A method and system for interpretation and diagnosis of nerve conduction study data

Abstract

The present application relates to the technical field of nerve electrophysiological examination data processing, in particular to a method and system for nerve conduction examination data interpretation and diagnosis, first performing field extraction, term standardization and other preprocessing on the examination data to obtain a detection item set, performing pre-determination through a preset 15-class peripheral nerve injury disease diagnosis rule base, directly outputting results and determination basis if the diagnosis rule is met, and entering a model interpretation process if not; based on nerve conduction anatomy prior identification of pairing relationship and generation of comparison features, constructing a direction consistent deviation matrix in combination with normal population data; fusing multi-dimensional features and constructing a nerve conduction anatomy relationship graph, using a relationship perception heterogeneous graph attention network coding with edge type embedding, and extracting global features through comparison perception attention pooling enhanced by pairing difference to realize disease classification. The present application can improve interpretation efficiency and accuracy, reduce subjective errors, and provide objective auxiliary diagnosis support for neuromuscular diseases.

Description

Technical Field

[0001] This invention relates to the field of neurophysiological examination data processing technology, and in particular to a method and system for interpreting and diagnosing nerve conduction examination data. Background Technology

[0002] Nerve conduction testing is a core technique in clinical neuroelectrophysiology for assessing peripheral nerve function. It is widely used in the diagnosis and differential diagnosis of carpal tunnel syndrome, cubital tunnel syndrome, peroneal nerve palsy, Guillain-Barré syndrome, and various peripheral neuropathy. This test involves applying electrical stimulation at different anatomical sites and recording the evoked responses to obtain a series of measurements, including amplitude, conduction velocity, latency, and F-wave latency. These measurements reflect axonal function, myelin conduction function, distal conduction time, and proximal conduction status. A complete nerve conduction test typically involves multiple nerves, multiple stimulation sites, and various test types, ultimately generating dozens or even hundreds of measurement data points. These data points have complex anatomical correlations and clinical logical relationships. For example, amplitude attenuation between distal and proximal stimulation sites of the same nerve can be used to determine conduction block, and differences between the same segments of the same nerve on both sides can be used to identify unilateral involvement. Therefore, the interpretation of nerve conduction tests is not essentially a comparison of isolated values, but a reasoning and decision-making process that requires comprehensive information from multiple dimensions, including individual measurement values, proximal-distal comparisons, left-right symmetry, and the anatomical pathway of the nerve.

[0003] Existing technologies objectively suffer from the following shortcomings: Current methods typically treat various nerve conduction test indicators as independent feature inputs to the model, lacking explicit modeling of the clinical pairing relationships between test items. This prevents the effective utilization of crucial diagnostic criteria in nerve conduction, such as amplitude comparison between proximal and distal ends and left-right contralateral symmetry, making it difficult for the model to capture pathological patterns dependent on pairwise comparisons, such as conduction block and segmental demyelination. Furthermore, existing methods often employ uniform normalization or standardization when processing measurements of different test items and indicators, failing to fully consider the differences in the clinical abnormality direction of different indicators. For example, while both decreased amplitude and prolonged latency are abnormal, their numerical changes are opposite. Uniform scaling can confuse the model's abnormality interpretation direction, weakening its ability to represent the degree of deviation from true pathology. Graph neural network-based methods often treat edges from different sources as the same connection type when constructing the relationship graph between detection items, failing to semantically distinguish between different clinical relationships such as proximal-distal pairings, contralateral pairings, and adjacent segments. This results in the graph network being unable to aggregate information with a focus on different diagnostic logics during message transmission, limiting the model's ability to learn complex neurotransmission pathological patterns. Existing methods often use simple strategies such as average pooling or max pooling when aggregating detection-level features into patient-level global features, without considering the differences in the contribution of different detection items to the final diagnosis. In particular, they fail to highlight key detection items with significant pairing differences, which can easily lead to the key abnormal signals of focal lesions being diluted by a large number of normal detection items, or being dominated by a single extreme value while ignoring the cumulative diagnostic value of multiple mild abnormalities.

[0004] Furthermore, existing automatic interpretation schemes mostly adopt a single model architecture driven by pure data, without incorporating clinically recognized classic diagnostic rules as a pre-verification step. For common peripheral nerve injury diseases with typical characteristics, all deep learning calculations still need to be completed, and the interpretation efficiency needs to be improved. Moreover, the output results lack direct support from classic diagnostic logic, resulting in insufficient interpretability.

[0005] Therefore, this invention proposes a method and system for interpreting and diagnosing nerve conduction test data to solve the above problems. Summary of the Invention

[0006] This invention addresses the shortcomings of existing technologies by proposing a method and system for interpreting and diagnosing nerve conduction test data. This invention can improve interpretation efficiency and accuracy, reduce subjective errors, and provide objective auxiliary diagnostic support for neuromuscular diseases.

[0007] On the one hand, the technical solution of the present invention to solve the technical problem is a method for interpreting and diagnosing nerve conduction test data, comprising the following steps: S1. Perform field extraction, terminology standardization, measurement validity marking, and test item sorting on the patient's nerve conduction examination data in sequence to obtain a set of test items. Each test item includes the quasi-nerve name, left and right sides, nerve type, stimulation site, recording site, and at least one measurement index. First, the test items are judged according to the preset diagnostic rule library for peripheral nerve injury diseases. If the test item meets the diagnostic rule of any disease, the rule diagnosis result of the disease and the corresponding judgment basis are directly output; otherwise, the subsequent operation is performed. S2. Based on the prior rules of neural conduction anatomy, automatically identify the proximal and distal pairing relationships and the contralateral pairing relationships in the set of detection items, convert the paired differences into explicit contrast features, and generate a contrast feature matrix; S3. Construct a grouped normal reference distribution using a historical normal population database, convert patient test values ​​into directional deviation codes, concatenate all the code values ​​of each test item into a vector in a fixed order, and then stack all the vectors of the test items in a standard order to form a directional deviation matrix. S4. Robust scaling results are obtained by performing robust scaling on the original measurement data of the test item set. This result, the contrast feature matrix, the direction consistency deviation matrix, and the measurement validity label are then mapped to the same hidden dimension and fused to generate the test item feature sequence. At the same time, a neural conduction anatomical relationship map is constructed based on the anatomical relationship between the test items. S5. Using the feature sequence of the detection item as input, a trainable edge type embedding is introduced into the graph attention mechanism. The attention score between nodes is calculated through the relation-aware heterogeneous graph attention network, and the graph-encoded node feature matrix is ​​output. S6. Based on the difference between the graph coding node features of the detection item and the paired detection item, the paired difference enhancement feature vector is obtained. It is added to the contrast perception attention pooling weight calculation. The weighted summation of all graph coding node features is used to obtain the patient-level global feature vector. Based on this, the disease category prediction probability vector is calculated, and the disease category corresponding to the highest probability is selected as the output result. S7. Train the entire model using training samples with real disease category labels to obtain a trained automatic interpretation model for neural conduction examination data. Use this model to perform online interpretation and diagnosis.

[0008] S1 is as follows: The patient's nerve conduction examination data was read and extracted, and organized into a set of test items. The measurement indicators included amplitude measurement, conduction velocity measurement, latency measurement, and F-wave latency measurement. A terminology standardization dictionary was established, and the original names in the patient's nerve conduction examination data were converted into standard terms. For each indicator type of each test item, an effective indicator tag was generated to indicate whether there was a measurement value that could be used for calculation. The measurement units were then standardized, and finally, the test item set was sorted according to the standardized order of the test items. The system includes a pre-defined diagnostic rule base for 15 types of peripheral nerve injuries. Each type of injury has its own specific test criteria and threshold values. The 15 types of peripheral nerve injuries are: carpal tunnel syndrome, cubital tunnel syndrome, proximal median nerve injury, carpal tunnel syndrome combined with cubital tunnel syndrome, axonal peripheral neuropathy, axonal peripheral neuropathy combined with carpal tunnel syndrome, axonal peripheral neuropathy combined with cubital tunnel syndrome, acute demyelinating Guillain-Barré syndrome, acute motor axonal neuropathy-type Guillain-Barré syndrome, chronic Guillain-Barré syndrome, radial nerve injury, common peroneal nerve injury, tibial nerve ankle injury, and femoral nerve injury. The specific diagnostic rules are as follows: (1) Admission criteria for test items: The minimum set of test items required for the corresponding disease determination is pre-stored. Each test item in the minimum set of test items is determined by the standard nerve name, nerve type, stimulation site, and recording site respectively. The rule matching process for the disease is initiated only when the patient's set of test items covers the minimum set of test items corresponding to the disease. (2) Multi-branch judgment logic: There is at least one type of absolute value judgment branch and relative difference judgment branch. Each type may contain one or more judgment branches. The judgment branches are logically ORed. If all the limiting conditions of any judgment branch are met, it is considered to meet the diagnostic rules of the disease. The judgment logic of the absolute value judgment branch is: the specified measurement index value of the target detection item exceeds the preset normal threshold range, and the corresponding measurement index value of the reference detection item used as a control meets the preset normal threshold range. The judgment logic of the relative difference judgment branch is: the calculated difference of the same measurement index of two paired test items with a control relationship exceeds the preset difference threshold.

[0009] The specific process of constructing the contrast feature matrix in S2 is as follows: S2.1. Establish a preset anatomical segment sequence table based on preset neural conduction anatomy rules, record the proximal and distal order of stimulation sites under the same standard nerve name, nerve type, and recording site, and retrieve proximal and distal paired test items for each test item in the test item set according to the standard nerve name, left and right sides, nerve type, stimulation site, and recording site, while recording the distal and proximal test items in the paired relationship; then, retrieve contralateral paired test items for each test item according to the standard nerve name, left and right sides, nerve type, stimulation site, and recording site, and generate valid proximal and distal paired markers and valid contralateral paired markers; S2.2 Based on the identified pairing relationships, calculate the clinically significant contrast features for each test item from the original measurements. Explicitly construct these key diagnostic criteria into contrast feature vectors. Stack the contrast feature vectors of all test items row by row according to the standardized order of the test items to obtain the contrast feature matrix.

[0010] The construction of the directional deviation matrix in S3 is as follows: S3.1 Generate a normal reference group index based on the standard nerve name, nerve type, stimulation site and recording site. Retrieve the historical normal sample values ​​corresponding to the normal reference group index in the historical normal population database, and save the historical normal sample distribution according to the indicator type. When the number of historical normal samples under the indicator type of the normal reference group index is insufficient, obtain the backup normal reference distribution according to the preset merging rules. S3.2 For each test item and indicator type, extract the historical normal sample value of the same indicator type from the historical normal sample distribution corresponding to the normal reference group index, and compare the current original measurement value with the historical normal sample value in terms of rank; calculate the empirical rank probability based on the rank, and input the empirical rank probability into the inverse function of the cumulative distribution function of the standard normal distribution to obtain the rank normal quantile value; then, according to the abnormal direction of the indicator type, convert the rank normal quantile value into the direction consistency deviation code value. For indicator types that do not exist in each test item or are not clinically applicable, set the corresponding direction consistency deviation code value to 0. For indicator types that have a measurement value but exceed the range of historical normal samples, use the minimum rank or the maximum rank for encoding, and trim the direction consistency deviation code value to the preset range. S3.3. Concatenate all directional deviation codes of each detection item into a vector in a fixed order, and then stack all the vectors of the detection items into a directional deviation matrix in a standard order.

[0011] The specific process of generating the feature sequence of the detection item in S4 is as follows: By utilizing group quantile information, the original measurement values ​​with different dimensions and ranges are uniformly scaled to the [0,1] interval to generate robust scaling values; The robust scaling vector, contrast feature vector, orientation consistency deviation encoding vector, and effective label vector of each detection item are concatenated along the feature dimension in a fixed order to obtain a multi-granularity numerical feature vector. This multi-granularity numerical feature vector is then input into a linear projection layer and mapped to a unified hidden dimension to obtain a numerical fusion representation. Next, trainable embedding tables are established for standard nerve names, left / right sides, nerve types, stimulation sites, and recording sites, and the corresponding embedding vectors are read from these tables. Then, anatomical sequence position encoding vectors are generated based on a preset anatomical segment order table. The numerical fusion representation, each embedding vector, and the anatomical sequence position encoding vector are added together to obtain the comprehensive feature vector for each detection item. Finally, all comprehensive feature vectors for each detection item are stacked row by row according to the standardized order of the detection items to form a detection item feature sequence.

[0012] The specific process of constructing the neural conduction anatomy diagram in S4 is as follows: Each test item is treated as a node, and a node set is constructed. A set of proximal-distal paired edges is established based on the indexes of distal and proximal paired distal test items and the indexes of proximal-distal paired proximal test items. A set of contralateral paired edges is established based on the index of contralateral paired test items. A set of edges for adjacent segments of the same nerve is established based on a preset anatomical segment sequence table. A set of edges for nerve fibers with the same name is established based on whether they have a clinical relationship. A set of self-loop edges is established for each test item, resulting in a set of self-loop edges. All edge sets are merged and combined with the node set to form a neural conduction anatomical relationship diagram.

[0013] S5 is detailed below: S5.1. The feature sequence X of the detection item is used as the input of the relation-aware heterogeneous graph attention network. Each attention head performs linear projection on the node feature vector of the target detection item and the node feature vector of the neighbor detection item to obtain the query representation, key representation and value representation. A trainable edge type embedding is introduced into the graph attention mechanism. The query representation, key representation and edge type embedding vector are input into the attention scoring function to obtain the relation-aware attention score. S5.2. Softmax exponential normalization is performed on the relation perception scores calculated for all neighbors of each node, and the scores are converted into attention weights with a sum of 1. S5.3. Based on the calculated attention weights, the features of all neighboring nodes are weighted and summed to obtain the relationship aggregation results under each attention head. The relationship aggregation results of multiple attention heads are concatenated and used as the candidate node update features for the next layer. Then, the standard operation of the Transformer architecture and deep network in deep learning is used to obtain the node feature vector of each detection item in the next layer. S5.4 By stacking multiple layers of relation-aware heterogeneous graph attention layers, each detection node gradually aggregates information from its multi-hop neighbors, outputting graph-coded node features rich in anatomical relationships and contextual information, thus forming a graph-coded node feature matrix.

[0014] S6 is detailed below: S6.1. Each detection item generates a pairing difference enhancement feature by subtracting its own graph encoding node features from the features of the clinical pairing item, and then concatenates it with its own graph encoding node features to obtain the pairing difference enhancement feature vector. S6.2. A trainable pooling scoring function is adopted, with paired differential enhancement features as input, to automatically learn and calculate the importance of each detection item to the final diagnosis, and generate pooling weights after normalization. S6.3. Based on the contrast-aware pooling weights, the graph coding node features of all detection items are weighted and summed to aggregate a fixed-length patient-level global feature vector. The patient-level global feature vector is input into the fully connected classification head to obtain the disease category score. The disease category score is then input into the normalized exponential function to obtain the disease category prediction probability vector. Finally, the disease category with the highest probability in the disease category prediction probability vector is taken as the preliminary judgment result, and the detection items with higher contrast-aware attention pooling weights are output as judgment criteria.

[0015] S7 is detailed below: Each training patient sample is organized into a training test item set, and the true disease category label is determined by clinical diagnosis results or expert review results; the historical normal sample distribution, low quantile reference value and high quantile reference value are statistically analyzed from normal patient samples or normal population samples confirmed to have no evidence of peripheral neuropathy in the training set; The loss function is calculated based on the predicted structure obtained from the training set data, and the trainable parameters are updated through iterative training. The optimal parameters are retained to obtain a trained automatic interpretation model for neural conduction examination data.

[0016] On the other hand, the present invention also provides a system for interpreting and diagnosing nerve conduction test data, including a module for executing processing instructions for each step of a method for interpreting and diagnosing nerve conduction test data, comprising: The data preprocessing module is used to perform field extraction, terminology standardization, measurement validity marking, and test item sorting on the patient's nerve conduction examination data to obtain a set of test items. Each test item includes the quasi-nerve name, left and right sides, nerve type, stimulation site, recording site, and at least one measurement index. The pairing relationship identification and contrast feature generation module is used to automatically identify proximal and contralateral pairing relationships in the set of test items based on prior rules of neural conduction anatomy, and convert pairwise differences into explicit contrast features to obtain a contrast feature matrix; The directional deviation matrix construction module is used to construct a grouped normal reference distribution using a historical normal population database, convert patient test values ​​into directional deviation codes, concatenate all directional deviation code values ​​of each test item into a vector in a fixed order, and then stack all vectors of test items into a directional deviation matrix in a standard order. The multi-feature fusion and anatomical relationship map construction module is used to robustly scale the original measurement data in the detection item set to obtain robust scaling results. After mapping the results, the contrast feature matrix, the direction consistency deviation matrix and the measurement value validity label to the same hidden dimension, the module is fused to generate the detection item feature sequence and construct the neural conduction anatomical relationship map based on the anatomical relationship between the detection items. The heterogeneous graph encoding module takes the feature sequence of the detection item as input, introduces trainable edge type embedding in the graph attention mechanism, calculates the attention score between nodes through the relation-aware heterogeneous graph attention network, and outputs the graph-encoded node feature matrix. The global feature extraction and disease prediction module is used to obtain the pairing difference enhancement feature vector based on the difference between the graph coding node features of the detection item and the paired detection item, add it to the contrast perception attention pooling weight calculation, and obtain the patient-level global feature vector by weighted summation of the features of all graph coding nodes. Then, the disease category prediction probability vector is calculated based on this, and the disease category corresponding to the highest probability is selected as the output result. The model training and online interpretation module is used to train the model of the entire system using training samples with real disease category labels, and obtain a trained automatic interpretation model for nerve conduction examination data. This model is then used for online interpretation and diagnosis.

[0017] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects: This invention discloses a method and system for interpreting and diagnosing nerve conduction test data. Through a test item relationship graph construction method based on anatomical priors, the unstructured list of nerve conduction test items is transformed into a heterogeneous graph structure containing proximal-distal pairing edges, contralateral pairing edges, edges between adjacent segments of the same nerve, and edges related to the same nerve fibers. This allows the model to explicitly utilize pairing relationship information such as amplitude attenuation, decreased conduction velocity, and left-right asymmetry, which are crucial for clinical diagnosis, rather than treating all test items as independent inputs. A direction consistency deviation coding mechanism based on the rank distribution of normal populations is proposed. This mechanism compares the patient's measured values ​​with the historical normal sample distribution of their corresponding groups, generates rank-normal quantile values ​​using the inverse transformation of the standard normal distribution, and then corrects the direction based on the clinical abnormality direction of each indicator. This allows different abnormal directions, such as decreased amplitude and prolonged latency, to be uniformly represented as positive abnormalities and negative normalities, with comparable codes. This paper addresses the inconsistencies in the dimensions, normal ranges, and abnormal directions of different detection items and indicators. It proposes a heterogeneous graph attention network with edge type awareness. Building upon the standard graph attention mechanism, it maps different clinical relationships, such as proximal-distal pairing and contralateral pairing, into trainable edge type embedding vectors, which participate in the calculation of attention scores. This enables the message transmission process to distinguish the differentiated impact of different anatomical relationships on diagnosis, avoiding the aggregation of information by treating edges with different clinical meanings as homogeneous relationships. Furthermore, it proposes a contrast-aware attention pooling mechanism. When generating patient-level global features, it explicitly concatenates the graph encoding feature differences between each detection item and its clinically paired detection item into a pairing difference enhancement feature. A trainable pooling scoring function is used to give higher pooling weights to detection items with significant proximal-distal amplitude attenuation or left-right asymmetry, thus avoiding the dilution of key lesion information by average pooling or the neglect of accumulated evidence of multiple mild to moderate abnormalities by max pooling.

[0018] By using a rule base of 15 types of peripheral nerve injury diseases for pre-judgment, the system directly outputs diagnostic results and judgment criteria for typical cases that conform to classical diagnostic logic. This not only improves the efficiency of interpreting common diseases but also ensures the interpretability of classical diagnostic logic. At the same time, rule-based processing reduces the processing pressure on the deep learning model, allowing the model to focus on the feature discrimination of complex and atypical cases, further improving the overall accuracy and robustness of the interpretation. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0020] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0021] Figure 2 This is a heatmap corresponding to the direction-consistent deviation matrix.

[0022] Figure 3 This is a diagram showing the anatomical relationships of nerve conduction.

[0023] Figure 4 The scaling values ​​for the 12 detection items are given under amplitude scaling, conduction velocity scaling, latency scaling, and F-wave latency scaling. Detailed Implementation

[0024] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0025] Example 1 like Figure 1 As shown, a method for interpreting and diagnosing nerve conduction test data includes the following steps: S1. Perform field extraction, terminology standardization, measurement validity marking, and test item sorting on the patient's nerve conduction examination data in sequence to obtain a set of test items. Each test item includes the quasi-nerve name, left and right sides, nerve type, stimulation site, recording site, and at least one measurement index. First, the test items are judged according to the preset diagnostic rule library for peripheral nerve injury diseases. If the test item meets the diagnostic rule of any disease, the rule diagnosis result of the disease and the corresponding judgment basis are directly output; otherwise, the subsequent operation is performed. S2. Based on the prior rules of neural conduction anatomy, automatically identify the proximal and distal pairing relationships and the contralateral pairing relationships in the set of detection items, convert the paired differences into explicit contrast features, and generate a contrast feature matrix; S3. Construct a grouped normal reference distribution using a historical normal population database, convert patient test values ​​into directional deviation codes, concatenate all the code values ​​of each test item into a vector in a fixed order, and then stack all the vectors of the test items in a standard order to form a directional deviation matrix. S4. Robust scaling results are obtained by performing robust scaling on the original measurement data of the test item set. This result, the contrast feature matrix, the direction consistency deviation matrix, and the measurement validity label are then mapped to the same hidden dimension and fused to generate the test item feature sequence. At the same time, a neural conduction anatomical relationship map is constructed based on the anatomical relationship between the test items. S5. Using the feature sequence of the detection item as input, a trainable edge type embedding is introduced into the graph attention mechanism. The attention score between nodes is calculated through the relation-aware heterogeneous graph attention network, and the graph-encoded node feature matrix is ​​output. S6. Based on the difference between the graph coding node features of the detection item and the paired detection item, the paired difference enhancement feature vector is obtained. It is added to the contrast perception attention pooling weight calculation. The weighted summation of all graph coding node features is used to obtain the patient-level global feature vector. Based on this, the disease category prediction probability vector is calculated, and the disease category corresponding to the highest probability is selected as the output result. S7. Train the entire model using training samples with real disease category labels to obtain a trained automatic interpretation model for neural conduction examination data. Use this model to perform online interpretation and diagnosis.

[0026] In a specific implementation, S1 involves the collection and standardization of patient nerve conduction examination data and test items. This invention performs field extraction, terminology standardization, measurement value validity marking, and test item sorting on the patient nerve conduction examination data to obtain a set of test items. The specific steps are as follows: S1.1 Extraction of patient nerve conduction test data fields 1) Read the patient's nerve conduction test data and extract the test item name, nerve name, left and right sides, nerve type, stimulation site, recording site, amplitude measurement value, conduction velocity measurement value, latency measurement value, F wave latency measurement value and waveform quality label from the patient's nerve conduction test data.

[0027] Patient nerve conduction test data can come from structured tables exported by the device or from test reports that have been manually reviewed. If the patient's nerve conduction test data comes from an unstructured report, the nerve name, left / right side, nerve type, stimulation site, and recording site can be identified first using rule templates or natural language entity extraction methods before proceeding to the subsequent standardization process.

[0028] 2) Organize the patient's nerve conduction examination data into a test item set Ja. Test item set Ja refers to the data set consisting of all nerve conduction tests actually completed by the patient. Test item set Ja includes L test items, where L represents the number of test items actually completed by the patient. The value of the test item index i ranges from 1 to L. The i-th test item must include at least the standard nerve name. , left and right sides Neurological type Stimulation sites Recording sites And at least one measurement indicator.

[0029] Among them, standard nerve names Used to differentiate the origins of different nerves such as the median nerve, ulnar nerve, common peroneal nerve, tibial nerve, sural nerve, and radial nerve; left and right sides. Used to distinguish between the left and right sides; nerve type for distinguishing motor nerve conduction test, sensory nerve conduction test and F-wave test; stimulation site for identifying the anatomical position where electrical stimulation is applied; recording site for identifying the muscle or sensory recording area where the evoked response signal is collected.

[0030] 3) Establish an indicator type set Q for the i-th test item, where the indicator type set Q includes an amplitude indicator type, a conduction velocity indicator type, a latency indicator type and an F-wave latency indicator type. The original measured value of the i-th test item under the indicator type q is denoted as , where q represents the index of the indicator type, and q may take the value of the amplitude indicator type, the conduction velocity indicator type, the latency indicator type, or the F-wave latency indicator type.

[0031] The amplitude indicator type is used to reflect axonal function and the number of effective conductive fibers; the conduction velocity indicator type is used to reflect myelin conduction function and segmental conduction efficiency; the latency indicator type is used to reflect distal conduction time and the degree of neuromuscular conduction delay; the F-wave latency indicator type is used to reflect proximal nerve conduction and the proximal conduction state of the motor nerve root.

[0032] S1.2. Test item term standardization and measured value validity marking 1) Establish a term standardization dictionary, and use the term standardization dictionary to convert the original names in the patient's nerve conduction test data into standard terms. The term standardization dictionary includes a nerve name dictionary, a left-right side dictionary, a nerve type dictionary, a stimulation site dictionary, and a recording site dictionary.

[0033] In one embodiment, the patient's nerve conduction test data usually comes from a test form, a structured report or a manually entered record exported from an electromyography device. Different hospitals, different devices and different operators may use different naming methods for the same test item. For example, "Median", "median N" and "median nerve" may refer to the same standard nerve name, and "APB" and "abductor pollicis brevis" may refer to the same recording site.

[0034] In this embodiment, the nerve name dictionary is used to unify "Median", "median" and "median N" into "median nerve", and unify "Ulnar" and "ulnar N" into "ulnar nerve"; the recording site dictionary is used to unify "APB" and "abductor pollicis brevis" into "abductor pollicis brevis", and unify "ADM" and "abductor digiti minimi" into "abductor digiti minimi"; the stimulation site dictionary is used to unify "wrist", "wrist" and "wrist portion" into "wrist portion". Through this processing, the same clinical test item will not be mistakenly regarded as different test items in the subsequent process due to different name writing methods.

[0035] 2) Generate an indicator validity mark for each indicator type q of the i-th test item Valid indicator marking This is used to characterize whether the i-th detection item has a measurable value that can be calculated under index type q. When the i-th detection item has a measurable value under index type q, the unit is valid, and the waveform quality label is not a serious artifact, the index is marked as valid. The value is set to 1; when the measured value is missing, the unit cannot be identified, or the waveform quality mark is a serious artifact, the indicator is marked as valid. The value is 0.

[0036] In practice, motor nerve conduction testing typically includes amplitude, latency, and conduction velocity measurements; sensory nerve conduction testing typically includes amplitude, latency, and conduction velocity measurements; F-wave testing typically includes F-wave latency measurements, but not necessarily conventional amplitude and conduction velocity measurements. For indicator types that are not clinically applicable, effective markers are used. Also set it to 0 to prevent the model from misinterpreting non-existent indicators as abnormally low values.

[0037] 3) Standardize the measurement units. Unify the amplitude index type to millivolt or microvolt before proceeding to subsequent calculations. Specifically, the amplitude of motor nerve waves is preferably unified to millivolt, and the amplitude of sensory nerve waves is preferably unified to microvolt. The conduction velocity index type is unified to meters per second. The latency index type and the F wave latency index type are unified to milliseconds.

[0038] It should be noted that motor nerve amplitude and sensory nerve amplitude can retain their respective clinically commonly used units, because subsequent normal reference grouping and robust scaling are all performed within the same standard nerve name, the same nerve type, the same stimulation site, and the same recording site, and will not directly mix and compare motor nerve millivolt amplitude and sensory nerve microvolt amplitude.

[0039] 4) Sort the test item set Ja according to the standardized order of the test items. The standardized order of the test items refers to the pre-set stable sorting rules, which prioritize the arrangement of test items according to the standard nerve name, left and right sides, nerve type, recording site, and proximal and distal ends of the stimulation site.

[0040] In one implementation, nerves are first categorized into upper limb nerves, lower limb nerves, and other nerves. Then, within the same standard nerve name, they are sorted by left and right sides. Next, within the same left / right side, they are sorted by motor nerve conduction detection, sensory nerve conduction detection, and F-wave detection. Finally, within the same nerve type, they are sorted from distal to proximal according to a pre-defined anatomical segment order table. Based on this, the set of test items Ja has a consistent arrangement logic across different patients, facilitating the subsequent construction of the test item feature sequence.

[0041] In one embodiment, as an example, suppose the test item set Ja includes test items for left upper limb median nerve motor conduction at the wrist and elbow, and test items for right upper limb ulnar nerve sensory conduction at the wrist. Following a standardized order, the nerves are first sorted by major nerve categories, with upper limb nerves (median and ulnar nerves) preceding lower limb nerves; within the same nerve, the left side precedes the right; within the same side, motor nerve conduction tests precede sensory nerve conduction tests; and within the same nerve type, according to a preset anatomical segment order table, the wrist (distal) precedes the elbow (proximal). Therefore, the final order is: left median nerve motor wrist, left median nerve motor elbow, and right ulnar nerve sensory wrist.

[0042] S1.3, Diagnostic rule base for peripheral nerve injury diseases First, the test items are judged according to the preset diagnostic rule library for peripheral nerve injury diseases. If the test item meets the diagnostic rule of any disease, the rule diagnosis result of the disease and the corresponding judgment basis are directly output; otherwise, the subsequent operation is performed. The system includes a pre-defined diagnostic rule base for 15 types of peripheral nerve injuries. Each type of injury has its own specific test criteria and threshold values. The 15 types of peripheral nerve injuries are: carpal tunnel syndrome, cubital tunnel syndrome, proximal median nerve injury, carpal tunnel syndrome combined with cubital tunnel syndrome, axonal peripheral neuropathy, axonal peripheral neuropathy combined with carpal tunnel syndrome, axonal peripheral neuropathy combined with cubital tunnel syndrome, acute demyelinating Guillain-Barré syndrome, acute motor axonal neuropathy-type Guillain-Barré syndrome, chronic Guillain-Barré syndrome, radial nerve injury, common peroneal nerve injury, tibial nerve ankle injury, and femoral nerve injury. The diagnostic rule base is based on literature such as *Electromyography and Neuromuscular Disorders* and clinical practice. The specific diagnostic rules are as follows: (1) Admission criteria for test items: The minimum set of test items required for the corresponding disease determination is pre-stored. Each test item in the minimum set of test items is determined by the standard nerve name, nerve type, stimulation site, and recording site respectively. The rule matching process for the disease is initiated only when the patient's set of test items covers the minimum set of test items corresponding to the disease. (2) Multi-branch judgment logic: There is at least one type of absolute value judgment branch and relative difference judgment branch. Each type may contain one or more judgment branches. The judgment branches are logically ORed. If all the limiting conditions of any judgment branch are met, it is considered to meet the diagnostic rules of the disease. The judgment logic of the absolute value judgment branch is: the specified measurement index value of the target detection item exceeds the preset normal threshold range, and the corresponding measurement index value of the reference detection item used as a control meets the preset normal threshold range. The judgment logic of the relative difference judgment branch is: the calculated difference of the same measurement index of two pairs of paired test items with a control relationship exceeds the preset difference threshold.

[0043] This rule-based pre-processing mechanism achieves a complementary advantage between classical clinical diagnostic logic and deep learning models. Typical cases rely on rules to quickly and interpretably output conclusions, while atypical and complex cases rely on deep learning models to uncover deeper pathological features, thus balancing interpretation efficiency, accuracy, and clinical interpretability.

[0044] In one embodiment, the diagnostic criteria for carpal tunnel syndrome specifically include: (1) Admission criteria for test items: The test item set shall include at least the median nerve sensory wrist-index finger test item, the median nerve motor wrist-abductor pollicis brevis test item, the ulnar nerve sensory wrist-little finger test item, and the ulnar nerve motor wrist-abductor little finger test item; (2) Multi-branch decision logic: 1) First diagnostic branch: If the latency measurement value of the median nerve sensory wrist-index finger test item is greater than 4.4ms, and the measurement values ​​of the corresponding sensory and motor test items of the ulnar nerve are within the normal threshold range, then it is determined to be carpal tunnel syndrome. 2) Second diagnostic branch: If the latency measurement value of the median nerve sensory wrist-index finger test is less than or equal to 4.4 ms, and any of the following conditions are met, it is diagnosed as carpal tunnel syndrome: Absolute value determination branch: The test item set includes both the median nerve sensory wrist-ring finger test item and the ulnar nerve sensory wrist-ring finger test item, and the difference between the two latency measurements is greater than 0.4ms; Relative difference determination branch: The test item set includes both the median nerve sensory wrist-thumb test item and the radial nerve sensory forearm-thumb test item, and the difference between the two latency measurements is greater than 0.4ms.

[0045] In a specific implementation, step S2 involves identifying the pairing relationships of detection items and extracting contrastive features. Since the diagnostic basis for nerve conduction examination data includes not only the measured values ​​of individual detection items, but also the proximal-distal differences between different stimulation sites of the same nerve, and the symmetry differences between the same segments of the left and right contralateral nerves, directly inputting detection items as independent data into the model can easily lead to the loss of crucial diagnostic information such as amplitude attenuation, decreased conduction velocity, prolonged latency, and left-right asymmetry. Therefore, this invention, based on prior rules of nerve conduction anatomy, automatically identifies the proximal-distal pairing relationships and contralateral pairing relationships in the detection item set Ja, and converts paired differences into explicit contrastive features to obtain a contrastive feature matrix. The specific steps are as follows: S2.1 Identification of Detection Item Pairing Relationships This step, based on preset neural conduction anatomy rules, automatically identifies the proximal and distal pairs and contralateral pairs for each test item in the test item set Ja, and generates corresponding pairing relationship indexes and validity markers, transforming the unstructured test item list into structured data with clear clinical relevance. The specific steps are as follows: 1) Establish a preset anatomical segment sequence table. The preset anatomical segment sequence table is used to record the proximal and distal order of stimulation sites under the same standard nerve name, the same nerve type, and the same recording site.

[0046] In practical implementation, median nerve motor conduction testing can be set to stimulation sites at the wrist, elbow, and axilla; ulnar nerve motor conduction testing can be set to stimulation sites at the wrist, below the elbow, above the elbow, and axilla; and tibial nerve motor conduction testing can be set to stimulation sites at the ankle and popliteal fossa. The preset anatomical segment order table is not determined based on the order of appearance of the test items in the report, but rather on the nerve conduction pathway and clinical examination habits, thereby avoiding matching errors caused by differences in report ordering.

[0047] 2) According to standard nerve names , left and right sides Neurological type Stimulation sites and recording sites , for the first Each detection item retrieves paired detection items for both near and far ends.

[0048] In practice, the system first filters candidate detection items from the detection item set Ja that have the same standard nerve name, same left / right side, same nerve type, same recording site, or equivalent recording site as the i-th detection item. Then, it determines whether the stimulation site of the candidate detection item is located in the same preset anatomical segment sequence table as the stimulation site of the i-th detection item. If there are multiple candidate detection items, the candidate detection item adjacent to the stimulation site of the i-th detection item is selected as the proximal-distal paired detection item to highlight local segmental changes. If the clinical rules specify a fixed combination, such as a fixed pairing of wrist stimulation sites and elbow stimulation sites, the proximal-distal paired detection items are determined according to the fixed combination.

[0049] Furthermore, to avoid ambiguity regarding the near and far orientations, this invention records not only the paired detection items but also the far-end and near-end detection items within that pairing relationship. Specifically, for the i-th detection item, the far-end pairing far-end detection item index is denoted as... The index of the proximal detection item paired with the far-near end is denoted as .in, Used to point to the far-end detection item that is in the same near-far pairing relationship as the i-th detection item. This is used to point to the near-end detection item that is in the same near-far pairing relationship as the i-th detection item. By simultaneously storing the far-end detection item index and the near-end detection item index, it can be ensured that the subsequent amplitude attenuation ratio and latency extension ratio are always calculated in a clear direction.

[0050] In one embodiment, as an example, for test item i, which involves left median nerve motor conduction, wrist stimulation, and recording in the abductor pollicis brevis muscle, the system retrieves test item j from Ja that has the same standard nerve name, side, nerve type, and recording site, and whose stimulation site is at the elbow. According to a preset anatomical segment sequence table, the wrist is distal and the elbow is proximal. Therefore, the proximal-distal pairing index for test item i... Pointing to itself i, the index of the near-end detection item Pointing to j; the detection item j Pointing to i, Points to itself, j.

[0051] 3) According to standard nerve names , left and right sides Neurological type Stimulation sites and recording sites , retrieve the counterpart detection item for the i-th detection item.

[0052] In the specific implementation, firstly, candidate detection items are selected from the detection item set Ja that have the same standard nerve name, opposite left / right side, same nerve type, same stimulation site or corresponding anatomical segment, same recording site or equivalent recording site as the i-th detection item. If multiple candidate detection items exist, the detection item with complete measurement value, qualified waveform quality labeling, and closest examination time is selected as the contralateral paired detection item. The index of the contralateral paired detection item is denoted as . , Used to point to the detection items corresponding to the left and right sides of the i-th detection item.

[0053] For example, when testing the left median nerve motor conduction at the wrist stimulation site and recording the abductor pollicis brevis response, the contralateral paired test item should be the right median nerve motor conduction at the wrist stimulation site and recording the abductor pollicis brevis response, and the right ulnar nerve wrist stimulation test item should not be misjudged as the contralateral paired test item.

[0054] 4) Generate a valid marker for near-far pairing for the i-th detection item. Pairing valid markers with the opposite side Effective markers for proximal and distal pairing. The value is 1 when a proximal-distal pairing exists, and 0 when no proximal-distal pairing exists; contralateral pairing valid marker. The value is 1 when there is a paired detection item on the opposite side, and 0 when there is no paired detection item on the opposite side.

[0055] In one embodiment, for example, the i-th detection item is a wrist stimulation site detection item for left median nerve motor conduction detection. If the detection item set Ja contains an elbow stimulation site detection item for left median nerve motor conduction detection, then the wrist stimulation site detection item is determined as a distal detection item, and the elbow stimulation site detection item is determined as a proximal detection item, and let... If the test item set Ja contains a wrist stimulation site test item for right median nerve motor nerve conduction detection, then this right wrist stimulation site test item is determined as the contralateral paired test item, and let... .

[0056] S2.2, Feature Extraction for Detection Items This step, based on the identified pairing relationships, calculates clinically significant contrast features from the raw measurements, such as the proximal-distal amplitude attenuation ratio and the difference in contralateral conduction velocity. These key diagnostic criteria are explicitly constructed as feature vectors, allowing the model to directly utilize this differential information without implicitly learning from the raw data. The specific steps are as follows: 1) Read the amplitude measurement value, conduction velocity measurement value and latency measurement value of the i-th detection item and its paired detection item, and determine whether the corresponding indicator can participate in the comparison calculation based on the valid indicator mark.

[0057] In practice, the corresponding far-near contrast feature is calculated only when the i-th detection item, the far-near paired detection item, and the far-near paired detection item all have valid labels for their corresponding indicators; similarly, the corresponding opposite-side contrast feature is calculated only when the i-th detection item and the opposite-side paired detection item both have valid labels for their corresponding indicators. If any necessary measurement value is missing, the contrast feature is set to 0, and the model is prompted to determine whether the contrast feature is truly usable based on the valid labels for the far-near paired detection item or the valid labels for the opposite-side paired detection item.

[0058] 2) When the proximal and distal ends are paired, a valid marker is used. When the value is 1, the near-far amplitude attenuation ratio, the near-far conduction velocity decrease value, and the near-far latency extension ratio are calculated based on the near-far paired far-far detection term and the near-far paired near-far detection term.

[0059] In practical implementation, the proximal-distal amplitude attenuation ratio is calculated by dividing the proximal amplitude measurement by the distal amplitude measurement, used to characterize whether the evoked response obtained from proximal stimulation shows a significant attenuation compared to the evoked response obtained from distal stimulation. The proximal-distal conduction velocity decrease value is calculated by subtracting the proximal related conduction velocity measurement from the distal related conduction velocity measurement, used to characterize whether the proximal segment conduction velocity decreases relatively. The proximal-distal latency prolongation ratio is calculated by dividing the proximal latency measurement by the distal latency measurement, used to characterize whether the latency corresponding to a longer conduction path shows an abnormal prolongation. When calculating the ratio, a numerical stability constant ε is added to the denominator as a minimum constant to prevent the denominator from being zero; a value of ε can be taken as... .

[0060] In one embodiment, during the left median nerve motor conduction detection, the amplitude at the wrist stimulation site was 8.0 mV and the amplitude at the elbow stimulation site was 4.0 mV. The amplitude attenuation ratio between the proximal and distal ends was 4.0 / (8.0+ε), which is approximately 0.5, indicating that the amplitude decreased significantly relative to the distal end after proximal stimulation. If the conduction velocity from the wrist to the elbow was 55 m / s, while the conduction velocity of the more proximal segment was 40 m / s, the decrease in conduction velocity between the proximal and distal ends was 15 m / s, indicating that the decrease in local conduction velocity was more significant.

[0061] 3) When the opposite side is paired with a valid marker When the value is 1, the amplitude ratio, conduction velocity difference, and latency difference of the opposite side are calculated based on the i-th detection item and the paired detection item on the opposite side.

[0062] In practical implementation, the contralateral amplitude ratio is calculated by dividing the amplitude measurement value of the i-th test item by the amplitude measurement value of the contralateral paired test item. This is used to characterize the amplitude symmetry between the same segments of the left and right corresponding nerves. For example, if the amplitude of the left median nerve motor wrist stimulation is 8.0 mV, and the amplitude of its contralateral paired test item (right median nerve motor wrist stimulation) is 7.5 mV, then the contralateral amplitude ratio is 8.0 / 7.5 ≈ 1.07. The contralateral conduction velocity difference is calculated by subtracting the conduction velocity measurement value of the contralateral paired test item from the conduction velocity measurement value of the i-th test item. This is used to characterize the amplitude symmetry between the same segments of the left and right corresponding nerves. The difference in conduction velocity is as follows: For example, the conduction velocity of the left ulnar nerve motor segment from below the elbow to the wrist is 55 m / s, while the conduction velocity of its contralateral paired test item is 58 m / s. Therefore, the difference in contralateral conduction velocity is 55 - 58 = -3 m / s. The difference in contralateral latency is calculated by subtracting the latency measurement of the contralateral paired test item from the latency measurement of the i-th test item. This is used to characterize the difference in distal conduction time between the same segments of the same nerve on the left and right sides. For example, the latency of stimulation at the ankle of the left common peroneal nerve motor segment is 4.5 ms, while the latency of its contralateral paired test item is 4.1 ms. Therefore, the difference in contralateral latency is 4.5 - 4.1 = 0.4 ms.

[0063] It should be noted that the contralateral amplitude ratio can retain directional information, or it can be further converted into an absolute deviation form in engineering implementation. For example, the distance between the contralateral amplitude ratio and 1 can be used as the left-right asymmetry intensity. If it is necessary to retain the directional information of the left side increasing or decreasing relative to the right side, then the directional contralateral amplitude ratio can be used directly.

[0064] 4) Calculate the near-far amplitude attenuation ratio, near-far conduction velocity decrease, near-far latency extension ratio, contralateral amplitude ratio, contralateral conduction velocity difference, contralateral latency difference, and effective markers for near-far pairing. Pairing valid markers with the opposite side By concatenating the data in a fixed order, we obtain the contrast feature vector of the i-th detection item. Compare feature vectors This refers to a feature vector with dimensions of 1×8, used to characterize the proximal-distal difference and the left-right contralateral difference of the i-th detection item. The first six components provide paired measurement differences, while the last two components indicate whether the corresponding pairing relationship truly exists. By using valid labels as common input features, the model can avoid misinterpreting "missing pairings set to 0" as "normal pairing differences."

[0065] 5) Following the standardized order of the detection items, stack the contrast feature vectors of all detection items row by row to obtain the contrast feature matrix D. The contrast feature matrix D is a matrix composed of the contrast feature vectors of all detection items, with a dimension of L×8, which is used to provide pairwise difference information to the subsequent model.

[0066] It should be noted that the proximal-distal amplitude attenuation ratio is more sensitive to conduction block, the proximal-distal conduction velocity decrease value is more sensitive to focal demyelination, the proximal-distal latency prolongation ratio is more sensitive to conduction time prolongation, and the contralateral amplitude ratio, contralateral conduction velocity difference, and contralateral latency difference can reduce the interference of individual patient baseline differences on the interpretation results.

[0067] In a specific implementation, step S3 involves encoding the directional deviation from the grouped normal reference distribution. Since different detection items are determined by the standard nerve name, nerve type, stimulation site, and recording site, their normal reference ranges differ significantly. Motor nerve amplitude is typically measured in millivolts, while sensory nerve amplitude is typically measured in microvolts, and the abnormal directions of latency and conduction velocity also differ. Directly using a uniform normalization method would weaken the true pathological deviation; relying solely on artificial normal value tables makes it difficult for the model to stably utilize normal reference information. Therefore, this invention utilizes a historical normal population database to construct a grouped normal reference distribution and converts patient detection values ​​into directional deviation codes, enabling comparable abnormal representations for different detection items and indicators. The specific steps are as follows: S3.1 Determination of Normal Reference Grouping Index This step generates a standardized grouping index for each test item and uses this index to retrieve a set of normal reference samples that perfectly match the clinical meaning from a historical normal population database. The specific steps are as follows: 1) According to standard nerve names Neurological type Stimulation sites and recording sites Generate normal reference grouping index Normal reference grouping index It refers to the grouping identifier used to retrieve the normal reference distribution that is consistent with the clinical meaning of the i-th test item from a historical normal population database.

[0068] In practice, the terminology of standard nerve names, nerve types, stimulation sites, and recording sites is first standardized, and then assembled in a fixed order to form a normal reference grouping index. For example, "median nerve-motor nerve conduction detection-wrist stimulation site-abductor pollicis brevis recording site" forms one normal reference group index; "sural nerve-sensory nerve conduction detection-lower leg stimulation site-lateral malleolus recording site" forms another normal reference group index. Based on this, historical normal samples within the same normal reference group have consistent clinical reference meaning.

[0069] 2) Retrieve the normal reference group index from the historical normal population database. The corresponding historical normal sample values ​​are stored, and the distribution of historical normal samples is saved according to indicator type q. The historical normal population database refers to a database composed of normal samples with no evidence of peripheral neuropathy, qualified examination quality, and complete basic information.

[0070] In one implementation, the historical normal population database can be further stratified according to age range, height range, and limb temperature range. For example, sensory nerve amplitude can change with age, and conduction velocity can be affected by limb temperature; therefore, historical normal samples with similar age and limb temperature ranges can be prioritized for retrieval. When the number of samples after stratification is insufficient, it can then revert to the unstratified normal reference group. Based on this, the adaptability of the normal reference distribution to the current patient's physiological conditions can be improved.

[0071] In one embodiment, for example, a 65-year-old patient needs to retrieve the amplitude index of "sural nerve-sensory nerve conduction detection-lower leg stimulation site-lateral malleolus recording site" from a normal reference distribution. The system prioritizes retrieving historical normal samples from patients aged 60-70 years with limb temperatures between 32-34°C. If the number of samples within this stratum exceeds a threshold... If so, this portion of the sample is used as the normal reference distribution, so that the reference range is more in line with the patient's age and physiological state.

[0072] 3) When the normal reference grouping index When the number of historical normal samples under indicator type q is insufficient, a backup normal reference distribution is obtained according to the preset merging rules.

[0073] In practical implementation, the historical normal sample quantity threshold The value can be 100 or 200. If the number of historical normal samples under index type q for a certain normal reference grouping index is less than... If the sample size is insufficient, historical normal samples with the same standard nerve name and adjacent stimulation sites of the same nerve type will be merged first. If this is still insufficient, the overall reference distribution with the same standard nerve name and nerve type will be used. If this is still insufficient, backup reference distributions of the same type of upper limb nerve or lower limb nerve will be used. This avoids the instability of directional consistency deviation coding caused by small sample reference distributions.

[0074] In one embodiment, as an example, when calculating the conduction velocity index of "radial nerve-motor nerve conduction detection-forearm stimulation site-extensor digitorum recording site", if the number of historical normal samples under this precise group is insufficient... If there are fewer than 100 cases, normal samples from all stimulation sites (such as the upper arm and forearm) under the "radial nerve-motor nerve conduction test" will be combined as a backup distribution. If the sample size is still insufficient, the reference distribution of the broader category of "upper limb nerve-motor nerve conduction test" will be used in the end.

[0075] S3.2, Index Rank Normality Deviation Coding This step compares the patient's measurements with the rank of the corresponding normal reference distribution, and uses the inverse transformation of the standard normal distribution to convert them into comparable rank-normal quantiles. Finally, it corrects for the clinical abnormality direction of each indicator, generating a deviation code with a consistent direction. This allows indicators with different dimensions and abnormal directions to represent the degree of abnormality on a unified scale. The specific steps are as follows: 1) For the i-th detection item and indicator type q, start from the normal reference grouping index. Extract historical normal sample values ​​of the same indicator type q from the corresponding historical normal sample distribution, and then use the current original measurement value. Compare the rank with the historical normal sample values.

[0076] In practice, historical normal sample values ​​are sorted in ascending order, and the current raw measurement value is statistically analyzed. The insertion position within the ascending sequence. An earlier insertion position indicates that the current original measurement is closer to the lower end of the normal reference distribution; a later insertion position indicates that the current original measurement is closer to the higher end of the normal reference distribution. If the current original measurement is the same as multiple historical normal sample values, the average rank can be used to reduce discrete jumps caused by identical values.

[0077] In one embodiment, for example, the amplitude of the left median nerve motor wrist stimulation in a patient is 3.2 mV. The corresponding normal reference distribution contains 200 historical normal sample values ​​(unit: mV), which are arranged in ascending order as [4.5, 4.7, 4.8, ..., 12.1]. 3.2 mV is lower than all normal values, so its insertion position in the ascending sequence is the 1st position, with a rank of 1.

[0078] 2) Calculate the empirical rank probability based on the rank, and input the empirical rank probability into the inverse function of the cumulative distribution function of the standard normal distribution to obtain the rank normal quantile value. The empirical rank probability is not directly set to 0 or 1 to avoid the inverse function of the cumulative distribution function of the standard normal distribution producing an infinite value.

[0079] In one implementation, the first The rank-normal quantile of each test item under index type q is calculated by the inverse transformation of the empirical rank probability and the standard normal distribution, and is expressed as follows: ; in, This represents the rank normal quantile of the i-th detection item under index type q, used to characterize the position of the current raw measurement value relative to the normal reference distribution; It represents the inverse function of the cumulative distribution function of the standard normal distribution; Indicates the current raw measurement value Compared to normal reference grouping index The ascending rank corresponding to the historical normal sample distribution; Indicates normal reference grouping index The number of historical normal samples under indicator type q.

[0080] It should be noted that, Refers to the current original measurement value In normal reference grouping index The rank of the corresponding indicator type q among all historical normal sample values, arranged in ascending order; if the measured values ​​are repeated, the average rank is taken. For example, the rank of amplitude measurements. The number of samples corresponding to the normal reference distribution The empirical rank probability is 1 / (200+1)≈0.00498. This probability value is used as input and calculated using the inverse function of the cumulative distribution function of the standard normal distribution. .

[0081] It should also be noted that the cumulative distribution function of the standard normal distribution refers to the probability that a random variable takes a value less than or equal to a certain value in a normal distribution with a mean of 0 and a standard deviation of 1; and the inverse function of the cumulative distribution function of the standard normal distribution refers to finding the corresponding standard normal distribution quantile given a probability value.

[0082] 3) Based on the abnormal direction of indicator type q, convert the rank normal quantile value into a direction-consistent deviation code value. Direction consistency deviation coding value It refers to the deviation value after abnormal direction correction. Positive values ​​are used to indicate deviation in the abnormal direction, while negative values ​​are used to indicate deviation in the non-abnormal direction or deviation above the normal median but not in the main abnormal direction of the indicator.

[0083] In one embodiment, for example, if the amplitude of a certain detection item is Because its abnormal direction is downward, If the incubation period of a certain test item is Because its abnormal direction is elongation, Both are positive, and the model interprets this uniformly as "an anomaly exists".

[0084] In practical implementation, amplitude and propagation velocity indices typically exhibit a decreasing trend as the primary anomaly; therefore, the directional deviation coding value... Take the negative of the rank-normal quantile; the latency index type and the F-wave latency index type usually show prolonged latency as the main abnormal direction, therefore the directional deviation coding value We directly take the rank-normal quantile value. That is, when the amplitude is significantly lower than the normal reference distribution, the directional deviation value is encoded. The value is positive; when the latency period is significantly longer than the normal reference distribution, the directional deviation coding value is positive. It is also positive. Through this directional correction, the subsequent model (referring to the overall diagnostic model composed of the relation-aware heterogeneous graph attention network and the fully connected classification head) can uniformly understand the abnormal trends of different indicators as "the larger the positive value, the stronger the abnormal tendency".

[0085] 4) For indicator type q that is not present in the i-th test item or is not clinically applicable, assign the corresponding directional deviation code value. Setting it to 0 indicates that the indicator type is at the median level of the normal reference distribution, and does not provide anomaly alerts to the model. For indicator type q that has a measured value but exceeds the range of historical normal samples, it is encoded using the minimum or maximum rank, and the directional deviation encoding value is clipped to a preset range, such as [-4,4], to preserve obvious abnormal trends while avoiding extreme values ​​dominating model training.

[0086] In one embodiment, for example, the amplitude of the wrist stimulation site for left median nerve motor conduction detection is 3.2 mV. If this amplitude is located at the low end of the corresponding normal reference distribution, the rank-normal quantile is negative. Since a decrease in amplitude is an abnormal direction, taking the opposite number yields a positive direction consistency deviation encoding value, allowing the model to directly identify that the amplitude is lower than the normal reference distribution. If the latency of the same detection item is 6.0 ms and is higher than the high end of the normal reference distribution, the rank-normal quantile is positive. Since a prolonged latency is an abnormal direction, the direction consistency deviation encoding value is still positive, thus forming a consistent abnormal direction expression with the decrease in amplitude.

[0087] S3.3 Construction of the Direction Consistency Deviation Matrix This step concatenates the directional deviation codes of each detection item into a vector in a fixed order, and then stacks the vectors of all detection items into a matrix in a standard order. This provides the model with a global view of anomaly severity that is dimensionally uniform, has consistent anomaly directions, and is comparable across detection items. The main steps are as follows: 1) Following a fixed order of amplitude index type, propagation velocity index type, latency index type, and F-wave latency index type, concatenate the multiple directional consistency deviation codes of the i-th detection item into a directional consistency deviation code vector. Direction-consistent deviation encoding vector It refers to a vector that centrally represents the degree of deviation of the i-th detection item from the normal reference distribution under each index type, with a dimension of 1×4.

[0088] 2) Following the standardized order of the detection items, stack the direction consistency deviation encoding vectors of all detection items row by row to obtain the direction consistency deviation matrix S. The direction consistency deviation matrix S is a matrix composed of the direction consistency deviation encoding vectors of all detection items, with a dimension of L×4, which is used to provide the subsequent model with information on the degree of anomaly that can be compared across detection items and across indicators.

[0089] It should be noted that this invention does not simply use the upper and lower limits of artificial normal values ​​to determine whether something is normal or abnormal. Instead, it uses the rank-normal quantiles of the grouped normal reference distribution to describe the degree of continuous deviation, and further performs directional correction based on the direction of the abnormality. Based on this, it can both preserve the positional relationship of the patient's measurements relative to the normal population and ensure that the abnormal directions of different indicators remain consistent in the model input.

[0090] In one embodiment, such as Figure 2 The image shows a heatmap corresponding to the directional consistency deviation matrix. The rows of the heatmap represent 10 detection items, and the columns represent 4 indicators (amplitude, conduction velocity, latency, and F-wave latency). The color represents the directional consistency deviation code value; a positive value indicates a deviation in an abnormal direction, and a negative value indicates a deviation in a non-abnormal direction or within the normal range. Figure 2The horizontal axis represents the index type (unitless), the vertical axis represents the test item, and the color bar value represents the standard normal quantile value after direction correction (dimensionless).

[0091] In a specific implementation, step S4 involves constructing a neural conduction anatomical relationship map. Since neural conduction detection items simultaneously include raw measurements, pairwise contrast differences, normal reference deviations, and anatomical attribute information, the numerical scales and semantic levels of features from different sources differ. Directly concatenating these features into the input model can easily lead to inconsistencies in dimensions and a loss of anatomical location meaning. Therefore, this invention maps robustly scaled raw measurements, contrast feature vectors, directional consistency deviation encoding vectors, and effective indicator labels to the same hidden dimension, and constructs a neural conduction anatomical relationship map based on the anatomical relationships between detection items. The specific steps are as follows: S4.1 Robust scaling of original measurements This step utilizes grouped quantile information to uniformly scale the original measurement values ​​(mV, m / s, ms) with different dimensions and ranges to the [0,1] interval, generating robust scaling values. This eliminates the influence of dimensions and suppresses the interference of extreme outliers while preserving their relative normal range position information, providing a numerical basis for multi-source feature fusion. The specific steps are as follows: 1) Regarding the first Each detection item and indicator type q is used to obtain a normal reference group index from normal samples in the training set. Corresponding lower quantile reference value and high percentile reference value .

[0092] Among them, the lower quantile reference value Used to define a robust scaling lower bound, which can be the 1st percentile of historical normal samples; higher percentile reference value. To define a robust upper bound for scaling, the 99th percentile of historical normal samples can be used. Using quantiles as scaling boundaries can reduce the impact of extreme values ​​on the numerical scale.

[0093] 2) The original measurement values Converted to a robust scaling value within the [0,1] interval according to the quantile range of the corresponding normal reference group. Robust scaling value It refers to the value of the i-th test item after grouping and scaling by quantiles under index type q, which is used to preserve the relative position of the original measurement value within the normal reference range.

[0094] In the actual implementation, the original measurement value is calculated first. Compared to the lower quantile reference value and high percentile reference value The positional proportions are then used to truncate results less than 0 to 0 and results greater than 1 to 1. If the denominator is too small, a numerical stability constant ε is added to the denominator, where ε can be taken as... .

[0095] In one embodiment, for example, the original measured value of the amplitude of a certain detection item is 3.2 mV, the 1st percentile (lower quantile reference value) of its normal reference group is 5.0 mV, and the 99th percentile (higher quantile reference value) is 15.0 mV; the position ratio = (3.2 - 5.0) / (15.0 - 5.0 + ... The value is -0.18, which is less than 0 and is truncated to 0. Therefore, the robust scaling value... This indicates that the amplitude is below the lower end of the normal reference range.

[0096] 3) For indicator types q that are absent or clinically inapplicable in the i-th test item, the corresponding robust scaling value will be applied. Setting it to 0.5 indicates that the indicator type is in the middle of the robust scaling range, so that missing indicators do not produce numerical implications that are too low or too high.

[0097] 4) Following a fixed order of amplitude index type, propagation velocity index type, latency index type, and F-wave latency index type, the robust scaling values ​​of the i-th detection item are concatenated into a robust scaling vector. Robust scaling vectors It refers to a vector that centrally represents the robust scale position of the original measurement values ​​of each indicator of the i-th detection item, with a dimension of 1×4.

[0098] S4.2, Feature Fusion of Multi-Granularity Detection Items This step concatenates and linearly projects features of various granularities and semantics, including robust scaling of the original measurements, contrast features, deviation encoding, and anatomical attribute embeddings, into a comprehensive feature vector with unified dimensions and complete information. This provides a complete initial feature representation for constructing graph network nodes. The specific steps are as follows: 1) Robustly scale the vector of the i-th detection item Comparison of feature vectors Direction consistency deviation encoding vector and indicator effective label vector By concatenating the features along their dimensions in a fixed order, we obtain multi-granularity numerical feature vectors. Multi-granularity numerical feature vectors The dimension is 1×20, including 4 robust scaling components, 8 contrastive feature components, 4 orientation consistency deviation encoding components, and 4 index effective labeling components.

[0099] Among them, the effective label vector of the indicator It refers to a vector composed of the effective labels of the i-th detection item under the wave amplitude index type, transmission velocity index type, latency index type and F wave latency index type, with a dimension of 1×4.

[0100] 2) Multi-granularity numerical feature vectors The input linear projection layer is mapped to a unified hidden dimension d, resulting in a numerical fusion representation. The unified hidden dimension d controls the length of the feature representations within the model, and can be 64, 128, or 256. The linear projection layer includes a feature fusion projection matrix. and feature fusion bias vector Feature fusion projection matrix and feature fusion bias vector All of these are trainable parameters used to bring features from different sources into the same representation space.

[0101] In one implementation, the linear projection layer is a fully connected layer used to transform multi-granularity numerical feature vectors of dimension 1×20. Mapped to a unified hidden space of dimension 1×d, its input is the multi-granularity numerical feature vector of the i-th detection item. The dimension is 1×20, and the output is the numerical fusion representation of the i-th detection item, which has a dimension of 1×d.

[0102] 3) These are the standard nerve names. , left and right sides Neurological type Stimulation sites and recording sites Establish a trainable embedding table and read the nerve name embedding vector, left and right side embedding vector, nerve type embedding vector, stimulus site embedding vector, and recording site embedding vector from the corresponding embedding table.

[0103] Among them, the nerve name embedding vector is used to represent the differences in the anatomical origin of different nerves; the left and right lateral embedding vector is used to represent the lateral information of the left and right detection items; the nerve type embedding vector is used to represent the semantic differences of motor nerve conduction detection, sensory nerve conduction detection and F wave detection; the stimulus site embedding vector is used to represent the anatomical segment information corresponding to the stimulus location; and the recording site embedding vector is used to represent the anatomical meaning of the response signal acquisition location.

[0104] It should be noted that the trainable embedding table is a trainable lookup table matrix; for example, the dimension of the neuron name embedding table is [dimension number missing]. ,in, It is the total number of standard nerve names. This refers to the dimension of the embedding vector. During a lookup, inputting an integer index of a neuron name will return the corresponding row in the table. An embedding vector. For example, the index of "median nerve" is 0, and the index of "ulnar nerve" is 1; when reading, the vector is taken from the 0th row of the nerve name embedding table as the embedding representation of "median nerve".

[0105] 4) Generate anatomical sequence position encoding vectors based on the preset anatomical segment sequence table. Anatomical sequence position encoding vector Used to characterize the proximal-distal order of the i-th detection item in the same neural stimulation site sequence.

[0106] In practical implementation, for multiple stimulation sites under the same standard nerve name, same left / right side, same nerve type, and same recording site, segment numbers are assigned in descending order from distal to proximal. For example, in median nerve motor conduction detection, the wrist stimulation site, elbow stimulation site, and axillary stimulation site can be assigned as segments 1, 2, and 3, respectively. Anatomical sequence location encoding vector Either learnable positional coding or sinusoidal positional coding can be used. When the examination protocol is relatively fixed, learnable positional coding is more convenient for learning differences in specific segments; when the combination of patient test items varies greatly, sinusoidal positional coding is more convenient for handling segment numbers that do not appear regularly.

[0107] In one embodiment, for example, learnable location encoding is employed; for median nerve motor conduction, the embedding vector of the "wrist stimulation site" is set as follows: "Elbow stimulation site" is "Axillary stimulation sites" are Regardless of which items are included in the patient's actual examination, the anatomical sequence encoding vectors corresponding to these stimulation sites are fixed, and the model learns these three vectors to distinguish different segments.

[0108] 5) Add the numerical fusion representation, nerve name embedding vector, left / right side embedding vector, nerve type embedding vector, stimulus site embedding vector, recording site embedding vector, and anatomical order position encoding vector to obtain the comprehensive feature vector of the i-th detection item. Comprehensive feature vector of detection items It refers to a vector that simultaneously contains measurement values, pairwise comparison differences, deviation from normal reference, missing status, and anatomical attribute information, with a dimension of 1×d.

[0109] 6) Following the standardized order of the detection items, stack the comprehensive feature vectors of all detection items row by row to form the detection item feature sequence X. The detection item feature sequence X is a sequence matrix composed of the comprehensive feature vectors of all detection items, with a dimension of L×d, which is used as the node input features of the relation-aware heterogeneous graph attention network.

[0110] S4.3 Construction of the Anatomical Relationship Diagram of Nerve Conduction This step defines each test item as a node, and establishes various types of edges between nodes based on the identified pairing relationships, the preset anatomical order, and nerve fiber connections. This transforms the unstructured set of test items into a structured heterogeneous graph, which is used to encode the anatomical and pathophysiological relationships between the items relied upon for clinical diagnosis. The specific steps are as follows: 1) Construct a node set Ve, where each detection item is treated as a node. The i-th node in the node set Ve corresponds to the i-th detection item, and the initial feature of the node is the comprehensive feature vector of the detection item. .

[0111] 2) Based on the index of remote detection items paired with near-far ends Pairing of proximal and distal detection items index Establish a set of near and far paired edges When the proximal and distal ends are paired, a valid marker is used. When the value is 1, a near-far pairing edge is established between the far-far pairing detection item and the near-far pairing detection item.

[0112] In the specific implementation, the set of detection items Ja is traversed. For detection item i, if its near and far end pairings are valid, it is considered as a valid marker. Then, starting from the detection item index i, pair the detection item indices to its nearest neighbor. Create a directed edge (or treat it as an undirected edge), and add the edge and its type identifier to the set. For example, the left median nerve motor wrist stimulation test item i. If the elbow stimulus detection term j is pointed to, then a near-far pairing edge is created between nodes i and j.

[0113] It should be noted that the proximal-distal pairing edge is used to connect the detection items corresponding to the distal and proximal stimuli at the same recording site of the same nerve, so that subsequent message transmission can explicitly utilize information such as amplitude attenuation, conduction velocity reduction, and latency prolongation.

[0114] 3) Based on the paired detection item index on the opposite side Establish the set of paired edges on opposite sides When the opposite side is paired with a valid marker. When the value is 1, a pairing edge is established between the i-th detection item and the opposite pairing detection item.

[0115] In the specific implementation, the set of detection items Ja is traversed. For detection item i, if its opposite pair is a valid tag... And the index of paired detection items on the opposite side. If the pointer points to the detection item k, then a pairing edge is created between nodes i and k, and added to the pairing edge. For example, the left ulnar nerve motor wrist test item i. If the wrist test item k points to the right ulnar nerve motor region, then a contralateral pairing is established between the two.

[0116] It should be noted that the contralateral pairing edge is used to connect the same segment of the same nerve on the left and right sides, so that subsequent message transmission can explicitly utilize the information of left-right symmetry; for unilateral focal lesions, the contralateral pairing edge can help the model determine whether the abnormality of a certain test item exceeds the patient's own contralateral baseline.

[0117] 4) Establish a set of adjacent segments of the same nerve based on the pre-set anatomical segment sequence table. Specifically, for multiple stimulation sites under the same standard nerve name, same left / right side, same nerve type, and same recording site, they are sorted in order from distal to proximal, and adjacent segments of the same nerve are established between corresponding detection items of adjacent stimulation sites.

[0118] In one embodiment, as an example, for the test item "left median nerve-motor-abductor pollicis brevis", it is ordered in the order of "wrist", "elbow", and "axillary". If there are three corresponding nodes a, b, and c, then two adjacent segments of the same nerve are established between adjacent node pairs (a, b) and (b, c), and added... .

[0119] It should be noted that the edges between adjacent segments of the same nerve are used to characterize the continuous trend of the lesion along the nerve conduction path. For example, the continuous edges between the wrist stimulation site and the elbow stimulation site, and between the elbow stimulation site and the axillary stimulation site, can help the model identify whether the abnormality is limited to a certain segment or persists along a longer conduction path.

[0120] 5) Establish a set of edges connecting nerve fibers with the same name Specifically, for motor nerve conduction test items and sensory nerve conduction test items under the same standard nerve name and the same left or right side, if the corresponding anatomical segments are clinically related, then a nerve fiber association edge with the same name is established between the corresponding test items.

[0121] In practical implementation, motor conduction and sensory conduction test items with the same nerve name and side are distinguished. For example, if the left median nerve motor conduction wrist test item 'i' and the left median nerve sensory conduction wrist test item 'm' belong to adjacent segments of the same nerve anatomically, then a corresponding nerve fiber association edge is established between 'i' and 'm', and added... .

[0122] It should be noted that the association edges of nerve fibers with the same name are used to characterize the combined involvement of motor and sensory fibers of the same nerve. By associating the motor and sensory functional states of the same nerve, it helps to distinguish between isolated motor involvement, isolated sensory involvement, and mixed motor and sensory involvement. For example, when abnormal motor nerve conduction tests of the median nerve are accompanied by abnormal sensory nerve conduction tests of the median nerve, it suggests more extensive median nerve involvement; if only distal sensory tests are abnormal, it may suggest early sensory fiber involvement or focal entrapment.

[0123] 6) Establish self-loop edges for each detection item to obtain the set of self-loop edges. Self-loop edges are used to preserve the contribution of the detection item's own features to the node update process, and to prevent the node features from being completely covered by neighbor information during the relationship message passing process.

[0124] It should be noted that a self-loop edge is an edge that points from a node to itself. This is used when constructing the edge set. At that time, it is a set of nodes. For each node i in the algorithm, an edge is created pointing from node i to node i. For example, the node representing "left median nerve motor wrist stimulation" has a self-loop edge connected to itself, ensuring that its own features can be effectively transmitted even when it has no neighbors or when neighbor information is not important.

[0125] 7) Set the pairs of near and far ends. , set of paired edges on opposite sides Adjacent segments of the same nerve are grouped together. Set of edges associated with the same name of nerve fibers and self-loop edge set Merging these elements yields the neural conduction anatomy graph Gra, which is a graph structure used to describe the structured clinical relationships between test items, including the set of nodes Ve and the set of edges Ev.

[0126] In the actual implementation, each edge stores an edge type identifier. Edge type identifier This is used to distinguish whether the relationship between the i-th and j-th test items is a proximal-distal pairing edge, a contralateral pairing edge, an edge between adjacent segments of the same nerve, an edge connecting nerve fibers of the same name, or a self-loop edge. If the same pair of test items satisfies multiple relationships, it is preferable to retain multiple edges with different edge type labels to avoid losing semantic differences after different clinical relationships are merged.

[0127] In one embodiment, as an example, for the three detection items of wrist stimulation site, elbow stimulation site and axillary stimulation site for left median nerve motor nerve conduction detection, adjacent segments of the same nerve can be established from the wrist stimulation site to the elbow stimulation site and from the elbow stimulation site to the axillary stimulation site. At the same time, proximal and distal pairings can be established between the wrist stimulation site and the elbow stimulation site. If there is a right median nerve motor nerve conduction detection item for wrist stimulation site, then contralateral pairings can be established between the left and right wrist stimulation site detection items. If there is also a corresponding segment detection item for left median nerve sensory nerve conduction detection, then corresponding nerve fiber associations can be established.

[0128] In one embodiment, such as Figure 3 The diagram illustrates the anatomical relationships of nerve conduction, with nodes representing each test item. Node labels include the nerve abbreviation, laterality, type, and stimulation site. Specifically, the first row contains the nerve abbreviation (e.g., "median," "ulnar," etc., using the first two Chinese characters of the standard nerve name); the second row contains the first character of the laterality ("left" or "right") plus the first character of the nerve type (e.g., "left lateral" or "right lateral"); and the third row contains the stimulation site (e.g., "wrist," "elbow," etc.). Edge types are distinguished by different colors: red for proximal-distal pairings, blue for contralateral pairings, and orange for edges associated with the same nerve fiber. This diagram constructs the clinical anatomical relationships between test items, demonstrating that the unstructured list of test items has been organized into a heterogeneous diagram containing multiple relationship types.

[0129] In one embodiment, such as Figure 4 As shown, robust scaling values ​​are displayed after mapping the original measurements to the [0,1] interval using grouped quantiles. The horizontal axis represents amplitude scaling, conduction velocity scaling, latency scaling, and F-wave latency scaling, while the vertical axis represents the detection item number. Quantile scaling of the original measurements preserves their relative position in the normal population while eliminating dimensional differences and extreme value interference, allowing different indicators to be fused in the same numerical space and providing a numerical basis for multi-granularity feature stitching.

[0130] In a specific implementation, step S5 involves encoding a neural conduction graph based on a relation-aware heterogeneous graph attention network. Since the interpretation of neural conduction test results depends not only on individual test values ​​but also on the clinical relationships between test items, a significant decrease in proximal amplitude compared to distal amplitude may indicate conduction block, significant differences between the left and right contralateral segments of the same nerve may indicate unilateral involvement, and the trend of changes in adjacent segments of the same nerve may indicate lesion localization. Therefore, this invention introduces edge type embedding in graph attention calculation, enabling the relation-aware heterogeneous graph attention network to distinguish between proximal-distal paired edges, contralateral paired edges, adjacent segments of the same nerve, edges associated with the same nerve fiber, and self-loop edges during message transmission. The specific steps are as follows: S5.1 Relationship-Aware Attention Score Calculation This step introduces trainable edge-type embeddings into the graph attention mechanism, enabling the model to distinguish and learn the differentiated impact of different clinical relationships (such as proximal-distal pairing, contralateral pairing, etc.) on information transmission when calculating attention scores between nodes. This allows for more accurate aggregation of neighborhood information relevant to diagnosis. The specific steps are as follows: 1) Use the feature sequence X of the detection item as the feature matrix of the graph encoding node of the 0th layer of the relation-aware heterogeneous graph attention network. .

[0131] The relation-aware heterogeneous graph attention network consists of K relation-aware heterogeneous graph attention layers, where K represents the number of network layers, which can be 2 or 3; the node feature vector of the i-th detection item in the l-th layer is denoted as... Furthermore, l represents the layer index; for the initial input features, l = 0; the output of the l-th layer relation-aware heterogeneous graph attention layer is the node features of the (l+1)-th layer. The total number of network layers is K, and the value of l ranges from 0 to K-1. The final output is... .

[0132] It should be noted that the relation-aware heterogeneous graph attention layer is the basic computational unit of graph neural networks. Its core is to take the type of the edge connecting two nodes (i.e., clinical relation) as input to participate in the calculation of attention score on the basis of the standard multi-head graph attention mechanism, so that the information aggregation process can "perceive" different relational semantics.

[0133] 2) For the h-th attention head, perform linear projections on the node feature vectors of the target detection item and the node feature vectors of the neighboring detection items to obtain the query representation, key representation, and value representation. h represents the attention head index, ranging from 1 to H; H represents the number of attention heads, which can be 4 or 8. The query projection matrix, key projection matrix, and value projection matrix are all trainable parameters.

[0134] It should be noted that the attention head is a parallel computation channel in the multi-head attention mechanism. A complete multi-head attention module consists of H structurally identical but parameter-independent "heads". Each head independently learns a set of projection parameters for queries, keys, and values, allowing the model to capture the association patterns between nodes from different perspectives.

[0135] In the specific implementation, the query represents what type of neighborhood information the target detection item needs to receive; the key represents the relationship matching information that the neighbor detection item can provide; and the value represents the message content that the neighbor detection item actually passes to the target detection item.

[0136] 3) Based on the edge type identifier between the i-th detection item and the j-th neighbor detection item. Read the edge type embedding vector from the edge type embedding table. Edge type embedding vector It refers to a trainable vector used to characterize the semantic differences of different clinical relationships.

[0137] In practical implementation, the edge type embedding vectors corresponding to the proximal and distal paired edges make the model pay more attention to amplitude attenuation and segmental changes in conduction velocity; the edge type embedding vectors corresponding to the contralateral paired edges make the model pay more attention to the disruption of left-right symmetry; the edge type embedding vectors corresponding to the adjacent segments of the same nerve make the model pay more attention to the continuous changes of the lesion along the anatomical path; the edge type embedding vectors corresponding to the edges associated with the same nerve fibers make the model pay more attention to the joint involvement between motor fibers and sensory fibers; and the edge type embedding vectors corresponding to the self-loop edges make the model retain the evidence of individual detection values.

[0138] In one implementation, an edge type embedding table (a trainable parameter matrix) is first established, with the number of rows representing the total number of edge types (5 in this example) and the number of columns representing the embedding dimension; each edge type... Each vector corresponds to an integer index; by searching the table using that index, the corresponding embedding vector can be retrieved. For example, the index of the "far-near pairing edge" is 0, and the index of the "opposite pairing edge" is 1; when calculating the attention between two far-near nodes, the embedding vector corresponding to index 0 is read to participate in the calculation.

[0139] 4) Input the query representation, key representation, and edge type embedding vector into the attention scoring function to obtain the relation-aware attention score. Relationship-aware attention score It refers to the correlation score between the i-th detection item and the j-th neighbor detection item in the h-th attention head of the l-th layer, which is used to measure the contribution strength of the neighbor detection item to the current detection item update process.

[0140] In one implementation, relation-aware attention scoring It can be calculated by splicing scores, and expressed as follows: ; in, This represents the query projection matrix of the h-th attention head; This represents the key projection matrix of the h-th attention head; The edge type projection matrix represents the h-th attention head; Represents the scoring parameter vector for the h-th attention head; [ | | ] indicates feature concatenation operation; LeakyReLU( ) represents the linear unit activation function with leakage correction, used to preserve negative score information and improve training stability.

[0141] S5.2, Neighbor Attention Weight Normalization This step calculates the relationship-aware scores for all neighbors of each node, performs Softmax exponential normalization, and converts them into attention weights that sum to 1, so as to aggregate neighbor information proportionally. The specific steps are as follows: 1) For the i-th detection item, obtain the set of neighboring detection items in the neural conduction anatomy graph Gra. Neighbor detection item set This includes detection items connected to the i-th detection item via proximal-distal pairing edges, contralateral pairing edges, adjacent segments of the same nerve, nerve fiber association edges of the same name, and self-loop edges.

[0142] In one embodiment, as an example, for the detection item i of left median nerve motor elbow stimulation, its neighbor set It may include: wrist stimulation detection items connected by proximal-distal paired edges, right elbow stimulation detection items connected by contralateral paired edges, axillary stimulation detection items connected by adjacent segment edges, and detection item i itself (self-loop edge).

[0143] 2) Perform exponential normalization on the relation-aware attention scores of all neighboring detection items for the i-th detection item to obtain the attention weights. Attention weight This refers to the weight coefficient of the i-th detection item in the h-th attention head of layer l when it receives the message of the j-th neighbor detection item. Through normalization, the sum of the weights of all neighbor detection items for the same target detection item is 1.

[0144] In practical implementation, if a detection item has multiple neighboring detection items, the relation-aware heterogeneous graph attention network will jointly determine which neighboring detection items are more important based on node features and edge types. For example, if there is a significant difference in amplitude attenuation between a detection item and its near and far paired detection items, the neighboring detection items corresponding to the near and far paired edges may receive higher attention weights; if the difference between the left and right sides of the same detection item is not significant, the neighboring detection items corresponding to the paired edges on the opposite side may have lower weights.

[0145] It should be noted that exponential normalization is a standard operation in deep learning, used to convert a set of real scores into weights in the form of a probability distribution that sums to 1.

[0146] S5.3, Relationship Message Aggregation and Node Feature Update This step utilizes the attention weights obtained in the previous step to perform a weighted summation of the features of all neighboring nodes, thereby aggregating the relational messages. It then updates the features of the current node through mechanisms such as residual connections and layer normalization, integrating neighborhood information. The specific steps are as follows: 1) For the h-th attention head, project the node feature vectors of the neighbor detection terms through the value projection matrix. Convert to a neighbor message representation. Specifically, for neighbor node j, convert its l-th layer node feature vector. Projection matrix of the h-th head value Multiplying them together yields information about the neighboring household. .in, The node feature vector of the j-th neighbor detection item in the l-th layer is represented by the value projection matrix. is the trainable parameter of the h-th attention head, used to extract the message content passed from the neighbor detection item to the target detection item.

[0147] 2) Based on attention weights We perform a weighted summation on the neighbor messages to obtain the relation aggregation result under the h-th attention head, which is represented as: ; in, This represents the relation aggregation result obtained by the h-th attention head of the l-th layer for the i-th detection item.

[0148] In practice, the relationship aggregation results enable the proximal-far end pairing detection items, contralateral pairing detection items, adjacent segment detection items of the same nerve, nerve fiber association detection items of the same name, and the current detection item itself to participate in node updates according to different contribution ratios.

[0149] 3) Concatenate the relationships of multiple attention heads to obtain the candidate node update features for the i-th detection item in layer (l+1). Then, using the standard operations of the Transformer architecture and deep networks in deep learning, specifically performing residual connections, layer normalization, and non-linear activation on the candidate node update features, we obtain the node feature vector for the i-th detection item in layer (l+1). .

[0150] Among them, residual connection refers to adding the input of a module to its output to solve the gradient vanishing problem in deep networks; layer normalization refers to standardizing the activation values ​​of each neuron in a single sample to stabilize the training process; nonlinear activation can use functions such as ReLU activation function to introduce nonlinear expressive power into the model.

[0151] It should be noted that residual connections are used to preserve the original node features of the i-th detection item in the l-th layer, avoiding excessive smoothing of single detection evidence after multi-layer message passing; layer normalization is used to stabilize the feature scale fluctuations caused by changes in the number of detection items for different patients; and nonlinear activation is used to enhance the ability of the relation-aware heterogeneous graph attention network to express complex pathological patterns.

[0152] S5.4, Multi-layer Graph Coding Feature Output This step stacks multiple layers of relation-aware heterogeneous graph attention layers, enabling each detection node to progressively aggregate information from multi-hop neighbors. The final output is a graph-coded node feature matrix rich in anatomical relationships and contextual information, which can be used by downstream tasks. The specific steps are as follows: 1) Stack the relation-aware heterogeneous graph attention layer with K layers to enable the detection item to complete information transmission within the scope of multi-hop anatomical relationships. In one implementation, when K=2, the detection item can not only receive information from directly paired detection items, but also indirectly obtain the changing trends of more distant segments through the relationship between adjacent segments.

[0153] 2) Output the feature matrix of the Kth layer graph encoding nodes The feature matrix of the Kth layer graph encoding nodes It refers to the feature matrix of the detection item after being encoded by the relation-aware heterogeneous graph attention network. It has a dimension of L×d and is used to represent the state of the detection item after fusing single measurement information, pairwise comparison relationships and dissecting neighborhood relationships.

[0154] It should be noted that the relation-aware heterogeneous graph attention network is composed of K layers of relation-aware heterogeneous graph attention layers stacked together. Taking one of the layers as an example, its hierarchical structure is shown in Table 1. Table 1. Structural Relationships of Relation-Aware Heterogeneous Graph Attention Networks It should also be noted that the relation-aware heterogeneous graph attention network of this invention does not allow completely free interaction between arbitrary detection items. Instead, it first limits the scope of message transmission through a neural conduction anatomical relationship graph, and then distinguishes different clinical relationships through edge type embedding. Based on this, the risk of spurious relevance learning can be reduced, and the interpretability of the model output results can be enhanced.

[0155] In a specific implementation, step S6 involves patient-level global feature generation and disease category prediction, since the Kth layer graph encoding node feature matrix... While individual detection information and structured relationship information have been integrated, different detection items contribute differently to disease diagnosis. Average pooling dilutes key lesion segments, max pooling easily overlooks the cumulative evidence of multiple mild to moderate abnormalities, and ordinary attentional pooling cannot explicitly highlight locations with significant paired differences. Therefore, this invention incorporates the graph coding feature differences between detection items and paired detection items into the pooling weight calculation, enabling the model to prioritize key abnormal patterns such as abrupt drops in amplitude at proximal and distal ends, focal decreases in conduction velocity, and left-right contralateral asymmetry. The specific steps are as follows: S6.1 Construction of Paired Difference Enhancement Features This step generates a paired differential enhancement feature for each test item. By subtracting its own graph encoding feature from the feature of the clinical paired item, it highlights the differential information of individual test items under key clinical relationships (such as proximal / distal, left / right). The specific steps are as follows: 1) Select a pooling pair for the i-th test item. Specifically, pooling pairing tests are preferentially selected from proximal and distal pairs because conduction block and segmental demyelination usually depend on the proximal-distal difference for interpretation; when no proximal-distal pairing test item exists but a contralateral pairing test item exists, the contralateral pairing test item is selected; when neither proximal-distal nor contralateral pairing test items exist, the pooling pairing is marked as valid. Set to 0.

[0156] Pooled pair detection item index is denoted as , Used to point to the detection item for pooling difference comparison with the i-th detection item; pooling pair valid marker. This is used to control whether paired difference features participate in pooling scoring.

[0157] 2) When pooling pairing is validly marked When the value is 1, read the feature matrix of the graph encoding node of the i-th detection item in the K-th layer. Node feature vectors And read the node feature vector of the pooled pair detection item. The difference between the two is calculated, and this difference is used to characterize the graph coding feature difference between the current test item and the clinical paired test item.

[0158] In one embodiment, as an example, the detection item i Paired detection items with pooling of If the difference is significant (e.g., there is conduction blockage), the absolute value of some dimensions of the difference vector will be large, sending a strong signal to the model.

[0159] Furthermore, when pooling pairing is effectively marked When the value is 0, the difference part is set to an all-zero vector, so that detection items without pairing relationships can still participate in pooling based on their own node features, but will not generate false pairing difference signals.

[0160] In one embodiment, as an example, for certain test items such as "sural nerve sensory conduction" that do not have contralateral or proximal-distal pairings, their pooling pairings are effectively labeled. At this point, the paired difference enhancement feature vector of this detection item... From its own and a with It is composed of zero vectors of equal dimension, so that the model knows that there is no pairing difference information for this item.

[0161] 3) Concatenate the node feature vector of the i-th detection item with the paired difference feature vector to obtain the paired difference enhancement feature vector. Paired difference enhancement feature vector It refers to a vector that simultaneously represents the graph encoding state of the i-th detection item itself and the difference state of its relative paired detection item, with a dimension of 1×2d.

[0162] S6.2, Calculation of Contrastive Perception Attention Pooling Weights This step employs a trainable pooling scoring function, taking paired difference enhancement features as input, to automatically learn and calculate the importance of each detection item to the final diagnosis. After normalization, pooling weights are generated, enabling the model to focus on detection items with significant contrast differences. The specific steps are as follows: 1) Enhance the feature vector by pairing differences Input a trainable pooling scoring function to obtain the pooling score of the i-th detection item. The pooling scoring function includes the pooling projection matrix. Pooling bias vector And the pooling scoring vector u. Where, the pooling projection matrix... Pooling bias vector is used to extract diagnostic information from paired differential enhancement feature vectors. The baseline of the pooling projection result is used to adjust the pooling score vector u, which is used to convert the nonlinear transformation result into a scalar score.

[0163] In one implementation, the pooling scoring function is a small neural network that maps a 2D vector to a scalar score. The calculation process involves: first, obtaining an intermediate representation through a linear projection layer. Then, the scoring vector u is pooled. Convert to scalar score ;in, Represents paired difference augmentation feature vectors The intermediate representation vector after linear projection layer and hyperbolic tangent activation function; tanh( ) represents the hyperbolic tangent activation function, which maps the input value to a bounded interval of (-1,1) to increase the nonlinear expressive power of the model; The pooling projection matrix is ​​a trainable parameter matrix used to extract from... Extract diagnostic information from it; This represents the pooling bias vector, which is a trainable parameter vector used to adjust the baseline after linear projection. The transpose of u is a trainable parameter vector; The pooling score of the i-th test item is a scalar value representing the initial importance of the test item to the final diagnosis.

[0164] In one embodiment, for example, suppose there are two detection items, A and B. The pairing difference feature of A is very obvious, while that of B is very slight. After calculation by the pooling scoring function, A may get a score of 3.5 and B may get a score of 0.2, which means that the model determines that A is more valuable for diagnosis.

[0165] 2) Normalize the pooling scores of all detection items to obtain the contrastive perception attention pooling weights. Compare the perceived attention pooling weights This refers to the contribution weight of the i-th detection item to the patient-level global feature vector, with a value ranging from 0 to 1, and the sum of the contrast-perception attention pooling weights of all detection items is 1. The calculation method is expressed as follows: ; in, tanh( represents the paired difference enhancement feature vector of the j-th detection item); ) represents the hyperbolic tangent function, used to map the pooling projection result to a bounded nonlinear space; j represents the summation index of the detection term, with a value ranging from 1 to L; This represents the transpose of u.

[0166] It should be noted that the contrast perception attention pooling weights are jointly determined by the single detection state and the paired difference state. Therefore, it can improve the contribution of key detection items to the patient-level global feature vector when there are focal abnormalities, left-right asymmetric abnormalities, or multi-segment mild to moderate abnormalities.

[0167] S6.3, Patient-level Global Feature Aggregation This step uses contrastive pooling weights to perform a weighted summation of the graph-coded node features of all detection items, aggregating a fixed-length patient-level global feature vector. This integrates all information at the detection item level into a patient-level diagnostic representation. The specific steps are as follows: 1) Based on the comparative perception attention pooling weight, the node feature vectors of all detection items are weighted and summed to generate a fixed-length patient-level global feature vector z. The patient-level global feature vector z is a vector used to characterize the overall diagnostic information of the current patient's neural conduction examination results, with a dimension of 1×d.

[0168] In practice, if a certain test item has significant amplitude attenuation at the proximal and distal ends, significant decrease in conduction velocity, or significant contralateral asymmetry, the paired difference enhancement feature vector corresponding to that test item is more likely to obtain a higher pooling score, thereby allowing the patient-level global feature vector z to retain evidence of key lesion segments.

[0169] 2) Input the patient-level global feature vector z into the fully connected classifier head to obtain the disease category score. In one implementation, the fully connected classifier head can adopt a single hidden layer structure, including the first layer weight matrix, the first layer bias vector, the modified linear unit activation function, the second layer weight matrix, and the second layer bias vector.

[0170] 3) Input the disease category score into the normalized exponential function (i.e., the softmax function) to obtain the disease category prediction probability vector. Disease category prediction probability vector This refers to the predicted probability vector of the current patient belonging to each disease category, with a dimension of 1×C; C represents the number of disease categories, which can be set according to the actual diagnostic task, such as including normal category, axonal damage category, demyelinating damage category, focal entrapment category, and mixed damage category.

[0171] 4) Predict the probability vector of disease categories The disease category with the highest probability is used as the preliminary interpretation result, and at the same time, the detection items with higher comparative attention pooling weights are output as the basis for interpretation.

[0172] In one embodiment, as an example, if the disease category prediction probability vector If the fourth element corresponds to the focal entrapment category, the preliminary interpretation result is the focal entrapment category. If the detection items with higher comparative attention pooling weights are concentrated at the wrist stimulation site, elbow stimulation site and their contralateral paired detection items of the left median nerve, the interpretation basis suggests that "the differences between the proximal and distal ends of the left median nerve and the differences between the left and right contralateral ends contribute significantly".

[0173] In a specific implementation, step S7 involves model training, saving fixed reference statistics, and online interpretation. The model is trained under supervision using training samples labeled with real disease categories. The specific steps are as follows: S7.1 Training Sample Construction and Class Label Encoding 1) Organize each training patient sample into a training test item set, and determine the true disease category label based on clinical diagnosis results or expert review results. The true disease category label can be set according to the specific application task, such as normal category, axonal damage category, demyelinating damage category, focal entrapment category, and mixed damage category.

[0174] 2) Encode the true disease category label of the nth training patient sample into a one-hot label vector. Where n represents the index of the training patient sample; if the nth training patient sample belongs to the cth disease category, then The c-th element has a value of 1, and the rest have a value of 0; c represents the disease category index, with a value range from 1 to C.

[0175] 3) Statistical analysis was conducted on historical normal population samples from the training set, including normal patient samples or normal population samples confirmed to have no evidence of peripheral neuropathy. The analysis included historical normal sample distributions, lower quantile reference values, and higher quantile reference values. These reference statistics were saved after the training phase and remained constant throughout the validation, testing, and online interpretation phases.

[0176] It should be noted that the historical normal population database and quantile reference statistics are not used as trainable parameters for updating, to avoid the distribution of supervised labels changing the normal reference benchmark.

[0177] S7.2 Loss Function Calculation and Parameter Update 1) For a training batch, read The disease category prediction probability vector and the true one-hot label vector for each training patient sample. This indicates the number of patient samples in the training batch, which can be 16, 32, or 64.

[0178] 2) Calculate the class weights based on the number of samples for each disease category in the training set. The number of samples for the c-th disease category is denoted as... The maximum number of samples across all disease categories is denoted as The class weight of the c-th disease category is denoted as Category weights This is used to increase the contribution of a few disease categories to the training loss and reduce the risk of the model being biased towards categories with a larger number of samples.

[0179] In practical implementation, class weights can be set. and Proportional, of which, This represents the stability constant for calculating class weights, and can be taken as... If a certain type of sample is relatively small, its class weight is larger, thus receiving more attention during the training process.

[0180] 3) Calculate the category-weighted cross-entropy loss Loss based on the real unique hot label vector and the disease category prediction probability vector. The category-weighted cross-entropy loss Loss is used to measure the difference between the model's predicted probability and the real disease category label.

[0181] 4) The gradient optimization algorithm is used to update the set of trainable parameters Θ, which includes feature fusion projection parameters, anatomical attribute embedding parameters, edge type embedding parameters, relation-aware heterogeneous graph attention network parameters, contrast-aware attention pooling parameters, and fully connected classification head parameters.

[0182] In one implementation, the Adam optimizer is used to update the set of trainable parameters Θ. The initial learning rate can be 0.001, and the number of training epochs can be 100 to 200. During training, a learning rate decay strategy can be adopted based on the decrease in validation set loss. For example, if the validation set loss does not decrease for 5 consecutive training epochs, the learning rate is multiplied by 0.5.

[0183] 5) After each training cycle, evaluate the model performance using the validation set. Specifically, calculate the F1 score for each disease category, and take the arithmetic mean of the F1 scores for all disease categories to obtain the macro-average F1 score; save the set of trainable parameters Θ corresponding to the highest macro-average F1 score as the model parameters for the final automatic interpretation of neural conduction examination data.

[0184] It should be noted that the category-weighted cross-entropy loss is used to alleviate the problem that there are many normal samples and few samples of some disease categories in the actual examination data; the macro-average F1 score is used to avoid the model performing well only on high-frequency categories and ignoring a few disease categories.

[0185] S7.3 Online Interpretation and Result Output 1) Input the new patient's neural conduction test data into the trained model and perform the same data processing procedure as in the training phase.

[0186] Specifically, the process begins with extracting data fields and standardizing terminology from patient nerve conduction studies to obtain a set of test items, Ja. Next, it involves identifying test item pairings and extracting contrastive features to obtain a contrastive feature matrix, D. Finally, it performs direction consistency deviation encoding and robust scaling to obtain a direction consistency deviation matrix, S, and a robust scaling vector. A robust scaling matrix is ​​constructed; then the feature sequence X of the detection item and the neural conduction anatomy graph Gra are constructed; finally, the graphs are passed through a relation-aware heterogeneous graph attention network, a contrast-aware attention pooling network, and a fully connected classification head to obtain the disease category prediction probability vector. .

[0187] 2) During the online interpretation phase, the historical normal population database, the reference distribution corresponding to the normal reference group index, the lower quantile reference value, and the higher quantile reference value saved during the training phase must be used. The normal reference distribution or scaling range must not be recalculated based on the neural conduction test data of a single new patient. This restriction prevents abnormalities in the test items of new patients from altering the reference baseline and affecting the stability of the interpretation.

[0188] 3) Set a confidence threshold η, which represents the minimum predicted probability required to accept the model's interpretation; it can be between 0.6 and 0.8. If the disease category prediction probability vector... If the maximum predicted probability is not less than the confidence threshold η, the disease category corresponding to the maximum predicted probability is output; if the maximum predicted probability is less than the confidence threshold η, a low confidence level prompt is output, and it is recommended to make a comprehensive judgment based on clinical manifestations, needle electromyography, or follow-up results.

[0189] In one embodiment, as an example, if the disease category prediction probability vector Given a confidence threshold η = 0.6, the maximum predicted probability is 0.63, corresponding to the 4th disease category. The 4th disease category is then output as the automatic interpretation result. If the disease category prediction probability vector... If the maximum predicted probability is 0.28, which is less than the confidence threshold η, then a low confidence warning will be output.

[0190] 4) Generate auxiliary explanatory information based on the contrast perception attention pooling weight and the relationship perception attention weight. The auxiliary explanatory information includes the detection item with the highest contribution, the pairing relationship with the highest contribution, the index type with a high abnormal direction consistency deviation coding value, and the combination of detection items with obvious differences in near and far pairing or obvious differences in opposite pairing.

[0191] In practical implementation, the model can output the top k detection items with the highest contrastive attention pooling weights, where k can be 3 to 5. For each high-contribution detection item, it outputs its standard nerve name, left / right side, nerve type, stimulation site, recording site, main abnormality index type, orientation consistency deviation encoding value, and corresponding pairing difference. Through this processing, the model not only outputs the predicted probability of the disease category but also provides key detection items for doctors to review.

[0192] Example 2 A system for interpreting and diagnosing nerve conduction test data includes a module for executing processing instructions for each step of a method for interpreting and diagnosing nerve conduction test data, comprising: The data preprocessing module is used to perform field extraction, terminology standardization, measurement validity marking, and test item sorting on the patient's nerve conduction examination data to obtain a set of test items. Each test item includes the quasi-nerve name, left and right sides, nerve type, stimulation site, recording site, and at least one measurement index. The pairing relationship identification and contrast feature generation module is used to automatically identify proximal and contralateral pairing relationships in the set of test items based on prior rules of neural conduction anatomy, and convert pairwise differences into explicit contrast features to obtain a contrast feature matrix; The directional deviation matrix construction module is used to construct a grouped normal reference distribution using a historical normal population database, convert patient test values ​​into directional deviation codes, concatenate all directional deviation code values ​​of each test item into a vector in a fixed order, and then stack all vectors of test items into a directional deviation matrix in a standard order. The multi-feature fusion and anatomical relationship map construction module is used to robustly scale the original measurement data in the detection item set to obtain robust scaling results. After mapping the results, the contrast feature matrix, the direction consistency deviation matrix and the measurement value validity label to the same hidden dimension, the module is fused to generate the detection item feature sequence and construct the neural conduction anatomical relationship map based on the anatomical relationship between the detection items. The heterogeneous graph encoding module takes the feature sequence of the detection item as input, introduces trainable edge type embedding in the graph attention mechanism, calculates the attention score between nodes through the relation-aware heterogeneous graph attention network, and outputs the graph-encoded node feature matrix. The global feature extraction and disease prediction module is used to obtain the pairing difference enhancement feature vector based on the difference between the graph coding node features of the detection item and the paired detection item, add it to the contrast perception attention pooling weight calculation, and obtain the patient-level global feature vector by weighted summation of the features of all graph coding nodes. Then, the disease category prediction probability vector is calculated based on this, and the disease category corresponding to the highest probability is selected as the output result. The model training and online interpretation module is used to train the model of the entire system using training samples with real disease category labels, and obtain a trained automatic interpretation model for nerve conduction examination data. This model is then used for online interpretation and diagnosis.

[0193] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A method for interpreting and diagnosing nerve conduction test data, characterized in that, Includes the following steps: S1. Perform field extraction, terminology standardization, measurement value validity marking, and test item sorting on the patient's nerve conduction examination data in sequence to obtain a set of test items. Each test item includes the standard nerve name, left and right sides, nerve type, stimulation site, recording site, and at least one measurement index. First, the test items are judged according to the preset diagnostic rule library for peripheral nerve injury diseases. If the test item meets the diagnostic rule of any disease, the rule diagnosis result of the disease and the corresponding judgment basis are directly output; otherwise, the subsequent operation is performed. S2. Based on the prior rules of neural conduction anatomy, automatically identify the proximal and distal pairing relationships and the contralateral pairing relationships in the set of detection items, convert the paired differences into explicit contrast features, and generate a contrast feature matrix; S3. Construct a grouped normal reference distribution using a historical normal population database, convert patient test values ​​into directional deviation codes, concatenate all the code values ​​of each test item into a vector in a fixed order, and then stack all the vectors of the test items in a standard order to form a directional deviation matrix. The construction of the directional deviation matrix in S3 is as follows: S3.1 Generate a normal reference group index based on the standard nerve name, nerve type, stimulation site and recording site. Retrieve the historical normal sample values ​​corresponding to the normal reference group index in the historical normal population database, and save the historical normal sample distribution according to the indicator type. When the number of historical normal samples under the indicator type of the normal reference group index is insufficient, obtain the backup normal reference distribution according to the preset merging rules. S3.2 For each detection item and indicator type, extract the historical normal sample value of the same indicator type from the historical normal sample distribution corresponding to the normal reference group index, and compare the current original measurement value with the historical normal sample value in order. The empirical rank probability is calculated based on the rank, and the empirical rank probability is input into the inverse function of the cumulative distribution function of the standard normal distribution to obtain the rank normal quantile value. Then, according to the abnormal direction of the indicator type, the rank normal quantile value is converted into the direction consistency deviation code value. For indicator types that do not exist in each test item or are not clinically applicable, the corresponding direction consistency deviation code value is set to 0. For indicator types that have measured values ​​but exceed the range of historical normal samples, the minimum rank or the maximum rank is used for encoding, and the direction consistency deviation code value is clipped to the preset range. S3.

3. Concatenate all directional deviation codes of each detection item into a vector in a fixed order, and then stack the vectors of all detection items into a directional deviation matrix in a standard order. S4. Robust scaling results are obtained by performing robust scaling on the original measurement data of the test item set. This result, the contrast feature matrix, the direction consistency deviation matrix, and the measurement validity label are then mapped to the same hidden dimension and fused to generate the test item feature sequence. At the same time, a neural conduction anatomical relationship map is constructed based on the anatomical relationship between the test items. The specific process of constructing the neural conduction anatomy diagram in S4 is as follows: Construct a node set with each detection item as one node; establish a set of paired edges based on the index of the far-end detection item and the index of the near-end detection item; Establish a set of paired edges on the opposite side based on the index of the paired detection items on the opposite side; Establish a set of edges for adjacent segments of the same nerve based on a pre-defined anatomical segment sequence list; establish a set of edges for nerve fibers with the same name based on whether they have a clinical relationship; establish a self-loop edge for each test item to obtain a set of self-loop edges; merge all edge sets and combine them with the node set to form a neural conduction anatomical relationship diagram; S5. Using the feature sequence of the detection item as input, a trainable edge type embedding is introduced into the graph attention mechanism. The attention score between nodes is calculated through the relation-aware heterogeneous graph attention network, and the graph-encoded node feature matrix is ​​output. S6. Based on the difference between the graph coding node features of the detection item and the paired detection item, the paired difference enhancement feature vector is obtained. It is added to the contrast perception attention pooling weight calculation. The weighted summation of all graph coding node features is used to obtain the patient-level global feature vector. Based on this, the disease category prediction probability vector is calculated, and the disease category corresponding to the highest probability is selected as the output result. S7. Train the entire model using training samples with real disease category labels to obtain a trained automatic interpretation model for neural conduction examination data. Use this model to perform online interpretation and diagnosis.

2. The method for interpreting and diagnosing nerve conduction test data according to claim 1, characterized in that, S1 is as follows: The patient's nerve conduction examination data was read and extracted, and organized into a set of test items. The measurement indicators included amplitude measurement, conduction velocity measurement, latency measurement, and F-wave latency measurement. A terminology standardization dictionary was established, and the original names in the patient's nerve conduction examination data were converted into standard terms. For each indicator type of each test item, an effective indicator tag was generated to indicate whether there was a measurement value that could be used for calculation. The measurement units were then standardized, and finally, the test item set was sorted according to the standardized order of the test items. The system includes a pre-defined diagnostic rule base for 15 types of peripheral nerve injuries. Each type of injury has its own specific test criteria and threshold values. The 15 types of peripheral nerve injuries are: carpal tunnel syndrome, cubital tunnel syndrome, proximal median nerve injury, carpal tunnel syndrome combined with cubital tunnel syndrome, axonal peripheral neuropathy, axonal peripheral neuropathy combined with carpal tunnel syndrome, axonal peripheral neuropathy combined with cubital tunnel syndrome, acute demyelinating Guillain-Barré syndrome, acute motor axonal neuropathy-type Guillain-Barré syndrome, chronic Guillain-Barré syndrome, radial nerve injury, common peroneal nerve injury, tibial nerve ankle injury, and femoral nerve injury. The specific diagnostic rules are as follows: (1) Admission criteria for test items: The minimum set of test items required for the corresponding disease determination is pre-stored. Each test item in the minimum set of test items is determined by the standard nerve name, nerve type, stimulation site, and recording site respectively. The rule matching process for the disease is initiated only when the patient's set of test items covers the minimum set of test items corresponding to the disease. (2) Multi-branch judgment logic: There is at least one type of absolute value judgment branch and relative difference judgment branch. Each type may contain one or more judgment branches. The judgment branches are logically ORed. If all the limiting conditions of any judgment branch are met, it is considered to meet the diagnostic rules of the disease. The judgment logic of the absolute value judgment branch is: the specified measurement index value of the target detection item exceeds the preset normal threshold range, and the corresponding measurement index value of the reference detection item used as a control meets the preset normal threshold range. The judgment logic of the relative difference judgment branch is: the calculated difference of the same measurement index of two paired test items with a control relationship exceeds the preset difference threshold.

3. The method for interpreting and diagnosing nerve conduction test data according to claim 1, characterized in that the construction process of the comparison feature matrix in S2 is as follows: S2.

1. Establish a preset anatomical segment sequence table based on preset neural conduction anatomy rules, record the proximal and distal order of stimulation sites under the same standard nerve name, nerve type, and recording site, and retrieve proximal and distal paired test items for each test item in the test item set according to the standard nerve name, left and right sides, nerve type, stimulation site, and recording site, while recording the distal and proximal test items in the paired relationship; then, retrieve contralateral paired test items for each test item according to the standard nerve name, left and right sides, nerve type, stimulation site, and recording site, and generate valid proximal and distal paired markers and valid contralateral paired markers; S2.2 Based on the identified pairing relationships, calculate the clinically significant contrast features for each test item from the original measurements. Explicitly construct these key diagnostic criteria into contrast feature vectors. Stack the contrast feature vectors of all test items row by row according to the standardized order of the test items to obtain the contrast feature matrix.

4. The method for interpreting and diagnosing nerve conduction test data according to claim 1, characterized in that, The specific process of generating the feature sequence of the detection item in S4 is as follows: By utilizing group quantile information, the original measurement values ​​with different dimensions and ranges are uniformly scaled to the [0,1] interval to generate robust scaling values; The robust scaling vector, contrast feature vector, orientation consistency deviation encoding vector, and effective label vector of each detection item are concatenated along the feature dimension in a fixed order to obtain a multi-granularity numerical feature vector. This multi-granularity numerical feature vector is then input into a linear projection layer and mapped to a unified hidden dimension to obtain a numerical fusion representation. Next, trainable embedding tables are established for standard nerve names, left / right sides, nerve types, stimulation sites, and recording sites, and the corresponding embedding vectors are read from these tables. Then, anatomical sequence position encoding vectors are generated based on a preset anatomical segment order table. The numerical fusion representation, each embedding vector, and the anatomical sequence position encoding vector are added together to obtain the comprehensive feature vector for each detection item. Finally, all comprehensive feature vectors for each detection item are stacked row by row according to the standardized order of the detection items to form a detection item feature sequence.

5. The method for interpreting and diagnosing nerve conduction test data according to claim 1, characterized in that, S5 is detailed below: S5.

1. The feature sequence X of the detection item is used as the input of the relation-aware heterogeneous graph attention network. Each attention head performs linear projection on the node feature vector of the target detection item and the node feature vector of the neighbor detection item to obtain the query representation, key representation and value representation. A trainable edge type embedding is introduced into the graph attention mechanism. The query representation, key representation and edge type embedding vector are input into the attention scoring function to obtain the relation-aware attention score. S5.

2. Softmax exponential normalization is performed on the relation perception scores calculated for all neighbors of each node, and the scores are converted into attention weights with a sum of 1. S5.

3. Based on the calculated attention weights, the features of all neighboring nodes are weighted and summed to obtain the relationship aggregation results under each attention head. The relationship aggregation results of multiple attention heads are concatenated and used as the candidate node update features for the next layer. Then, the standard operation of the Transformer architecture and deep network in deep learning is used to obtain the node feature vector of each detection item in the next layer. S5.4 By stacking multiple layers of relation-aware heterogeneous graph attention layers, each detection node gradually aggregates information from its multi-hop neighbors, outputting graph-coded node features rich in anatomical relationships and contextual information, thus forming a graph-coded node feature matrix.

6. A method for interpreting and diagnosing nerve conduction test data according to claim 1, characterized in that, S6 Specifically as follows: S6.

1. Each detection item generates a pairing difference enhancement feature by subtracting its own graph encoding node features from the features of the clinical pairing item, and then concatenates it with its own graph encoding node features to obtain the pairing difference enhancement feature vector. S6.

2. A trainable pooling scoring function is adopted, with paired differential enhancement features as input, to automatically learn and calculate the importance of each detection item to the final diagnosis, and generate pooling weights after normalization. S6.

3. Based on the contrast-aware pooling weights, the graph coding node features of all detection items are weighted and summed to aggregate a fixed-length patient-level global feature vector. The patient-level global feature vector is input into the fully connected classification head to obtain the disease category score. The disease category score is then input into the normalized exponential function to obtain the disease category prediction probability vector. Finally, the disease category with the highest probability in the disease category prediction probability vector is taken as the preliminary interpretation result, and the detection items with higher comparative perceptual attention pooling weights are output as the interpretation basis prompts.

7. The method for interpreting and diagnosing nerve conduction test data according to claim 1, characterized in that, S7 is detailed below: Each training patient sample is organized into a training test item set, and the true disease category label is determined by clinical diagnosis results or expert review results; the historical normal sample distribution, low quantile reference value and high quantile reference value are statistically analyzed from normal patient samples or normal population samples confirmed to have no evidence of peripheral neuropathy in the training set; The loss function is calculated based on the predicted structure obtained from the training set data, and the trainable parameters are updated through iterative training. The optimal parameters are retained to obtain a trained automatic interpretation model for neural conduction examination data.

8. A system for interpreting and diagnosing nerve conduction test data, performing a method for interpreting and diagnosing nerve conduction test data as described in any one of claims 1-7, characterized in that, Includes the following modules: The data preprocessing module is used to perform field extraction, terminology standardization, measurement validity marking, and test item sorting on the patient's nerve conduction examination data to obtain a set of test items. Each test item includes a standard nerve name, left and right sides, nerve type, stimulation site, recording site, and at least one measurement index. The pairing relationship identification and contrast feature generation module is used to automatically identify proximal and contralateral pairing relationships in the set of test items based on prior rules of neural conduction anatomy, and convert pairwise differences into explicit contrast features to obtain a contrast feature matrix; The directional deviation matrix construction module is used to construct a grouped normal reference distribution using a historical normal population database, convert patient test values ​​into directional deviation codes, concatenate all directional deviation code values ​​of each test item into a vector in a fixed order, and then stack all vectors of test items into a directional deviation matrix in a standard order. The multi-feature fusion and anatomical relationship map construction module is used to robustly scale the original measurement data in the detection item set to obtain robust scaling results. After mapping the results, the contrast feature matrix, the direction consistency deviation matrix and the measurement value validity label to the same hidden dimension, the module is fused to generate the detection item feature sequence and construct the neural conduction anatomical relationship map based on the anatomical relationship between the detection items. The heterogeneous graph encoding module takes the feature sequence of the detection item as input, introduces trainable edge type embedding in the graph attention mechanism, calculates the attention score between nodes through the relation-aware heterogeneous graph attention network, and outputs the graph-encoded node feature matrix. The global feature extraction and disease prediction module is used to obtain the pairing difference enhancement feature vector based on the difference between the graph coding node features of the detection item and the paired detection item, add it to the contrast perception attention pooling weight calculation, and obtain the patient-level global feature vector by weighted summation of the features of all graph coding nodes. Then, the disease category prediction probability vector is calculated based on this, and the disease category corresponding to the highest probability is selected as the output result. The model training and online interpretation module is used to train the model of the entire system using training samples with real disease category labels, resulting in a trained automatic interpretation model for neural conduction examination data, which is then used for online interpretation and diagnosis.

Images