surface and acquire the first bioelectrical signal through a multi-channel bioelectrical impedance acquisition circuit; S2: An impedance dynamic compensation mechanism is used to adapt the first bioelectric signal to individual skin impedance to obtain the first key signal; S3: Based on the first key signal, use a deep learning algorithm to perform multi-scale decomposition and extract the first signal features corresponding to the skin measurement points; S4: Calculate the pairwise interdependence between the first signal features using the correlation matrix; S5: An attention-guided cross-modal fusion network is used to integrate the first signal features corresponding to the skin measurement points, and the information flow is regulated by a differentiable gating mechanism to obtain the second signal features that distinguish between benign and malignant tumors; S6: Calculate the absolute value of the correlation of the second signal feature based on historical diagnostic data, and select the third signal feature; S7: Based on the third signal feature, the ROC-AUC value is dynamically obtained and weighted through a reinforcement learning strategy, and a tumor risk prediction model that matches the physiological state of the target patient is selected.

[0021] Specifically, in S1, a non-invasive skin measurement point is selected on the target patient's body surface, and a multi-modal bioelectric signal, i.e., a first bioelectric signal, is acquired by using a multi-channel bioelectrical impedance acquisition circuit.

[0022] In a possible implementation, the conditions to be met for selection of the non-invasive skin measurement point include: the main measurement points are located at at least three core acupoints of the lung meridian of Hand-Taiyin, i.e., Zhongfu, Chize, Kongzhi, Taeyin and Shaoshang, and are supplemented by acupoints of the lung meridian branch; the number of measurement points is less than or equal to 12 and the distribution conforms to the topological direction of the lung meridian; and the acupoints are adaptively selected according to the individual constitution type of the target patient.

[0023] It should be noted that the acupoints of the lung meridian of Hand-Taiyin include Yunmen, Tianfu, Xiaobai, Lieque, Jingqu, Yujing, Zhongfu, Chize, Kongzhi, Taeyin, Shaoshang; in this embodiment, if the individual constitution of the target patient is the type of yin deficiency and dry heat, the Taixi acupoint and the Zhaohai acupoint are added, and the interference of virtual fire on the signal is reduced; if the individual constitution of the target patient is the type of phlegm-dampness accumulating in the lung, the Fenglong acupoint and the Yinlingquan acupoint are added, and the influence of phlegm-dampness on the signal is reduced from the source.

[0024] Further, the multi-channel bioelectrical impedance acquisition circuit adopts a multi-channel architecture to synchronously acquire signals of multiple measurement points, and each channel includes a silver chloride sintered electrode, an instrument amplifier and a fourth-order Bessel band-pass filter; in this embodiment, the number of channels of the multi-channel bioelectrical impedance acquisition circuit is set to be less than or equal to 12, and the sampling rate is set to be 500 Hz.

[0025] Still further, the multi-channel bioelectrical impedance acquisition circuit is coated with a conductive paste and a hydrogel composite layer at the interface between the electrode and the skin, and is connected by a three-wire connection method.

[0026] In this embodiment, the connection method of the three-wire connection method includes: directly contacting the skin through the excitation electrode; selecting one measurement electrode to contact the skin, and the distance between the measurement electrode and the excitation electrode is 5 mm; selecting another measurement electrode to contact the skin, and the distance between the measurement electrode and the excitation electrode is 10 mm; the excitation current of the excitation electrode is set to be a sinusoidal signal of 50 μA and 50 kHz, and the sampling frequency is set to be 500 Hz to synchronously acquire data of 12 channels.

[0027] Specifically, in S2, an impedance dynamic compensation mechanism is used to adapt the first bioelectric signal to the individual skin impedance, thereby generating a measurement curve corresponding to each non-invasive skin measurement point, i.e., a first key signal.

[0028] It should be noted that real-time data preprocessing is performed before the first bioelectric signal is adapted to the individual skin impedance; the real-time data preprocessing includes but is not limited to eliminating abnormal signals, segmenting and caching, signal conditioning, environmental interference protection, individual difference pre-compensation, etc.

[0029] In the embodiment, the abnormal signals are removed, if the signal amplitude exceeds 3 times of the standard deviation or the frequency is higher than 150 Hz, the frame data is discarded immediately and the re-sampling is triggered to remove the abnormal data.

[0030] In the embodiment, the segmented cache is set for each channel of the multi-channel bioelectric impedance acquisition circuit, and the data is segmented and cached in units of 200 ms.

[0031] In the embodiment, the signal conditioning adopts a three-step filter chain to optimize the signal, including removing pulse noise by median filtering, retaining waveform characteristics while smoothing data by Savitzky-Golay, and eliminating baseline drift caused by physiological activities such as respiration by adaptive baseline correction.

[0032] In the embodiment, the environmental interference protection constructs a three-level shielding system during acquisition to ensure signal purity, including using a twisted shielded wire with a coverage rate of 95% as the electrode line, using an aluminum-magnesium alloy shell with a shielding effectiveness of >60 dB as the acquisition box, and setting a Faraday cage in the operation room.

[0033] It should be noted that the individual difference pre-compensation takes pre-compensation measures for signal variation introduced by patient physical differences, and calculates the real-time skin impedance pre-compensation coefficient based on real-time skin impedance , reference impedance and real-time skin temperature , which is specifically expressed as: Among them, represents the rate of change of skin impedance with temperature, represents the reference temperature; the reference temperature is taken as the physiological state and electrical characteristics of the skin being considered stable and standard, and the reference temperature in the embodiment is taken as 25℃.

[0034] In a possible implementation, the impedance dynamic compensation mechanism includes: completing the static correction of patient physical typing through the skin impedance pre-compensation coefficient; using adaptive weighted filtering to dynamically obtain an adaptive weight factor reflecting the constitution sensitivity and the impedance change rate based on the correlation between the skin impedance pre-compensation coefficient and the constitution typing, and calculating the compensation value of the first bioelectric signal, which is specifically expressed as: Among them, ​a first bioelectric signal pre-compensated by individual difference, a body constitution coefficient, for quantifying the sensitivity of different body constitutions to impedance change, a change amount of real-time skin impedance relative to reference impedance, a dynamic response factor, for adjusting the response speed to impedance change rate, a change rate of real-time skin impedance over time; in this embodiment, the body constitution coefficient is 0.7 for phlegm-dampness accumulation in the lung type and 1.2 for yin deficiency and dryness-heat type; the dynamic response factor is 0.05.

[0035] It should be noted that the skin impedance pre-compensation coefficient combines body constitution typing to pre-compensate, providing key input for dynamic compensation mechanism, including: directly associating signal characteristics of different body constitutions; in this embodiment, a +20% gain compensation is provided for yin deficiency and dryness-heat type patients in the low frequency band less than 1 Hz; a -15% gain compensation is provided for phlegm-dampness accumulation in the lung type patients in the medium frequency band of 1-10 Hz, all through the skin impedance pre-compensation coefficient to inject compensation logic in advance; directly determining the adjustment direction of dynamic compensation; in this embodiment, when the skin impedance pre-compensation coefficient is greater than 1, the dynamic compensation mechanism will adjust the gain of the impedance signal through ; when the skin impedance pre-compensation coefficient is less than 1, then attenuation adjustment is realized.

[0036] Further, the first bioelectric signal after impedance dynamic compensation is used to obtain a first key signal, i.e., a measurement curve corresponding to each skin measurement point, and the generation of the measurement curve uses the Overlap-add method to splice data segments of every 200 ms to form a continuous and smooth measurement curve.

[0037] Specifically, in S3, based on the first key signal after impedance dynamic compensation generated in S2, a deep learning algorithm is used to extract first signal features with physiological and pathological significance from the signal corresponding to each skin measurement point in combination with multi-scale decomposition technology.

[0038] In this embodiment, the first key signal is derived from 3 repeated tests, and 9 signal features are obtained from each measurement point of 208 subjects.

[0039] In one possible implementation, the first signal features include minimum value features and maximum value features corresponding to all non-invasive skin measurement points; wherein, the minimum value features are obtained by summarizing the measurement values ​​from multiple trials, extracting the minimum value of impedance, the area under curve A, the area under curve B, the total area of ​​the curve, the number of zero-crossing samples, and the minimum value of the rising edge slope; the maximum value features are obtained by summarizing the maximum values ​​of the falling edge slope and the negative abrupt change amplitude of all measurement points.

[0040] It should be noted that the maximum impedance value This value reflects cell density; an increased value indicates a correlation with cell proliferation in malignant tumors, corresponding to the phlegm-blood stasis syndrome in Traditional Chinese Medicine; the minimum impedance value... The area under the A curve is used to reflect the fullness of tissue fluid; a decrease in the value indicates an increased risk of tumors, corresponding to lung qi deficiency syndrome in traditional Chinese medicine. Used to reflect low-frequency energy, associated with deep tissue abnormalities, corresponding to the lung meridian obstruction syndrome in Traditional Chinese Medicine; the area under the B curve Used to reflect mid-frequency energy, associated with microcirculatory disorders, corresponding to Qi stagnation and blood stasis syndrome in Traditional Chinese Medicine; the total area of ​​the curve As a total energy indicator related to tumors, it reflects the strength of the syndrome of deficiency of vital energy and excess of pathogenic factors; the zero-crossing sampling number Used to reflect signal complexity, related to tissue heterogeneity, and corresponding to the degree of Yin-Yang imbalance in Traditional Chinese Medicine; the rising slope This value reflects the cell proliferation rate; an increase in the value indicates rapid growth of malignant tumors, corresponding to the syndrome of "excessive pathogenic factors" in Traditional Chinese Medicine; the slope of the descending edge... Reflecting the rate of tissue necrosis, an increase in the value indicates a correlation with tissue destruction in advanced tumors, corresponding to the syndrome of "decline of vital energy" in Traditional Chinese Medicine; the negative mutation amplitude... It reflects signs of blood vessel rupture, indicating the risk of tumor metastasis, corresponding to the syndrome of blood stasis outside the meridian in traditional Chinese medicine.

[0041] Furthermore, a dual-stream hybrid network architecture is constructed, comprising a 1D-CNN stream and a BiLSTM stream; the 1D-CNN stream is used to extract local spatial features of the signal to capture the fine structure of the acupoint signal; the BiLSTM stream is used to capture the long-term temporal dependencies of the signal and establish the dynamic process of meridian conduction.

[0042] In the embodiment, the wavelet packet-EMD hybrid method is adopted for multi-scale decomposition, and the decomposition process is as follows: db4 wavelet base is selected, and the signal band of 0-250 Hz is divided into 16 sub-bands; the sub-band containing the target physiological information is subjected to empirical mode decomposition, and the intrinsic mode function component is adaptively extracted; the frequency band is reconstructed according to the theory of traditional Chinese medicine, the IMF component corresponding to the main frequency band of the lung meridian is extracted, and the component of the high-frequency interference frequency band is excluded, so as to focus on the key physiological information related to the tumor, that is, the first signal feature.

[0043] Specifically, in S4, based on the first signal feature extracted in S3, a correlation matrix between the features is constructed and analyzed, and the degree of mutual dependence is calculated.

[0044] Further, the first signal feature is subjected to standardized processing according to traditional Chinese medicine, and different types of features are subjected to different normalization methods; in the embodiment, the standard deviation normalization is adopted for the maximum impedance and the minimum impedance, so as to eliminate the baseline drift caused by the physical differences, and the specific expression is as follows: wherein, is the first signal feature after the normalization processing, is the numerical value of the first signal feature, is the average value of the numerical value corresponding to the first signal feature, is the variance of the numerical value corresponding to the first signal feature; the energy feature of the AUC series is subjected to logarithmic transformation, so as to conform to the exponential growth law of the evil qi, and the specific expression is as follows: wherein, is the first signal feature after the normalization processing, is the numerical value of the first signal feature; the dynamic feature of the rising edge slope and the falling edge slope is subjected to nonlinear normalization based on the physical coefficient , so as to adapt to the dynamic characteristics under different physical conditions, and the specific expression is as follows: wherein, is the first signal feature after the normalization processing, is the numerical value of the first signal feature, is the maximum value of the numerical value corresponding to the first signal feature, is the physical coefficient; further, the first signal feature after the normalization is constructed in the form of a matrix, wherein the row of the matrix represents the acupoint measurement data corresponding to the measurement point of the target patient, and the column of the matrix represents the feature vector.

[0045] In a possible implementation, the pair-wise interdependence degree between the first signal features is calculated using a weighted correlation coefficient , which is specifically represented as: , wherein, and represent different first signal features, respectively, and represent the weighted average values corresponding to different pairs of first signal features, respectively, represents the number of data of the first signal features, represents the index of the first signal features, represents the meridian topology weight, and in this embodiment, the weight of the lung meridian acupoint is set to 1.0, the weight of the large intestine meridian acupoint is 0.3, and the weights of other meridian acupoints are 0.1.

[0046] Further, based on the calculated pair-wise interdependence degree , an interpretation table is established to map the numerical range of the pair-wise interdependence degree to the interdependence degree and the TCM pathogenesis explanation; in this embodiment, when the numerical range of the pair-wise interdependence degree is between 0.8 and 1.0, it represents a strong correlation, corresponding to the resonance between acupoints of the same meridian; when the numerical range of the pair-wise interdependence degree is between 0.5 and 0.8, it represents a strong correlation, reflecting the interaction between the front and back meridians; when the numerical range of the pair-wise interdependence degree is between 0.3 and 0.5, it represents a moderate correlation, which is related to the association between the meridians under the five-element generation and restraint relationship; and when the numerical range of the pair-wise interdependence degree is between 0 and 0.3, it represents a weak correlation, mainly reflecting the noise of irrelevant meridians.

[0047] Specifically, in S5, an attention-guided cross-modal fusion network is constructed to integrate the first signal features corresponding to the skin measurement points extracted in S3, and a differentiable gating mechanism is used to dynamically regulate the information flow to generate second signal features that can effectively distinguish between benign and malignant tumors.

[0048] It should be noted that the second signal features also include minimum value features and maximum value features; wherein the minimum value features are the minimum values of the impedance maximum value, the impedance minimum value, the A curve area, the B curve area, the total curve area, the zero-crossing point sampling number and the rising edge slope of the benign and malignant tumors, respectively, which are summarized from the historical data of the benign and malignant tumors; and the maximum value features are the maximum values of the falling edge slope and the negative mutation amplitude of all the historical data of the benign and malignant tumors.

[0049] In a possible implementation, generating the second signal features includes assigning weights based on the similarity between the first signal features corresponding to each measurement point and the tumor features, which is specifically represented as: wherein, the attention weight of the i-th measurement point, and respectively represent the matrix constructed by the first signal feature corresponding to the i-th and j-th measurement point,

[0050] It should be noted that the feature vector of the tumor region comes from the historical diagnosis database.

[0051] In one possible implementation, generating the second signal feature further includes integrating a differentiable gating mechanism using a gating unit structure, and the gating unit combines the physiological state and TCM constitution type of the patient.

[0052] In this embodiment, in the case of yin deficiency and dryness-heat constitution and motion artifact, the update gate tends to be closed and the reset gate tends to be opened, so as to retain the basic syndrome features learned from history, while resetting or ignoring the input information disturbed by the motion artifact; in the case of phlegm-dampness accumulation in the lung constitution and stable impedance, the update gate tends to be opened and the reset gate tends to be closed, so as to update the features related to phlegm and blood stasis sufficiently, while retaining the stable state learned from history.

[0053] Specifically, in S6, the second signal feature generated in S5 is subjected to correlation analysis using the large-scale historical diagnosis data stored in the historical diagnosis database, the correlation absolute value between the second signal feature and the tumor benignity and malignancy diagnosis result is calculated, and the number of features with predictive value, i.e., the third signal feature, is selected based on the correlation absolute value.

[0054] It should be noted that the historical diagnosis data comes from the historical diagnosis database, and the historical diagnosis database is used to save a large amount of multi-modal data related to meridian bioelectricity features, including acupoint-disease association model data established based on evidence-based medicine principles, pathological feature waveform library annotated by experts, and standardized individual difference compensation parameter set.

[0055] Further, the correlation absolute value between the second signal feature and the tumor benignity and malignancy diagnosis result is calculated , and the classical Spearman rank correlation coefficient is defined, which is specifically represented as: ​​​​​​​wherein, is expressed as a body constitution coefficient, is expressed as a difference between the second signal feature and the tumor benignity and malignancy diagnosis result, is expressed as a syndrome type deviation amount, is expressed as a syndrome type adjustment coefficient, is expressed as a number of the second signal features.

[0056] It should be noted that the difference between the second signal feature and the tumor benignity and malignancy diagnosis result reflects the degree of contradiction between the second signal feature and the tumor benignity and malignancy diagnosis result, when , the feature ranking is consistent; when > 0, it indicates that the disease is complex; the syndrome type deviation amount is a quantitative degree of syndrome transformation, when = 0, it indicates that the patient is in a stable period; > 0, it indicates that the patient is in a transformation period; the syndrome type adjustment coefficient is used to control the attenuation intensity of the correlation of syndrome transformation.

[0057] In this embodiment, based on the calculated correlation absolute value, a double-threshold dynamic screening mechanism is implemented to perform hierarchical feature screening, including: setting a global threshold, retaining strong correlation features with a correlation absolute value greater than 0.7 with the tumor diagnosis result; setting a constitution-specific threshold, for patients with yin deficiency, retaining features with a correlation absolute value greater than 0.6 and still significant after the AUCTotal drops in weight; for patients with phlegm-dampness, retaining features with a correlation absolute value greater than 0.65 and still significant after the Drop rises in weight.

[0058] Further, the feature number optimization control is performed by using the Pareto optimization rule, including: drawing a relationship curve of the feature number and the AUC value, and observing the change trend of the AUC value with the increase of the feature number; in this embodiment, when the feature number reaches 5, the AUC value reaches a peak value of 0.92, and when the feature number exceeds 7 dimensions, the AUC value decreases instead, indicating overfitting; then, 5 core features are retained as the third signal features, and the third signal features include the AUCTotal difference after attention weighting and gate processing, the smallest AUCTotal value in the lung meridian acupoint, the largest Drop value in the lung meridian acupoint, a quantitative value reflecting the degree of yin deficiency, and a correlation coefficient reflecting the degree of phlegm and blood stasis.

[0059] Specifically, in S7, based on the third signal features screened out in S6, the performance of a plurality of prediction models is dynamically evaluated by using a reinforcement learning strategy, and combined with the theory of traditional Chinese medicine, a tumor risk prediction model that best matches the physiological state of the current target patient is finally selected.

[0060] In the embodiment, when performing model selection and verification, the feature template data of the corresponding disease is called from the historical diagnosis data by using 500 resampling iterations of the Monte Carlo model verification method to complete online fine-tuning and decision credibility cross-validation of the tumor risk prediction model.

[0061] In a possible implementation, selecting the tumor risk prediction model matching the physiological state of the target patient comprises: selecting the tumor risk prediction model according to the physiological state of the patient by using a reinforcement learning strategy; an environment of the reinforcement learning strategy is defined as a physiological state space of the patient, and the physiological state space comprises a constitution type and an evolution of a syndrome; a state of the reinforcement learning strategy is constituted by the third signal feature; an action of the reinforcement learning strategy is to select a prediction model from a preset model library; in the preset model library, a tumor risk prediction model corresponding to the physiological state space is obtained, and the preset model library is used to save a corresponding relationship between the physiological state space and the tumor risk prediction model.

[0062] It should be noted that the preset model library is derived from a historical diagnosis database.

[0063] Further, an online AUC estimation method is used to calculate the AUC value of the selected tumor risk prediction model in real time based on a sliding window, so as to evaluate the performance of the tumor risk prediction model under different constitutions.

[0064] Further, based on the third signal feature and the physiological state of the patient, and according to the real-time AUC and the constitution differentiation benchmark, the model with the highest Q value in the current state is output as the tumor risk prediction model matching the physiological state of the target patient, and is specifically represented as: wherein, represents a function of the expected cumulative reward that can be obtained under a specific state under the action represents a learning rate parameter, and the value range is 0-1, represents a reward obtained immediately after the action ac is taken in the current state st, represents a discount factor, and the value range is also 0-1, represents a next state after state transition, represents a next state after state transition, represents a next state after state transition, represents a next state after state transition.

[0065] Secondly, only the structures related to the disclosed embodiments are involved in the drawings of the disclosed embodiments, and other structures can be referred to the general design, and the same embodiments and different embodiments of the present application can be combined with each other under the condition of no conflict; Finally: the above only for the preferred embodiments of the present application, and not for limiting the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application, should be included in the scope of protection of the present application.

Claims

1. A method for establishing a tumor risk prediction model based on bioelectrical impedance technology, characterized in that, The method comprises the following steps: S1: Selecting a non-invasive skin measurement point on the body surface of a target patient, and acquiring a first bioelectric signal through a multi-channel bioelectric impedance acquisition circuit; S2: Using an impedance dynamic compensation mechanism to adapt the first bioelectric signal to individual skin impedance, and acquiring a first key signal; S3: Based on the first key signal, using a deep learning algorithm for multi-scale decomposition to extract the first signal features corresponding to the skin measurement point; S4: Calculating the degree of mutual dependence between the first signal features through a correlation matrix; S5: Using an attention-guided cross-modal fusion network to integrate the first signal features corresponding to the skin measurement point, and regulating the information flow through a differentiable gating mechanism to obtain second signal features that can distinguish between benign and malignant tumors; S6: Based on historical diagnosis data, calculating the correlation absolute value of the second signal features, and selecting third signal features; S7: Based on the third signal features, dynamically obtaining the ROC-AUC value through a reinforcement learning strategy and weighting, and selecting a tumor risk prediction model that matches the physiological state of the target patient. 2.The method of claim 1, wherein the method further comprises: determining a tumor risk prediction model based on the bioelectrical impedance analysis. In the S1, the selection of the non-invasive skin measurement point needs to meet the following conditions: The main measurement points are located at at least 3 core acupoints of the middle-fu, chi ze, kong zui, tai yuan, and shao shang of the lung meridian, and are supplemented by the lung meridian branch meridian acupoints; The number of measurement points is ≤12 and the distribution conforms to the topological direction of the lung meridian; The acupoints are adaptively selected according to the individual constitution type of the target patient. 3.The method of claim 1, wherein the method further comprises: determining a tumor risk prediction model based on the bioelectrical impedance analysis. In the S2, the implementation of the impedance dynamic compensation mechanism comprises: Completing the static correction of the patient's constitution type through a skin impedance pre-compensation coefficient; Adopting self-adaptive weighted filtering, based on the correlation between the skin impedance pre-compensation coefficient and the body constitution type, dynamically obtaining the self-adaptive weight factor reflecting the body constitution sensitivity and the impedance change rate, calculating the compensation value of the first bioelectric signal , which is specifically represented as: wherein, is the first bioelectric signal pre-compensated for individual differences, is the body constitution coefficient, quantifying the sensitivity of impedance change to different body constitutions, is the change amount of real-time skin impedance relative to the reference impedance, is the dynamic response factor, adjusting the response speed to the impedance change rate, is the rate of change of real-time skin impedance over time. 4.The method of claim 3, wherein the method further comprises: determining a tumor risk prediction model based on the bioelectrical impedance data. Before implementing the impedance dynamic compensation mechanism, the skin impedance pre-compensation coefficient is calculated, which comprises: Real-time skin impedance based reference impedance and real-time skin temperature to calculate real-time skin impedance pre-compensation coefficient , which is specifically represented as: wherein, the rate of change of skin impedance with temperature, denotes the reference temperature. 5.The method of claim 1, wherein the method further comprises: determining a tumor risk prediction model based on the bioelectrical impedance analysis. In the S3, the first signal features include the minimum value features and the maximum value features corresponding to all non-invasive skin measurement points; The minimum value features are obtained by summarizing the measurement values, extracting the maximum impedance, the minimum impedance, the area under the A curve, the area under the B curve, the total area of the curve, the number of zero-crossing points, and the minimum value of the rising edge slope; The maximum value features are obtained by summarizing the maximum values of the falling edge slope and the negative mutation amplitude of all measurement points. 6.The method of claim 1, wherein the method further comprises: obtaining a plurality of bioelectrical impedance data of the subject; and determining a plurality of bioelectrical impedance parameters of the subject based on the plurality of bioelectrical impedance data. The S4 calculates the degree of pair-wise interdependence between the first signal features using a weighted correlation coefficient , and is specifically represented as: wherein, with are represented as different first signal features, with are represented as weighted averages of different first signal feature pairs, is represented as a data quantity of the first signal feature, is represented as an index of the first signal feature, is represented as a meridian topology weight. 7.The method of claim 1, wherein the method further comprises: determining a tumor risk prediction model based on the bioelectrical impedance analysis. In the S5, the generation of the second signal features comprises: Assigning weights based on the similarity between the first signal features corresponding to each measurement point and the tumor characteristics; Integrating a differentiable gating mechanism with a gating unit structure, and the gating unit combines the patient's physiological state and the constitution type of traditional Chinese medicine. 8.The method of claim 1, wherein the method further comprises: obtaining a plurality of bioelectrical impedance data of the subject; and determining a plurality of bioelectrical impedance parameters of the subject based on the plurality of bioelectrical impedance data. In the S7, the selection of the tumor risk prediction model comprises: Selecting a tumor risk prediction model using a reinforcement learning strategy according to the patient's state; The environment of the reinforcement learning strategy is defined as the physiological state space of the patient, which includes the constitution type and the evolution of the syndrome; The state of the reinforcement learning strategy is composed of the third signal features; The action of the reinforcement learning strategy is to select a prediction model from a preset model library. In the preset model library, an tumor risk prediction model corresponding to the physiological state space is acquired, and the preset model library is used to save the corresponding relationship between the physiological state space and the tumor risk prediction model.

Images