Cross-specialized diagnosis evaluation method and related device

Abstract

The invention discloses a cross-specialist diagnosis evaluation method and a related device, and relates to the technical field of data processing, and the method comprises the steps: obtaining a target case text of a plurality of specialists related to pathogenesis and a corresponding candidate diagnosis set, employing a specialist evaluation model of each specialist, and carrying out the independent evaluation of the candidate diagnosis set according to the target case text, the method comprises the steps of obtaining a specialist assessment result of each specialist, performing cross-specialist collaborative assessment based on all the specialist assessment results to obtain a cross-specialist collaborative assessment result corresponding to each specialist assessment result, and generating a fusion assessment result according to all the cross-specialist collaborative assessment results. The fusion evaluation result is used for representing comprehensive confidence distribution of each candidate diagnosis in the candidate diagnosis set after cross-specialist collaborative evaluation. According to the method, the evaluation process consistent with the real consultation process is adopted, so that the fusion evaluation result can more accurately perform confidence evaluation on each candidate diagnosis in the candidate diagnosis set, and the interpretability of the evaluation process is higher.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a cross-specialty diagnostic assessment method and related apparatus. Background Technology

[0002] With the deep integration of medical informatization and artificial intelligence technologies, AI-assisted diagnostic systems have gradually extended from screening for single diseases to comprehensive analysis of complex cases. Among them, the auxiliary assessment of complex cases with multi-specialty pathological associations (such as chest pain of unknown cause and abdominal pain of unknown cause) has become a key scenario that urgently needs to be broken through.

[0003] Currently, the auxiliary assessment process for complex cases is as follows: identify the multiple specialties associated with the complex case, then call the assessment models corresponding to each specialty in turn, generate possible candidate diagnoses, and finally summarize the candidate diagnoses output by each model to obtain a candidate diagnosis set.

[0004] However, this auxiliary assessment scheme has the following drawbacks: First, the assessment models for each specialty operate independently and are fragmented, making it impossible to reflect the interaction and collaboration between multiple specialties in real clinical practice. Second, the final output is only a few discrete candidate diagnoses, lacking a quantitative confidence assessment of the probability of each diagnosis. In other words, the above-mentioned auxiliary assessment scheme is essentially a simple parallel arrangement of multiple specialty "information silos," failing to achieve the integration of multi-specialty assessment information. This results in weak interpretability of the final output, failing to provide doctors with hierarchical and quantitative decision support, and limiting its practical application effectiveness in complex diagnostic scenarios. Summary of the Invention

[0005] In view of the above problems, this application provides a cross-specialty diagnostic assessment method and related apparatus to solve the problem that the assessment process in the prior art lacks multi-specialty collaboration and interaction and multi-specialty assessment information fusion for confidence assessment, resulting in weak interpretability of the output results. The specific solution is as follows:

[0006] The first aspect of this application provides a cross-specialty diagnostic assessment method, including:

[0007] Obtain target case texts involving multiple specialties related to pathogenesis, along with corresponding candidate diagnostic sets;

[0008] Using the specialty assessment model of each specialty, the candidate diagnosis set is independently evaluated based on the target case text to obtain the specialty assessment result of each specialty. The specialty assessment result is used to characterize the confidence distribution of each candidate diagnosis in the candidate diagnosis set under the medical logic of the corresponding specialty.

[0009] Based on all the specialty assessment results, cross-specialty collaborative assessment is conducted to obtain the cross-specialty collaborative assessment result corresponding to each of the aforementioned specialty assessment results;

[0010] A fusion assessment result is generated based on all the interdisciplinary collaborative assessment results. The fusion assessment result is used to characterize the overall confidence distribution of each candidate diagnosis in the candidate diagnosis set after interdisciplinary collaborative assessment.

[0011] In one possible implementation, both the specialty assessment results and the cross-specialty collaborative assessment results are in vector form. The step of performing cross-specialty collaborative assessment based on all specialty assessment results to obtain the cross-specialty collaborative assessment result corresponding to each specialty assessment result includes:

[0012] Linear mapping and encoding are performed on all the specialty assessment results to obtain high-dimensional assessment vectors corresponding to each specialty assessment result. The dimension of the high-dimensional assessment vectors is greater than the dimension of the specialty assessment results.

[0013] Generate the query matrix and key matrix based on all the high-dimensional evaluation vectors;

[0014] The attention score matrix is ​​calculated based on the query matrix and the key matrix, and the attention score matrix is ​​normalized to obtain the attention weight matrix.

[0015] Each row vector of the attention weight matrix is ​​used as an attention weight vector to obtain the attention weight vector of each of the multiple specialties. The attention weight vector is used to quantify the dependence weight of the corresponding specialty on each high-dimensional evaluation result in the cross-specialty collaborative evaluation.

[0016] For each specialty, based on the specialty's attention weight vector and all the high-dimensional evaluation vectors, a cross-specialty collaborative evaluation result corresponding to the specialty's evaluation result is generated.

[0017] In one possible implementation, generating the cross-specialty collaborative assessment result corresponding to the specialty assessment result based on the specialty's attention weight vector and all the high-dimensional assessment vectors includes:

[0018] The attention weight vector of the specialty is weighted and summed with all the high-dimensional evaluation vectors to obtain the collaborative state vector of the specialty. The collaborative state vector represents the evaluation deviation of the independent evaluation of the corresponding specialty compared with the cross-specialty collaborative evaluation.

[0019] By using the collaborative state vector of the specialty to compensate for the bias in the high-dimensional evaluation vector corresponding to the specialty evaluation result, the compensated state vector of the specialty is obtained.

[0020] The compensated state vector of the specialty is subjected to layer normalization to obtain the normalized state vector of the specialty.

[0021] The normalized state vector of the specialty is linearly mapped to a space of target dimension to obtain the cross-specialty collaborative assessment result corresponding to the specialty assessment result. The target dimension is the total number of candidate diagnoses included in the candidate diagnosis set.

[0022] In one possible implementation, generating a fusion assessment result based on all the cross-specialty collaborative assessment results includes:

[0023] The mean of all the cross-specialty collaborative assessment results across the candidate diagnostic dimension is calculated to obtain the global collaborative result;

[0024] For each candidate diagnosis in the candidate diagnosis set:

[0025] The confidence score of the candidate diagnosis is obtained from the cross-specialty collaborative assessment results corresponding to each of the aforementioned specialty assessment results, and is used as the cross-specialty confidence score corresponding to the specialty of each of the aforementioned specialty assessment results;

[0026] Obtain the importance weights of each of the multiple specialties, and then sum the obtained importance weights with the cross-specialty confidence scores of each of the multiple specialties to obtain the weighted score of the candidate diagnosis.

[0027] To obtain a weighted score result composed of the weighted scores of each candidate diagnosis in the candidate diagnosis set;

[0028] The global collaborative results and the weighted score results are weighted and fused in the candidate diagnostic dimension using a weight balancing factor to obtain the fusion evaluation result.

[0029] In one possible implementation, obtaining the importance weights of each of the plurality of specialties includes:

[0030] The target case text is semantically encoded using a configured semantic encoding model to obtain a global semantic feature vector, which represents the semantic information related to the diagnostic content in the target case text.

[0031] The global semantic feature vector is transformed and mapped using a configured multilayer perceptron to obtain a normalized weight vector with a dimension equal to the total number of specialties, wherein the total number of specialties is greater than or equal to the number of the multiple specialties.

[0032] The importance weights of each of the multiple specialties are obtained from the normalized weight vector.

[0033] In one possible implementation, the step of using a configured multilayer perceptron to perform feature transformation mapping on the global semantic feature vector to obtain a normalized weight vector with a dimension equal to the total number of specialties includes:

[0034] Semantic information related to specialties is extracted from the global semantic feature vector using a first fully connected layer and an activation function to obtain an intermediate semantic feature vector. The dimension of the intermediate semantic feature vector is less than the dimension of the global semantic feature vector and greater than the total number of specialties.

[0035] The intermediate semantic feature vector is mapped to an initial weight vector with a dimension equal to the total number of specialties using a second fully connected layer;

[0036] The initial weight vector is normalized using a preset normalization function to obtain the normalized weight vector.

[0037] In one possible implementation, the cross-specialty diagnostic assessment method is applied to a cross-specialty diagnostic assessment model, the training process of which includes:

[0038] Obtain user feedback information, which includes at least one of the following: sorting adjustment information for each candidate diagnostic sample in the candidate diagnostic sample set, candidate diagnostic samples determined by the user as the final diagnostic result of the case text sample, and adjustment information for the fusion evaluation result predicted by the cross-specialty diagnostic evaluation model, wherein the sorting adjustment information corresponds to the sorting result before adjustment, which is determined by the predicted fusion evaluation result.

[0039] Generate positive sample preference pairs and corresponding negative sample preference pairs. The priority relationship between the two candidate diagnostic samples in the positive sample preference pair is the same as the priority relationship determined based on the user feedback information. The priority relationship between the two candidate diagnostic samples in the negative sample preference pair is the opposite of the priority relationship determined based on the user feedback information.

[0040] A contrastive learning loss is generated based on the positive sample preference pair and the negative sample preference pair;

[0041] The cross-specialty diagnostic assessment model is trained using the contrastive learning loss.

[0042] In one possible implementation, training the cross-specialty diagnostic assessment model based on the contrastive learning loss includes:

[0043] Obtain a training dataset, wherein each training data in the training dataset includes: the case text sample, the candidate diagnosis sample set, and the corresponding fusion evaluation result label;

[0044] Determine the Fisher information matrix of the cross-specialty diagnostic assessment model on the training dataset;

[0045] Generate a resilient weighted consolidation loss based on the Fisher information matrix;

[0046] The total loss is generated based on the elastic weight consolidation loss and the contrastive learning loss;

[0047] The cross-specialty diagnostic assessment model is trained using the total loss.

[0048] One possible implementation also includes:

[0049] Based on the fusion evaluation results, the candidate diagnoses in the candidate diagnosis set are sorted, and the sorting results are output and displayed.

[0050] And / or, output and display the attention weight vectors and importance weights of each of the multiple specialties in the form of a heatmap;

[0051] And / or, display the weight balancing factor output.

[0052] A second aspect of this application provides a computer program product including computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the cross-specialty diagnostic assessment method of the first aspect or any implementation thereof.

[0053] A third aspect of this application provides an electronic device, comprising at least one processor and a memory connected to the processor, wherein:

[0054] The memory is used to store computer programs;

[0055] The processor is used to execute the computer program so that the electronic device can implement the cross-specialty diagnostic assessment method of the first aspect or any implementation thereof.

[0056] A fourth aspect of this application provides a computer storage medium carrying one or more computer programs, which, when executed by an electronic device, enable the electronic device to implement the cross-specialty diagnostic assessment method described in the first aspect or any implementation thereof.

[0057] Using the above technical solution, the cross-specialty diagnostic assessment method provided in this application obtains target case texts involving multiple specialties in pathogenesis, as well as corresponding candidate diagnosis sets. In order to assess the confidence level of the diagnostic rationality of each candidate diagnosis in the candidate diagnosis set, this embodiment can use the specialty assessment model of each specialty to independently assess the candidate diagnosis set based on the target case text, and obtain the specialty assessment result of each specialty. The specialty assessment result is used to characterize the confidence distribution of each candidate diagnosis in the candidate diagnosis set under the medical logic of the corresponding specialty.

[0058] Since the pathogenesis of a case involves multiple specialties, multi-specialty collaborative consultations are conducted in real clinical practice. Therefore, the single-specialty independent assessment method mentioned above deviates from the multi-specialty collaborative assessment method in real consultations. To compensate for this deviation, this application can conduct cross-specialty collaborative assessments based on all specialty assessment results to obtain the cross-specialty collaborative assessment results corresponding to each specialty assessment result. This makes the cross-specialty collaborative assessment results of each specialty closer to the real assessment results obtained in real consultations after referring to the diagnostic assessments of other specialties.

[0059] Finally, a fusion assessment result was generated based on all the cross-specialty collaborative assessment results, realizing the integration of multi-specialty assessment information. Since the entire assessment process follows the workflow of "preliminary specialist assessment, cross-discussion, and comprehensive decision-making," which is consistent with the actual consultation process, the fusion assessment result can more accurately assess the confidence level of each candidate diagnosis in the candidate diagnosis set, and the assessment process is more interpretable. Attached Figure Description

[0060] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0061] Figure 1 A schematic diagram of a system architecture provided for this application;

[0062] Figure 2 A flowchart illustrating a cross-specialty diagnostic assessment method provided in this application;

[0063] Figure 3 A schematic diagram illustrating the process of structuring original medical record text provided in this application;

[0064] Figure 4 A schematic diagram of a cross-specialty diagnostic assessment device provided in this application;

[0065] Figure 5This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation

[0066] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0067] The embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.

[0068] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0069] It is understood that before using the technical solutions disclosed in the various embodiments of this application, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this application in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0070] It is understood that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and related provisions.

[0071] This application provides a cross-specialty diagnostic assessment method and related device, which can be applied to scenarios that assist doctors in disease diagnosis.

[0072] Optionally, the cross-specialty diagnostic assessment method provided in this application can be applied to, for example... Figure 1 The system architecture shown includes a terminal 100 and a server 200. The server 200 may include one or more servers (…). Figure 1 (This example uses a server as an illustration).

[0073] Either terminal 100 or server 200 can be used independently to execute the cross-specialty diagnostic assessment method provided in the embodiments of this application. Alternatively, terminal 100 and server 200 can also be used collaboratively to execute the cross-specialty diagnostic assessment method provided in the embodiments of this application.

[0074] For example, after terminal 100 obtains the target case text involving multiple specialties in the pathogenesis, as well as the corresponding candidate diagnosis set, it sends the target case text and the candidate diagnosis set to server 200. Server 200 performs cross-specialty diagnostic evaluation on each candidate diagnosis in the candidate diagnosis set based on the target case text, and obtains the fusion evaluation result that represents the comprehensive confidence distribution of each candidate diagnosis. The result is then returned to terminal 100 for output, so that doctors can make a comprehensive diagnostic decision based on the fusion evaluation result, combined with their own clinical experience and other examination results.

[0075] The following description Figure 1 The product form of the mid-terminal 100;

[0076] The terminal 100 in this application embodiment can be a mobile phone, tablet computer, wearable device, vehicle-mounted device, augmented reality (AR) / virtual reality (VR) device, laptop computer, ultra-mobile personal computer (UMPC), netbook, personal digital assistant (PDA), etc., and this application embodiment does not impose any restrictions on it.

[0077] To enable those skilled in the art to better understand this application, the cross-specialty diagnostic assessment method of this application will be described in detail below with reference to the accompanying drawings.

[0078] Reference Figure 2 , Figure 2 A flowchart illustrating a cross-specialty diagnostic assessment method provided in this application embodiment is shown below. Figure 2 As shown, this interdisciplinary diagnostic assessment method may include:

[0079] Step S101: Obtain target case texts involving multiple specialties related to the pathogenesis, as well as the corresponding set of candidate diagnoses.

[0080] In this embodiment, cases involving multiple specialties in pathogenesis are defined as target cases. The original case text of a target case includes, but is not limited to, the following information: basic patient information, such as age, gender, and basic medical history; chief complaint information, such as recurrent chest pain for 3 days; symptoms and signs, such as squeezing pain behind the sternum accompanied by sweating; and examination results, such as troponin 0.05 ng / mL and low T wave levels in leads V1-V3 of the electrocardiogram.

[0081] Understandably, the original case texts are usually semi-structured or unstructured clinical data. To facilitate subsequent processing, the original case texts can be structured to obtain structured case texts, which can then be used as target case texts.

[0082] Figure 3 This is a schematic diagram illustrating a process for structuring original medical record text, as provided in this application. Figure 3 First, key medical entities such as disease name, symptom terms, examination indicators and values ​​can be extracted from the original medical record text to ensure that no information related to the diagnosis is omitted from the original medical record text. Here, the information related to the diagnosis includes, but is not limited to, the following: the diagnosis itself and information used to identify suspected diagnoses.

[0083] Optionally, a BERT-CRF (Bidirectional Encoder Representations from Transformers - Conditional Random Fields, a sequence labeling model based on BERT and CRF) pre-trained model with fine-tuned medical corpus can be used to extract key medical entities from the original case text.

[0084] Furthermore, key medical entities extracted can be preprocessed with semantic cleaning and standardization. For example, redundant information irrelevant to diagnosis, such as patient identification and medical record numbers, can be removed. At the same time, the units of examination indicators can be standardized (e.g., "troponin 0.05ng / mL" and "troponin 50pg / mL" can be unified into "ng / mL"), and terminology can be standardized to avoid misjudgments due to differences in units or terminology.

[0085] Next, considering that key medical entities in the medical field may be very long, such as "acute exacerbation of chronic obstructive pulmonary disease complicated with type II respiratory failure", in order to ensure the accuracy of subsequent processing, the pre-processed key medical entities can be segmented.

[0086] To avoid errors caused by the mechanical segmentation of key medical entities by general word segmentation tools, it is preferable to use specialized word segmentation tools for the medical field to accurately segment key medical entities. For example, "chronic obstructive pulmonary disease acute exacerbation complicated with type II respiratory failure" can be segmented into "chronic obstructive pulmonary disease", "acute exacerbation", "combined with", and "type II respiratory failure".

[0087] Finally, the word segmentation and extracted values ​​can be structurally normalized to obtain a structured case feature table of "entity-value" association, which serves as the target case text in this embodiment. For example, {"symptoms: chest pain, onset time: 3 days", "indicator: troponin, value: 0.05ng / mL"}.

[0088] In this embodiment, multiple candidate diagnoses corresponding to the target case text can also be obtained. Optionally, the multiple candidate diagnoses can come from clinicians or from other auxiliary diagnostic systems. For example, candidate diagnoses for chest pain cases include "angina pectoris", "pulmonary embolism", and "gastroesophageal reflux".

[0089] Considering that due to individual differences, different doctors or auxiliary diagnostic systems may use different terminology for the same diagnostic result, in order to avoid errors in subsequent evaluation caused by differences in terminology, multiple candidate diagnoses can be standardized as a string array. That is, the naming conventions and format of candidate diagnoses are unified to eliminate ambiguity and redundancy and form a string array with a consistent structure.

[0090] For example, multiple candidate diagnoses for a case of "chest pain of unknown origin" may include: myocardial infarction, pulmonary embolism, gastroesophageal reflux disease (GERD), stable angina, pulmonary embolism, stable angina, and GERD. After standardization, the resulting set of candidate diagnoses may include: myocardial infarction, pulmonary embolism, gastroesophageal reflux, and angina.

[0091] It is understandable that the pathogenesis of different cases involves different specialties. In order to identify the multiple specialties involved in the pathogenesis of the target case, this embodiment can determine the multiple specialties most relevant to the target case from all specialties based on the importance weight of each specialty supported by the target case text evaluation system.

[0092] Optionally, the process of "evaluating the importance weights of all specialties supported by the system based on the target case text" may include: semantically encoding the target case text using a configured semantic encoding model to obtain a global semantic feature vector; performing feature transformation mapping on the global semantic feature vector using a configured multi-layer perceptron (MLP) to obtain a normalized weight vector with a dimension equal to the total number of specialties, where the total number of specialties is greater than or equal to the number of multiple specialties; and obtaining the importance weights of each specialty from the normalized weight vector.

[0093] Optionally, the semantic encoding model can be a BioBERT-base (Bidirectional Encoder Representations from Transformers for Biomedical Text Mining - Base Version) pre-trained model. This model has been pre-trained on 1.5 million medical documents from a public medical center (PubMed Central, PMC), naturally possessing the advantage of understanding the semantics of medical text. In specific processing, this model can encode the target case text and finally output a 768-dimensional global semantic feature vector from the CLS position of the text encoding result. This vector fully covers the semantic information related to the diagnosis in the target case text, such as symptoms, examination indicators, and the correlation between different specialties.

[0094] Optionally, the multilayer perceptron can be a network architecture with two fully connected layers, an activation function, and a normalization function. In this embodiment, the first fully connected layer and the activation function can be used to extract semantic information related to specialties from the global semantic feature vector to obtain an intermediate semantic feature vector. Then, the second fully connected layer is used to map the intermediate semantic feature vector to an initial weight vector with a dimension equal to the total number of specialties. Finally, the initial weight vector is normalized using a preset normalization function to obtain a normalized weight vector. The dimension of the intermediate semantic feature vector is less than the dimension of the global semantic feature vector and greater than the total number of specialties.

[0095] More specifically, a rectified linear unit (ReLU) can be used as the activation function between the first fully connected layer and the second fully connected layer to avoid the gradient vanishing problem. After processing by the first fully connected layer and the activation function, the global semantic feature vector can be reduced to a lower-dimensional space. During the dimensionality reduction process, the key information on the semantic relationship between the case and the specialty is extracted to more accurately measure the importance of each specialty.

[0096] For example, after processing by the first fully connected layer and ReLU, the 768-dimensional global semantic feature vector can be compressed into a 256-dimensional intermediate semantic feature vector.

[0097] The second fully connected layer further maps the intermediate semantic feature vectors into k-dimensional initial weight vectors, where k represents the total number of specialties supported by the system, and k ≥ m (where m represents the number of multiple specialties in this embodiment, i.e., the number of core specialties more relevant to the target case). For example, K=5 corresponds to cardiology, respiratory medicine, gastroenterology, emergency medicine, and neurology. Of course, K=5 is merely an example and is not intended to limit this application.

[0098] Furthermore, to ensure a reasonable weight allocation for each specialty, a normalization function such as the softmax function can be used to normalize the k-dimensional initial weight vector, so that the sum of the importance weights of all specialties is 1. The specific formula is as follows:

[0099] Formula (1);

[0100] in, This represents the importance weight of the i-th specialty among all specialties. This represents the original score of the i-th specialty (i.e., the i-th dimension of the initial weight vector).

[0101] Taking the target medical record of "chest pain for 3 days + acid reflux + troponin 0.05 ng / mL" as an example, the initial weight vector after the above MLP processing is [cardiology 0.5, respiratory 0.25, gastroenterology 0.15, emergency department 0.05, neurology 0.05]. In order to select multiple specialties from them, this embodiment can preset the weight threshold, such as 0.1, so that multiple specialties are cardiology, respiratory and gastroenterology, and the importance weights of the three specialties can be obtained, which are 0.5, 0.25 and 0.15 respectively.

[0102] After the above processing, multiple specialties (such as the three specialties mentioned above) can be identified as the core of the target case. Subsequently, the evaluation of each candidate diagnosis in the candidate diagnosis set can be carried out around these multiple specialties, thereby focusing more on the key specialty dimensions, reducing the interference of irrelevant specialties on subsequent cross-specialty collaborative evaluation, improving the overall evaluation efficiency, and improving the interpretability of the final evaluation.

[0103] Step S102: Using the specialty assessment model for each specialty, independently assess the candidate diagnosis set based on the target case text to obtain the specialty assessment results for each specialty.

[0104] Considering that the actual consultation process is usually "preliminary specialist assessment, cross-disciplinary discussion, and comprehensive decision-making", in order to reproduce this consultation process, this embodiment pre-trains a specialist assessment model for each medical specialty to realize the confidence assessment process of "preliminary specialist assessment".

[0105] Optionally, the specialty assessment model can adopt a Transformer-based architecture, which can simultaneously process textual features (such as case descriptions, examination reports, etc.) and structured features (such as indicator values, symptom duration, etc.), making it more suitable for the diagnostic data needs of each specialty in this application.

[0106] The training data used for training the aforementioned specialty assessment model may include: case text samples, a set of candidate diagnostic samples, and corresponding specialty assessment result labels. The specialty assessment result labels may be diagnostic scores representing the confidence scores of each candidate diagnostic sample from the perspective of the corresponding specialty, obtained by doctors of the corresponding medical specialty or by a large model based on the medical logic (such as diagnostic logic) of the corresponding specialty, independently evaluating the rationality of the candidate diagnostic sample set according to the case text samples.

[0107] More specifically, thousands of specialist case texts (including but not limited to outpatient and inpatient cases) can be used as training data. The labels for specialist assessment results can be annotated by humans and large models. When there are disagreements between human annotation and large model annotation, a third party (such as another large model with higher accuracy) can be introduced for arbitration to ensure consistency of annotation.

[0108] Optionally, this embodiment may also use methods such as text synonym replacement and slight perturbation of index values ​​to expand the above training data in order to avoid overfitting of the specialty evaluation model.

[0109] After obtaining sufficient training data, case text samples and candidate diagnosis sample sets can be input into the specialty assessment model to obtain the specialty assessment results predicted by the model. Then, the mean squared error loss is calculated between the specialty assessment results predicted by the model and the specialty assessment result labels, and the model is trained to optimize the scoring accuracy of the specialty assessment model.

[0110] After training the specialty assessment model, this embodiment can input the target case text and the candidate diagnosis set into the specialty assessment model of each specialty, so as to automatically focus on the key information of the specialty assessment through the text feature attention mechanism in the model and obtain the specialty assessment result of each specialty.

[0111] Here, the specialist assessment results are used to characterize the confidence distribution of each candidate diagnosis in the candidate diagnosis set under the medical logic of the corresponding specialty. That is, it characterizes the degree of fit and rationality between the candidate diagnosis and the case characteristics from the perspective of the corresponding specialty (i.e., the rationality of the candidate diagnosis as the candidate diagnosis corresponding to the target case text), rather than the probability distribution of the candidate diagnosis as the real disease.

[0112] In one possible implementation, the specialist assessment results can be in vector form, with dimension D, where D represents the number of candidate diagnoses in the candidate diagnosis set. Each dimension of the D-dimensional vector represents the confidence score of a candidate diagnosis, ranging from 0 to 1. The higher the confidence score, the better the fit between the candidate diagnosis and the target case text.

[0113] Taking a case of "chest pain of unknown origin" (assuming the candidate diagnosis is [angina pectoris, pulmonary embolism, gastroesophageal reflux]) as an example, each specialty assessment model completes an independent assessment based on a unified architecture: the cardiology assessment model outputs specialty assessment results [0.8 (angina pectoris), 0.3 (pulmonary embolism), 0.1 (gastroesophageal reflux)]; the pulmonology model outputs specialty assessment results [0.4 (angina pectoris), 0.7 (pulmonary embolism), 0.2 (gastroesophageal reflux)]; and the gastroenterology model outputs specialty assessment results [0.2 (angina pectoris), 0.1 (pulmonary embolism), 0.8 (gastroesophageal reflux)].

[0114] Step S103: Based on all the specialty assessment results, conduct cross-specialty collaborative assessment to obtain the cross-specialty collaborative assessment result corresponding to each specialty assessment result.

[0115] All the above-mentioned specialty assessment results refer to the individual specialty assessment results of multiple specialties, that is, the specialty assessment results output by the individual specialty assessment models of multiple specialties; cross-specialty collaborative assessment results refer to the final assessment results of each candidate diagnosis by the corresponding specialty after referring to the diagnostic assessments of other specialties.

[0116] As mentioned earlier, the second stage of the actual consultation process is "cross-discussion". Since the pathogenesis of the target case involves multiple specialties, the actual evaluation result of each specialty is the evaluation result after cross-discussion with other specialties. Based on this, step S102 adopts the medical logic of a single specialty for independent evaluation. The obtained specialty evaluation results usually deviate from the actual evaluation results under the specialty consultation.

[0117] In order to obtain cross-specialty collaborative assessment results that are as close as possible to the actual assessment results of the specialties, this embodiment can conduct cross-specialty collaborative assessment based on the specialty assessment results of each specialty obtained above. This allows each specialty to refer to the diagnostic assessments of other specialties to correct its own specialty assessment results and obtain the corrected assessment results, which are the cross-specialty collaborative assessment results.

[0118] Similar to the results of the previous specialist assessment, the results of the cross-specialty collaborative assessment can also be a D-dimensional vector. Each dimension represents the confidence score of a candidate diagnosis after being updated by the multi-specialty collaborative assessment. The value ranges from 0 to 1. The higher the confidence score, the higher the fit between the candidate diagnosis and the target case text, and the higher the rationality of the candidate diagnosis.

[0119] Step S104: Generate integrated assessment results based on all interdisciplinary collaborative assessment results.

[0120] Among them, the fusion assessment results are used to characterize the overall confidence distribution of each candidate diagnosis in the candidate diagnosis set after cross-specialty collaborative assessment; all cross-specialty collaborative assessment results refer to the cross-specialty collaborative assessment results corresponding to each of the specialty assessment results.

[0121] Through the process described above, we can obtain the rationality of each specialist's diagnosis for each candidate diagnosis. Therefore, we can integrate the cross-specialty collaborative assessment results corresponding to the assessment results of multiple specialties to obtain the comprehensive confidence assessment results of multiple specialties for each candidate diagnosis, thus realizing the "comprehensive decision" for each candidate diagnosis in the candidate diagnosis set.

[0122] The cross-specialty diagnostic assessment method provided in this application obtains target case texts involving multiple specialties in pathogenesis, as well as a corresponding set of candidate diagnoses. In order to assess the confidence level of the diagnostic rationality of each candidate diagnosis in the candidate diagnosis set, this embodiment can use the specialty assessment model of each specialty to independently assess the candidate diagnosis set based on the target case text, and obtain the specialty assessment result of each specialty. The specialty assessment result is used to characterize the confidence distribution of each candidate diagnosis in the candidate diagnosis set under the medical logic of the corresponding specialty.

[0123] Since the pathogenesis of a case involves multiple specialties, multi-specialty collaborative consultations are conducted in real clinical practice. Therefore, the single-specialty independent assessment method mentioned above deviates from the multi-specialty collaborative assessment method in real consultations. To compensate for this deviation, this application can conduct cross-specialty collaborative assessments based on all specialty assessment results to obtain the cross-specialty collaborative assessment results corresponding to each specialty assessment result. This makes the cross-specialty collaborative assessment results of each specialty closer to the real assessment results obtained in real consultations after referring to the diagnostic assessments of other specialties.

[0124] Finally, a fusion assessment result was generated based on all the cross-specialty collaborative assessment results, realizing the integration of multi-specialty assessment information. Since the entire assessment process follows the workflow of "preliminary specialist assessment, cross-discussion, and comprehensive decision-making," which is consistent with the actual consultation process, the fusion assessment result can more accurately assess the confidence level of each candidate diagnosis in the candidate diagnosis set, and the assessment process is more interpretable.

[0125] In some embodiments of this application, the process of step S103, "to perform cross-specialty collaborative evaluation based on all the specialty evaluation results and to obtain the cross-specialty collaborative evaluation result corresponding to each specialty evaluation result," is described with the premise that both the specialty evaluation results and the cross-specialty collaborative evaluation results are in vector form.

[0126] It should be understood that the number of candidate diagnoses in a typical candidate diagnosis set is not large, resulting in a low dimensionality of the specialist assessment results. Considering that low-dimensional vectors are insufficient to fully represent the assessment preferences of the specialist assessment model, leading to insufficient information, in order to more accurately reflect the assessment preferences of the specialist assessment model, optionally, this embodiment can perform linear mapping and encoding on all specialist assessment results to obtain high-dimensional assessment vectors corresponding to each specialist assessment result. Here, the dimension of the high-dimensional assessment vector is greater than the dimension of the specialist assessment result, that is, the dimension of the high-dimensional assessment vector is greater than D.

[0127] Optionally, the linear mapping process can be implemented using a pre-configured linear layer, for example, mapping each specialty assessment result to a higher-dimensional h-dimensional hidden space (e.g., h=512). Since the higher-dimensional space is richer in information, mapping the specialty assessment results to a higher-dimensional space can enhance its representation of the specialty assessment model's assessment preferences, that is, enhance its representation of the corresponding specialty assessment preferences.

[0128] Optionally, the encoding process can be implemented using one or more fully connected neural networks. That is, one or more cascaded fully connected neural networks encode the vector mapped to the high-dimensional space to obtain the encoded feature vector, and then the sigmoid function is used to perform nonlinear activation on the encoded feature vector to obtain the high-dimensional evaluation vector.

[0129] For example, if the specialist assessment results for cardiology (identity identifier 1) are [0.8 (angina pectoris), 0.3 (pulmonary embolism), 0.1 (gastroesophageal reflux)], then the high-dimensional assessment vector is: This high-dimensional evaluation vector may place more emphasis on the preference weight for "angina".

[0130] To determine how much information from other specialties each specialty needs to reference during cross-specialty collaborative assessment, this application provides a method for cross-specialty collaborative assessment using a cross-attention mechanism. Specifically, the assessment results of each specialty are used as the query (representing "based on my initial judgment, what information should I focus on from other departments"), and the assessment results of all specialties are used as keys and values ​​(representing a collective knowledge base of multiple specialties). Cross-attention calculation is then performed to obtain the attention weight vector corresponding to each specialty assessment result.

[0131] To quickly generate the attention weight vector corresponding to each specialty assessment result, this embodiment can further generate a query matrix and a key matrix based on all high-dimensional assessment vectors. Taking multiple specialties including cardiology, respiratory medicine, and gastroenterology as an example, with identity identifiers 1, 2, and 3 respectively, the query matrix can be: The key matrix and the query matrix can be the same. As can be seen, each row of the query matrix corresponds to a high-dimensional evaluation vector for a specific specialty, where... This represents the high-dimensional assessment vector corresponding to the cardiology specialty assessment results. This represents the high-dimensional assessment vector corresponding to the specialty assessment results in the respiratory department. This represents the high-dimensional evaluation vector corresponding to the specialty evaluation results of the gastroenterology department.

[0132] After obtaining the query matrix and key matrix, the attention score matrix can be calculated based on the query matrix and key matrix. For example, the attention score matrix can be calculated using the following formula (2).

[0133] Formula (2);

[0134] in, Let T represent the attention score matrix, and let T denote the transpose. This represents the preset scaling factor, used to prevent the softmax gradient from disappearing during subsequent normalization due to an excessively large dimension h.

[0135] In the above formula, , m represents the number of specialties, where the j-th element in the i-th row is... This represents the initial level of interest that the i-th specialty has in the j-th specialty;

[0136] It should be understood The numerical range of the attention score matrix is ​​unstable and may affect subsequent processing. Therefore, it is necessary to normalize the attention score matrix to obtain the attention weight matrix.

[0137] Optionally, the attention score matrix can be normalized using the softmax function, i.e.:

[0138] Formula (3);

[0139] in, This represents the attention weight matrix.

[0140] It is worth noting that, similar to the query matrix mentioned earlier, the attention weight matrix is ​​also an m×m matrix, i.e. Since each row of the query matrix corresponds to a high-dimensional evaluation vector for a specialty, each row vector of the attention weight matrix is ​​an attention weight vector, i.e. The j-th element in the i-th row This represents the proportion of attention (i.e., dependence weight) that the i-th specialty pays to the j-th specialty during the cross-specialty collaborative assessment process. Based on this, the attention weight vectors of each specialty can be obtained from the attention weight matrix. Here, the attention weight vectors are used to quantify the dependence weight of the corresponding specialty on each high-dimensional assessment result in the cross-specialty collaborative assessment.

[0141] Next, for each specialty, cross-specialty collaborative evaluation results can be generated based on the specialty's attention weight vector and all high-dimensional evaluation vectors.

[0142] Specifically, as mentioned earlier, each specialty has a dependency weight on each high-dimensional evaluation vector. Therefore, each high-dimensional evaluation vector can be used as a value vector to perform a weighted sum of the specialty's attention weight vector and all high-dimensional evaluation vectors, thus obtaining the specialty's collaborative state vector. Here, the collaborative state vector represents the evaluation deviation between independent evaluation of the corresponding specialty and cross-specialty collaborative evaluation.

[0143] Here, the process of weighted summation can be referred to as formula (4).

[0144] Formula (4);

[0145] in, This represents the collaborative state vector of the i-th specialty. Let represent the high-dimensional evaluation vector corresponding to the evaluation result of the j-th specialty.

[0146] Taking a case of "chest pain of unknown origin" as an example, assume that the attention weight of the cardiology department (i=1) to the respiratory department (j=2) is as follows: The weight for the gastroenterology department (j=3) is: ,but In other words, in the cardiology department's collaborative state vector, 60% of the information comes from the respiratory department, 30% comes from itself, and 10% comes from the gastroenterology department. Ultimately, the updated cross-specialty collaborative assessment results of the cardiology department will include assessment information from the respiratory and gastroenterology departments, achieving a dynamic interactive effect of "cardiology referencing assessments from the respiratory and gastroenterology departments."

[0147] Since the collaborative state vector can characterize the evaluation deviation of the corresponding specialty independent evaluation compared with cross-specialty collaborative evaluation, this application can use the collaborative state vector of the specialty to compensate for the deviation of the high-dimensional evaluation vector corresponding to the specialty evaluation result, and obtain the compensated state vector of the specialty.

[0148] Optionally, the process of “using the collaborative state vector of the specialty to compensate for the deviation of the high-dimensional evaluation vector corresponding to the specialty evaluation result of the specialty, and obtaining the compensated state vector of the specialty” can be as follows: add the collaborative state vector of the specialty to the high-dimensional evaluation vector corresponding to the specialty evaluation result of the specialty, and use the sum as the compensated state vector of the specialty, as shown in the following formula (5).

[0149] Formula (5);

[0150] in, This represents the compensated state vector of the i-th specialty. This represents the collaborative state vector of the i-th specialty. Let represent the high-dimensional evaluation vector corresponding to the evaluation result of the i-th specialty.

[0151] In this embodiment, a residual connection method is used to add the collaborative state vector of the specialty to the high-dimensional evaluation vector corresponding to the specialty evaluation result. This can avoid losing the original evaluation logic of the specialty during the collaboration process. For example, the cardiology department will not ignore its own attention to angina pectoris because it refers to the information of the respiratory department, thus improving the accuracy of subsequent evaluations.

[0152] Next, in this embodiment, the compensated state vector of the specialty can be subjected to layer normalization to obtain the normalized state vector of the specialty, as shown in the following formula (6).

[0153] Formula (6);

[0154] in, This represents the normalized state vector of the i-th specialty. This represents the mean of the compensated state vectors of multiple specialties. This represents the variance of the compensated state vectors for each of the multiple specialties. The default value (1e-5) is used when this application is applied to a cross-specialty diagnostic assessment model. It can prevent gradient explosion during model training and ensure stable model convergence.

[0155] As mentioned earlier, the high-dimensional evaluation vector is an h-dimensional high-dimensional vector. Therefore, the normalized state vector is also an h-dimensional high-dimensional vector. In order to obtain the D-dimensional cross-specialty collaborative evaluation result, this embodiment can linearly map the normalized state vector of the specialty to the space of the target dimension to obtain the cross-specialty collaborative evaluation result corresponding to the specialty evaluation result. The target dimension is the total number of candidate diagnoses contained in the candidate diagnosis set, i.e., D.

[0156] Optionally, another pre-configured linear layer can be used to linearly map the normalized state vector of the specialty to a space of the target dimension, thereby obtaining the cross-specialty collaborative assessment result corresponding to the specialty assessment result. .

[0157] In this embodiment, the process of bias compensation for the results of specialty assessments is carried out in a high-dimensional space, which can obtain high-precision compensation in a richer form of information and improve the assessment accuracy of cross-specialty collaborative assessment results.

[0158] In another embodiment of this application, the process of step S104, "generating a fusion assessment result based on all interdisciplinary collaborative assessment results," is described.

[0159] In this embodiment, each cross-specialty collaborative assessment result includes the cross-specialty confidence score of each candidate diagnosis in the candidate diagnosis set. Therefore, in one possible implementation, the average of all cross-specialty collaborative assessment results along the candidate diagnosis dimension can be directly calculated to obtain the global collaborative result C, which is then used as the fusion assessment result. Here, the global collaborative result reflects the comprehensive assessment after collaboration across all specialties.

[0160] For example, taking a candidate diagnostic set including angina pectoris, pulmonary embolism, and gastroesophageal reflux, the interdisciplinary collaborative assessment results in cardiology are as follows: (0.7 represents the cross-specialty confidence score for angina pectoris, 0.5 for pulmonary embolism, and 0.2 for gastroesophageal reflux disease; the same applies below.) The cross-specialty collaborative assessment results from the respiratory department are as follows: The interdisciplinary collaborative assessment results of the gastroenterology department were: The global collaboration result is: .

[0161] Since the overall collaborative results are the consensus reached after incorporating cross-specialty discussions, using the overall collaborative results as the fusion evaluation results can more accurately capture cross-specialty related information (such as using the respiratory department's judgment on pulmonary embolism to correct the initial bias of the cardiology department), avoiding the one-sidedness caused by the dominance of a certain specialty.

[0162] In another possible implementation, considering that the average of all cross-specialty collaborative assessment results is taken directly, multiple specialties are assumed to be of equal status. However, in reality, for any given case, the correct diagnosis will necessarily be biased towards one or more specialties. That is, multiple specialties have different levels of importance in relation to the final assessment result (i.e., the fusion assessment result).

[0163] Therefore, in this embodiment, for each candidate diagnosis in the candidate diagnosis set, the confidence score of the candidate diagnosis can be obtained from the cross-specialty collaborative assessment results corresponding to each specialty assessment result, which serves as the cross-specialty confidence score corresponding to each specialty assessment result. After obtaining the importance weights of multiple specialties, the obtained importance weights are weighted and summed with the cross-specialty confidence scores of multiple specialties to obtain the weighted score of the candidate diagnosis.

[0164] Here, the process of “obtaining the importance weights of each of the multiple specialties” may include: obtaining the importance weights of each of the multiple specialties from the normalized weight vector mentioned above.

[0165] For example, the weighted score of the d-th candidate diagnosis can be obtained using the following formula.

[0166] Formula (7);

[0167] in, This represents the weighted score of the d-th candidate diagnosis in the candidate diagnosis set. This represents the importance weight of the i-th specialty. This represents the d-th dimension in the cross-specialty collaborative assessment result corresponding to the specialty assessment result of the i-th specialty, which is also the cross-specialty confidence score corresponding to the i-th specialty.

[0168] Taking a case of "chest pain of unknown origin" as an example, the importance weights of cardiology, respiratory medicine, and gastroenterology are 0.5, 0.3, and 0.2, respectively. Based on the results of the three interdisciplinary collaborative assessments mentioned above, the following are observed:

[0169] ;

[0170] ;

[0171] .

[0172] In this embodiment, the weighted scores of each candidate diagnosis in the candidate diagnosis set can be combined to form a weighted score result. For example, this weighted score result is in vector form, denoted by P. This reflects the prior logic that places higher weight on multiple specialties.

[0173] In this embodiment, the weighted score result can be used as the fusion evaluation result, which can ensure that the final fusion evaluation result is more in line with the priority of specialty diagnosis, avoid averaging and ignoring the evaluation information of important specialties, which would lead to the deviation of the fusion evaluation result, and improve the accuracy of the fusion evaluation result.

[0174] In another possible implementation, a weight balancing factor can be introduced. The weighted balancing factor is used to weight and fuse the global collaborative results and the weighted score results to obtain the fusion evaluation results.

[0175] Taking the determination of the comprehensive confidence score of the d-th candidate diagnosis in the fusion assessment results as an example, it can optionally be calculated using the following formula.

[0176] Formula (8);

[0177] in, This represents the overall confidence score of the d-th candidate diagnosis in the fusion assessment results. This represents the weighted score of the d-th candidate diagnosis in the weighted score results. This represents the global collaboration score of the d-th candidate diagnosis in the global collaboration results.

[0178] It should be noted that if the cross-specialty diagnostic assessment method provided in this application is applied to the cross-specialty diagnostic assessment model, then the weight balancing factor... It can be a learnable parameter with an initial value of 0.5, which is continuously optimized during the model training process to adapt to the diagnostic needs of different cases, neither denying the value of multiple specialties nor ignoring the significance of cross-specialty collaboration.

[0179] Optionally, the candidate diagnoses in the candidate diagnosis set can be sorted according to the fusion assessment results. For example, the corresponding candidate diagnoses can be sorted from high to low according to the comprehensive confidence of each dimension in the fusion assessment results, and the sorting results can be output and displayed so that doctors can refer to the sorting results and make their own diagnoses in combination with their own clinical experience and other examination results.

[0180] For example, with , , For example, the fusion evaluation result can be:

[0181] ;

[0182] ;

[0183] .

[0184] Therefore, the ranking results are: [Angina pectoris 0.524, pulmonary embolism 0.446, gastroesophageal reflux 0.372].

[0185] Optionally, to improve the interpretability of the fusion evaluation results, this embodiment may also output and display the attention weight vectors and importance weights of each of the multiple specialties in the calculation process in the form of a heatmap, and / or output and display the weight balancing factor.

[0186] The aforementioned importance weight, attention weight vector, and weight balance factor correspond to the three processes of "specialty screening → collaborative interaction → result aggregation". The importance weight is used to indicate which specialties are involved in the evaluation, the attention weight vector explains how specialties interact, and the weight balance factor explains how "specialty priority" and "multi-specialty collaborative consensus" are aggregated. Users (such as doctors) can fully trace the evaluation logic by viewing these three types of information and are passively accepted by the results, which improves the interpretability and trustworthiness of the integrated evaluation results.

[0187] In some other embodiments of this application, the complete process of the above-described interdisciplinary diagnostic assessment can be implemented using an interdisciplinary diagnostic assessment model. To enable those skilled in the art to better understand this interdisciplinary diagnostic assessment model, its training process is described below.

[0188] Considering the objective differences in diagnostic preferences among different hospitals, doctors, or other users, in order to make the cross-specialty diagnostic assessment model more adaptable to individual preference differences and more accurate, this embodiment can obtain user feedback information on the fusion assessment results predicted by the model, hereinafter referred to as user feedback information.

[0189] Optionally, the user feedback information includes at least one of the following: sorting adjustment information for each candidate diagnostic sample in the candidate diagnostic sample set, candidate diagnostic samples determined by the user as the final diagnostic results of the case text samples, and adjustment information for the fusion evaluation results predicted by the cross-specialty diagnostic evaluation model. The sorting adjustment information corresponds to the sorting result before adjustment (i.e. the sorting result mentioned above), which is determined by the fusion evaluation results predicted by the model.

[0190] For example, if a doctor, based on clinical experience or other examination results, believes that the ranking results shown above may be inaccurate, the ranking results can be adjusted, for example, moving pulmonary embolism from second to first place; or, if a doctor believes that the overall confidence score of a candidate diagnosis in the fusion assessment results is inaccurate, it can be corrected to the score it believes to be correct, such as correcting the overall confidence score of angina pectoris from 0.524 to 0.9; at the same time, after referring to the fusion assessment results or ranking results, the doctor will make a final decision, for example, selecting angina pectoris as the final diagnosis.

[0191] In this embodiment, the above-mentioned types of user feedback information can be collected. Of course, other user feedback information may also be collected, and this application does not impose specific limitations.

[0192] Furthermore, this application can generate several positive sample preference pairs based on user feedback information. And the corresponding negative sample preference pair Here, each preference pair contains two candidate diagnostic samples from the candidate diagnostic sample set. For ease of description, the positive sample preference pair will be referred to as positive sample preference pair in the following text. This indicates that the corresponding negative sample preference pair is used... This means that A, which is common to both, represents the anchor point candidate diagnostic sample, and B represents the candidate diagnostic sample that forms a positive sample preference pair with anchor point A. This represents the candidate diagnostic samples that form a negative sample preference pair with anchor point A.

[0193] In this embodiment, the priority relationship between the two candidate diagnostic samples in the positive sample preference pair is the same as the priority relationship determined based on user feedback information, while the priority relationship between the two candidate diagnostic samples in the negative sample preference pair is the opposite of the priority relationship determined based on user feedback information.

[0194] For example, if the doctor adjusts the sorting (e.g., the system outputs [a,b,c], and the doctor changes it to [b,a,c]), then three positive sample preference pairs are generated: b>a, a>c, and b>c. If the doctor specifies a candidate diagnosis sample as the final diagnosis result (e.g., selecting b), then two positive sample preference pairs are generated: b>a and b>c (the confirmed diagnosis is superior to other candidates). If the doctor corrects the overall confidence score (e.g., b increases from 0.3 to 0.9, and a decreases from 0.4 to 0.2), then one positive sample preference pair (b>a) is generated.

[0195] At the same time, for each positive sample preference pair, one or more corresponding negative sample preference pairs can be generated. For example, for the preference pair b>a, c>b can be generated as a negative sample preference pair.

[0196] Therefore, in this embodiment, a contrastive learning loss can be generated based on positive sample preference pairs and negative sample preference pairs, and the cross-specialty diagnostic assessment model can be trained based on the contrastive learning loss.

[0197] For example, optionally, the contrastive learning loss can use the InfoNCE loss function, as shown in the formula:

[0198] Formula (9);

[0199] in, Indicates the contrast learning loss. This represents the similarity between A and B (calculated based on the comprehensive confidence scores of A and B predicted by the cross-specialty diagnostic assessment model). The temperature parameter is used to control the similarity discrimination, and in this embodiment it can be set to 0.1.

[0200] In this embodiment, a cross-specialty diagnostic assessment model is trained using contrastive learning loss, enabling the model to learn the ability to "have a higher similarity to the diagnosis pairs preferred by doctors," thereby quickly adapting to doctors' diagnostic habits.

[0201] Considering that when using contrastive learning loss to train a cross-specialty diagnostic assessment model, the model may forget old knowledge while learning new knowledge, in order to avoid catastrophic forgetting of the model, this embodiment also provides the following Elastic Weight Consolidation (EWC) loss.

[0202] In this embodiment, a training dataset can be obtained, where each training data point includes: case text samples, a set of candidate diagnostic samples, and corresponding fusion evaluation result labels. For example, incremental training can be triggered every 1000 feedback cases collected, resulting in a training dataset containing 1000 training data points.

[0203] Furthermore, the Fisher information matrix F of the cross-specialty diagnostic assessment model on the training dataset can be determined, and then the elastic weight consolidation loss can be generated based on the Fisher information matrix. The total loss can be generated based on the elastic weight consolidation loss and the contrastive learning loss, and the cross-specialty diagnostic assessment model can be trained using the total loss.

[0204] For example, the formula for calculating the total loss can be:

[0205] Formula (10);

[0206] in, Indicates the total loss. An element of the Fisher information matrix F represents the model parameters during incremental training. Importance of model performance This represents the initial model parameters before incremental training begins.

[0207] In one possible implementation, to avoid performance degradation of the model trained with the total loss mentioned above, this embodiment can adopt an offline training and online update method during the incremental training process. After each incremental training, multiple (e.g., 100) new cases (covering different specialties and different case types) that were not involved in the training are used for validation. The model evaluation indicators include, but are not limited to, the following indicators: the ranking accuracy of candidate diagnoses, the identification rate of high-risk diseases, and the consistency with doctors' decisions.

[0208] If at least one of the above model evaluation metrics improves by more than the first percentage threshold (e.g., 2%) compared to the previous model version and no evaluation metric decreases, then the online model is updated according to the model parameters trained offline. Otherwise, the model parameters are rolled back and restored to the model parameters of the previous version.

[0209] Of course, other strategies can also be used to decide whether to update model parameters or not. For example, if at least one of the multiple model evaluation metrics decreases by more than the second preset number (e.g., 1) compared to the previous model version by more than the second percentage threshold (e.g., 5%), then the model parameters will be rolled back directly to restore the model parameters of the previous version. Conversely, if neither of the above two rules is met, then the offline training model parameters will be retained but not updated to the online model, so that the next incremental training can be carried out based on the retained model parameters. If the above two rules are not met for a total of 3 times, then the model parameters will be rolled back to restore the model parameters of the previous version to avoid frequent updates that may cause model instability.

[0210] This application's embodiments employ a comparative learning approach between EWC and InfoNCE, combined with a verification mechanism of "updating when assessment indicators meet the criteria and rolling back when they do not," which solves the problem that static models cannot adapt to clinical changes and improves the user experience.

[0211] The above describes a cross-specialty diagnostic assessment method provided by the embodiments of this application. The following describes the apparatus for performing the above-described cross-specialty diagnostic assessment method.

[0212] Please see Figure 4 , Figure 4 This is a schematic diagram of a cross-specialty diagnostic assessment device provided in an embodiment of this application. Figure 4 As shown, the interdisciplinary diagnostic assessment device may include:

[0213] The information acquisition unit 401 is used to acquire target case texts involving multiple specialties in pathogenesis, as well as the corresponding set of candidate diagnoses;

[0214] Independent evaluation unit 402 is used to independently evaluate the candidate diagnosis set based on the target case text using the specialty evaluation model of each specialty, and obtain the specialty evaluation result of each specialty. The specialty evaluation result is used to characterize the confidence distribution of each candidate diagnosis in the candidate diagnosis set under the medical logic of the corresponding specialty.

[0215] The collaborative assessment unit 403 is used to conduct cross-specialty collaborative assessment based on all specialty assessment results, and obtain the cross-specialty collaborative assessment result corresponding to each specialty assessment result.

[0216] Information fusion unit 404 is used to generate fusion assessment results based on all cross-specialty collaborative assessment results. The fusion assessment results are used to characterize the overall confidence distribution of each candidate diagnosis in the candidate diagnosis set after cross-specialty collaborative assessment.

[0217] Each module in the aforementioned interdisciplinary diagnostic assessment device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0218] This application also provides an electronic device, which may include at least one processor and a memory connected to the processor, wherein:

[0219] Memory is used to store computer programs;

[0220] The processor is used to execute computer programs to enable the electronic device to implement any of the cross-specialty diagnostic assessment methods provided in the embodiments of this application.

[0221] refer to Figure 5 The diagram illustrates a structural schematic suitable for implementing the electronic device in the embodiments of this application. The electronic device in the embodiments of this application may include, but is not limited to, fixed terminals such as mobile phones, laptops, PDAs (personal digital assistants), PADs (tablet computers), desktop computers, etc. Figure 5 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0222] like Figure 5 As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. When the electronic device is powered on, the RAM 603 also stores various programs and data required for the operation of the electronic device. The processing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0223] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 608 including, for example, memory cards, hard drives, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 5 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have instead.

[0224] This application also provides a computer program product including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement any of the cross-specialty diagnostic assessment methods provided in this application.

[0225] This application also provides a computer-readable storage medium carrying one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the cross-specialty diagnostic assessment methods provided in this application.

[0226] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0227] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0228] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0229] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. A cross-specialty diagnostic assessment method, characterized in that, include: Obtain target case texts involving multiple specialties related to pathogenesis, along with corresponding candidate diagnostic sets; Using the specialty assessment model of each specialty, the candidate diagnosis set is independently evaluated based on the target case text to obtain the specialty assessment result of each specialty. The specialty assessment result is used to characterize the confidence distribution of each candidate diagnosis in the candidate diagnosis set under the medical logic of the corresponding specialty. Based on all the specialty assessment results, cross-specialty collaborative assessment is conducted to obtain the cross-specialty collaborative assessment result corresponding to each of the aforementioned specialty assessment results; A fusion assessment result is generated based on all the cross-specialty collaborative assessment results. The fusion assessment result is used to characterize the overall confidence distribution of each candidate diagnosis in the candidate diagnosis set after cross-specialty collaborative assessment.

2. The cross-specialty diagnostic assessment method according to claim 1, characterized in that, Both the specialty assessment results and the cross-specialty collaborative assessment results are in vector form. The cross-specialty collaborative assessment based on all specialty assessment results yields a cross-specialty collaborative assessment result corresponding to each specialty assessment result, including: Linear mapping and encoding are performed on all the specialty assessment results to obtain high-dimensional assessment vectors corresponding to each specialty assessment result. The dimension of the high-dimensional assessment vectors is greater than the dimension of the specialty assessment results. Generate the query matrix and key matrix based on all the high-dimensional evaluation vectors; The attention score matrix is ​​calculated based on the query matrix and the key matrix, and the attention score matrix is ​​normalized to obtain the attention weight matrix. Each row vector of the attention weight matrix is ​​used as an attention weight vector to obtain the attention weight vector of each of the multiple specialties. The attention weight vector is used to quantify the dependence weight of the corresponding specialty on each high-dimensional evaluation result in the cross-specialty collaborative evaluation. For each specialty, based on the specialty's attention weight vector and all the high-dimensional evaluation vectors, a cross-specialty collaborative evaluation result corresponding to the specialty's evaluation result is generated.

3. The cross-specialty diagnostic assessment method according to claim 2, characterized in that, The process of generating cross-specialty collaborative assessment results corresponding to the specialty assessment results based on the specialty's attention weight vector and all the high-dimensional assessment vectors includes: The attention weight vector of the specialty is weighted and summed with all the high-dimensional evaluation vectors to obtain the collaborative state vector of the specialty. The collaborative state vector represents the evaluation deviation of the independent evaluation of the corresponding specialty compared with the cross-specialty collaborative evaluation. By using the collaborative state vector of the specialty to compensate for the bias in the high-dimensional evaluation vector corresponding to the specialty evaluation result, the compensated state vector of the specialty is obtained. The compensated state vector of the specialty is subjected to layer normalization to obtain the normalized state vector of the specialty. The normalized state vector of the specialty is linearly mapped to a space of target dimension to obtain the cross-specialty collaborative evaluation result corresponding to the specialty evaluation result. The target dimension is the total number of candidate diagnoses included in the candidate diagnosis set.

4. The cross-specialty diagnostic assessment method according to claim 1, characterized in that, The process of generating a fusion assessment result based on all the cross-specialty collaborative assessment results includes: The mean of all the cross-specialty collaborative assessment results across the candidate diagnostic dimension is calculated to obtain the global collaborative result; For each candidate diagnosis in the candidate diagnosis set: The confidence score of the candidate diagnosis is obtained from the cross-specialty collaborative assessment results corresponding to each of the aforementioned specialty assessment results, and is used as the cross-specialty confidence score corresponding to the specialty for each of the aforementioned specialty assessment results. Obtain the importance weights of each of the multiple specialties, and then sum the obtained importance weights with the cross-specialty confidence scores of each of the multiple specialties to obtain the weighted score of the candidate diagnosis. To obtain a weighted score result composed of the weighted scores of each candidate diagnosis in the candidate diagnosis set; The global collaborative results and the weighted score results are weighted and fused in the candidate diagnostic dimension using a weight balancing factor to obtain the fusion evaluation result.

5. The cross-specialty diagnostic assessment method according to claim 4, characterized in that, The process of obtaining the importance weights of each of the multiple specialties includes: The target case text is semantically encoded using a configured semantic encoding model to obtain a global semantic feature vector, which represents the semantic information related to the diagnostic content in the target case text. The global semantic feature vector is transformed and mapped using a configured multilayer perceptron to obtain a normalized weight vector with a dimension equal to the total number of specialties, wherein the total number of specialties is greater than or equal to the number of the multiple specialties. The importance weights of each of the multiple specialties are obtained from the normalized weight vector.

6. The cross-specialty diagnostic assessment method according to claim 5, characterized in that, The step of using a configured multilayer perceptron to perform feature transformation mapping on the global semantic feature vector to obtain a normalized weight vector with a dimension equal to the total number of specialties includes: Semantic information related to specialties is extracted from the global semantic feature vector using a first fully connected layer and an activation function to obtain an intermediate semantic feature vector. The dimension of the intermediate semantic feature vector is less than the dimension of the global semantic feature vector and greater than the total number of specialties. The intermediate semantic feature vector is mapped to an initial weight vector with a dimension equal to the total number of specialties using a second fully connected layer; The initial weight vector is normalized using a preset normalization function to obtain the normalized weight vector.

7. The interdisciplinary diagnostic assessment method according to any one of claims 1-6, characterized in that, The cross-specialty diagnostic assessment method is applied to the cross-specialty diagnostic assessment model, and the training process of the cross-specialty diagnostic assessment model includes: Obtain user feedback information, which includes at least one of the following: sorting adjustment information for each candidate diagnostic sample in the candidate diagnostic sample set, candidate diagnostic samples determined by the user as the final diagnostic result of the case text sample, and adjustment information for the fusion evaluation result predicted by the cross-specialty diagnostic evaluation model, wherein the sorting adjustment information corresponds to the sorting result before adjustment, which is determined by the predicted fusion evaluation result. Generate positive sample preference pairs and corresponding negative sample preference pairs. The priority relationship between the two candidate diagnostic samples in the positive sample preference pair is the same as the priority relationship determined based on the user feedback information. The priority relationship between the two candidate diagnostic samples in the negative sample preference pair is the opposite of the priority relationship determined based on the user feedback information. A contrastive learning loss is generated based on the positive sample preference pair and the negative sample preference pair; The cross-specialty diagnostic assessment model is trained using the contrastive learning loss.

8. The cross-specialty diagnostic assessment method according to claim 7, characterized in that, The step of training the cross-specialty diagnostic assessment model based on the contrastive learning loss includes: Obtain a training dataset, wherein each training data in the training dataset includes: the case text sample, the candidate diagnosis sample set, and the corresponding fusion evaluation result label; Determine the Fisher information matrix of the cross-specialty diagnostic assessment model on the training dataset; Generate a resilient weighted consolidation loss based on the Fisher information matrix; The total loss is generated based on the elastic weight consolidation loss and the contrastive learning loss; The cross-specialty diagnostic assessment model is trained using the total loss.

9. The cross-specialty diagnostic assessment method according to claim 4, characterized in that, Also includes: Based on the fusion evaluation results, the candidate diagnoses in the candidate diagnosis set are sorted, and the sorting results are output and displayed. And / or, output and display the attention weight vectors and importance weights of each of the multiple specialties in the form of a heatmap; And / or, display the weight balancing factor output.

10. A computer program product, characterized in that, Includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the cross-specialty diagnostic assessment method as described in any one of claims 1 to 9.

11. An electronic device, characterized in that, It includes at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program to enable the electronic device to implement the cross-specialty diagnostic assessment method as described in any one of claims 1 to 9.

12. A computer storage medium, characterized in that, The storage medium carries one or more computer programs that, when executed by an electronic device, enable the electronic device to implement the cross-specialty diagnostic assessment method as described in any one of claims 1 to 9.

Images