A method and system for constructing a reasoning agent for intelligent medical guidance

Abstract

The present invention relates to the technical field of artificial intelligence and medical decision-making systems, and discloses a method and system for constructing a reasoning agent for intelligent medical guidance. The method comprises: constructing a hierarchical modality completion network and a modality correlation knowledge graph; constructing a medical feature cross-modality mapping network based on contrastive learning to map medical data of different modalities to a shared feature space; constructing an uncertainty quantification model and automatically adjusting the diagnostic confidence threshold based on the uncertainty of the completed data; constructing a disease modality association matrix and a modality reliability assessment network for different disease types and symptom combinations; and realizing cross-validation and collaborative decision-making of multi-specialty knowledge through weighted voting, evidence aggregation and specialist authority assessment to form a final diagnostic recommendation. The present invention solves the problem of limited diagnostic capabilities caused by missing medical data in an environment with limited medical resources.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and medical decision-making systems, and specifically discloses a method and system for constructing a reasoning agent for intelligent medical guidance. Background Art

[0002] As the application of artificial intelligence (AI) technology in the medical field continues to deepen, intelligent medical systems are playing an increasingly important role in assisting diagnosis and formulating treatment plans. However, in areas where medical resources are unevenly distributed, particularly in underdeveloped and remote areas, the lack of high-end medical equipment often makes it difficult to obtain complete multimodal medical data, which seriously restricts the effectiveness of existing intelligent medical systems.

[0003] Currently, existing intelligent medical systems face significant limitations when processing incomplete medical data. For one thing, most systems rely on complete multimodal data input, and their performance plummets when data is missing. Furthermore, existing cross-modal learning methods typically focus solely on mapping a single modality to another, failing to build comprehensive multimodal data completion capabilities. Furthermore, existing systems lack a reliability assessment mechanism for completed data, leading to unreliable reasoning based on completed data and an inability to dynamically adjust reliance on data from different modalities based on different disease types and symptom combinations.

[0004] Therefore, how to build an intelligent medical system that can process incomplete medical data and provide high-quality diagnostic services in an environment with limited medical resources has become a technical problem that needs to be solved urgently. Summary of the Invention

[0005] The present invention provides a method and system for constructing a reasoning agent for intelligent medical guidance, which solves the technical problem in related technologies of limited diagnostic capabilities caused by the inability to obtain complete multimodal data in areas with limited medical resources.

[0006] The first aspect of the present invention provides a method for constructing a reasoning agent for intelligent medical guidance, which specifically includes the following steps:

[0007] Construct a hierarchical modality completion network and modality correlation knowledge graph to achieve intelligent completion and quality assessment of missing medical data modalities in an environment with limited medical resources;

[0008] Based on the completed medical data, contrastive learning is used to build a medical feature cross-modal mapping network, mapping different modal medical data into a shared feature space to achieve multimodal medical data fusion and self-calibration;

[0009] For the fused multimodal data, an uncertainty quantification model is constructed based on the uncertainty of the supplemented data, and the diagnostic confidence threshold is automatically adjusted to achieve adaptive decision threshold adjustment;

[0010] Combined with the adjusted confidence threshold, a disease modality association matrix and a modality reliability assessment network are constructed for different disease types and symptom combinations, dynamically adjusting the weight of each modality in the inference process;

[0011] Based on dynamically adjusted modal weights, through weighted voting, evidence aggregation and expert authority assessment, cross-validation and collaborative decision-making of multi-specialty knowledge are achieved to form final diagnostic recommendations.

[0012] Furthermore, the steps of constructing a hierarchical modality completion network and a modality correlation knowledge graph include:

[0013] Construct a modality completion subnetwork with an encoder-decoder structure to complete different types of missing medical data;

[0014] By analyzing large-scale medical data sets, we quantify the mutual predictive power and dependencies between different medical data modalities and construct a modality correlation knowledge graph.

[0015] Perform quality assessment on the generated modal data and quantify its reliability and uncertainty bounds through the quality assessment function:

[0016] ;

[0017] in, For the generated data, is the observed data, Evaluate the rationality of the content, Assess consistency with observed data, Assess the degree of compliance with medical laws, 、 、 They represent the weight coefficients for evaluating the rationality of the content, the consistency with observed data, and the degree of compliance with medical laws, respectively, and meet the , represents the quality assessment function of the generated data conditioned on the observed data.

[0018] Furthermore, the steps of constructing a medical feature cross-modal mapping network based on contrastive learning include:

[0019] Construct a medical feature cross-modal mapping network to map medical data of different modalities into a shared feature space;

[0020] By optimizing the network parameters through contrast loss function and medical feature fidelity loss, the mapping results of different modal data from the same medical case are similar.

[0021] Construct a modal reliability assessment network to automatically assess the reliability of each modal data by analyzing historical accuracy and current information completeness;

[0022] Implement inter-modal consistency scoring and self-calibration mechanism. When the inter-modal consistency score is detected to be lower than the threshold, the self-calibration mechanism is triggered to adjust the weight of each modality or request additional information.

[0023] Furthermore, the step of automatically adjusting the diagnosis confidence threshold according to the uncertainty of the completed data includes:

[0024] Construct an uncertainty quantification model based on the Monte Carlo random dropout method to quantify epistemic uncertainty and random uncertainty;

[0025] Adaptively adjust the diagnostic decision threshold based on the completed data quality score and the output of the uncertainty quantification model;

[0026] Implement a progressive reasoning strategy based on completion quality, divide the completion data into three categories according to their quality: high quality, medium quality, and low quality, and adopt different reasoning strategies.

[0027] Furthermore, the steps for implementing dynamic modal weight optimization for different disease types and symptom combinations include:

[0028] Construct disease modality association matrices and symptom modality association matrices to quantify the importance of different medical data modalities for the diagnosis of various diseases;

[0029] Based on the disease probability distribution and association matrix, the weight of each modality in the reasoning process is dynamically adjusted;

[0030] Normalize the adjusted weights to ensure that the sum of all weights is 1, and apply them to multimodal feature fusion;

[0031] Computing resources are dynamically allocated according to modal weights, high-weight modes are deeply analyzed, and low-weight modes are simplified.

[0032] Furthermore, the steps to achieve multi-specialty knowledge collaboration through weighted voting, evidence aggregation, and specialist authority assessment include:

[0033] Build dedicated knowledge representation models for different medical specialties, including specialty-specific concept systems, reasoning rule sets, and evidence weight matrices;

[0034] Based on multimodal medical data, each specialty knowledge model is activated to perform independent reasoning and generate specialty-specific diagnostic hypotheses and confidence levels;

[0035] Conduct specialist cross-validation, calculate the diagnostic consistency matrix, and identify diagnostic conflicts and complementary information between specialties;

[0036] The authoritative score of each specialty in the current case is calculated, and the final collaborative diagnosis result is formed through a weighted voting mechanism.

[0037] Furthermore, the progressive reasoning strategy is specifically as follows:

[0038] For high-quality completed regions, conventional reasoning strategies are used to make diagnostic decisions directly based on fused features;

[0039] For medium-quality completed regions, a conservative reasoning strategy is adopted to increase the decision threshold and generate multiple possible diagnostic hypotheses simultaneously;

[0040] For low-quality completed areas, the system refuses to make a definitive diagnosis and instead provides preliminary analysis results and recommendations for additional examinations;

[0041] For critical decisions involving life safety, set minimum data quality requirements. When the quality of all available data is below this threshold, it is recommended that the patient be referred to a higher-level medical institution.

[0042] Furthermore, the specialized knowledge model includes:

[0043] Specialty-specific conceptual frameworks, including commonly used medical terminology, disease classifications, and diagnostic criteria;

[0044] A set of inference rules, including unique diagnostic ideas and decision rules;

[0045] The weight of evidence matrix defines the importance of different types of evidence in the diagnosis of this specialty;

[0046] The concept mapping relationship between specialties connects the concept systems of different specialties through equivalence relationships, inclusion relationships and correlation relationships.

[0047] Furthermore, the method further includes the steps of automatically generating graded medical recommendations:

[0048] For high-confidence diagnoses, a complete treatment plan is provided;

[0049] For moderate-confidence diagnoses, preliminary treatment recommendations are provided and close follow-up is recommended;

[0050] For low-confidence diagnoses, provide symptom management advice and recommend further testing;

[0051] Intelligently adjust the level of detail and urgency of treatment recommendations based on different disease types and severity.

[0052] A second aspect of the present invention further provides a system for constructing a reasoning agent for intelligent medical guidance, which is used to execute the above-mentioned method for constructing a reasoning agent for intelligent medical guidance, including:

[0053] A hierarchical modality completion module, which is used to intelligently complete and assess the quality of missing medical data modalities in resource-constrained environments;

[0054] Medical feature cross-modal mapping module, used to map different modal medical data into a shared feature space, achieving multimodal medical data fusion and self-calibration;

[0055] Adaptive decision threshold adjustment module, used to automatically adjust the diagnosis confidence threshold according to the uncertainty of the completed data;

[0056] Dynamic modality weight optimization module, which is used to dynamically adjust the weight of each modality in the inference process according to different disease types and symptom combinations;

[0057] The multi-disciplinary knowledge collaboration module is used to achieve cross-validation and collaborative decision-making of multi-disciplinary knowledge through weighted voting, evidence aggregation and expert authority evaluation.

[0058] The beneficial effects of the present invention are: through modal completion and quality assessment technology, when only partial modal data is available, the system can still provide diagnostic services under conditions close to complete data; through dynamic modal weight optimization technology, the system can intelligently adjust the weights of each modality according to different disease types and symptom combinations, and has stronger adaptability to different disease manifestations; based on the uncertainty of the completed data, the diagnostic confidence threshold is automatically adjusted, and combined with the progressive reasoning strategy, the system can ensure the accuracy of the diagnosis while clearly pointing out the items that need further examination, avoiding overconfident misdiagnosis; through the multi-specialty knowledge collaboration mechanism, the system can analyze the condition from different specialist perspectives, reducing the limitations of a single specialist perspective; grassroots medical institutions do not need to be equipped with a full set of high-end medical equipment, and can also provide preliminary diagnostic services close to the level of tertiary hospitals through this system, greatly improving the efficiency of medical resource utilization and reducing patients' medical costs and time; the system reduces unnecessary referrals by accurately identifying cases that need referral. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A flowchart of a method for constructing a reasoning agent for intelligent medical guidance provided by the present invention. DETAILED DESCRIPTION

[0060] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0061] At least one embodiment of the present invention discloses a method for constructing a reasoning agent for intelligent medical guidance, such as Figure 1 As shown, the following steps are included:

[0062] Step 1: Construct a hierarchical modality completion network and modality correlation knowledge graph to achieve intelligent completion and quality assessment of missing medical data modalities in a resource-constrained environment;

[0063] It includes the following sub-steps:

[0064] Step 1.1: Construct a hierarchical modality completion network;

[0065] Dedicated modality completion subnetworks are constructed for different types of missing medical data (such as medical imaging and laboratory tests). Each subnetwork adopts an encoder-decoder structure, where the encoder converts the existing medical data modality into a latent feature representation, and the decoder generates data for the missing modality based on this feature representation.

[0066] In some embodiments, encoder-decoder combinations with different architectures can be configured for different medical scenarios. For example, for the common need of generating X-ray images from text descriptions in primary healthcare institutions, a Transformer-based text encoder and a U-Net-based image generator can be used. For scenarios requiring complex examination results to be predicted from simple examination items, a lightweight MLP encoder and a specialized disease predictor can be used.

[0067] For common medical data modality pairs, such as generating medical images from text descriptions or predicting laboratory test results from medical images, this application constructs corresponding modality conversion paths. For the text-to-image path, a text encoder and image generator are used; for the image-to-lab test result path, an image encoder and a numerical feature predictor are used; and for the structured data-to-time series physiological signal path, a structured data encoder and a time series signal generator are used.

[0068] Each sub-network learns collaboratively by jointly training the optimization function:

[0069] ;

[0070] in Indicates the The loss function of the modality completion sub-network, is the corresponding weight coefficient, is the total number of modal types. Here, represents the joint loss function, represents the summation symbol, Indicates the modal index, from 1 to Traverse all modes. Optionally, in practical applications, the weight coefficients can be dynamically adjusted according to the importance of different modes. ,give higher weights to modalities that are more critical for diagnosis, for example, in the diagnosis of heart disease, the electrocardiogram modality can be given a higher weight, while for respiratory diseases, the lung imaging modality is given a higher weight;

[0071] The loss function of each sub-network Different calculation methods can be used according to the specific completion task, including reconstruction error (such as mean square error MSE), adversarial loss (such as GAN loss), perceptual loss, etc. For example, for medical image modality completion, It can be defined as:

[0072] ;

[0073] in, For the The data generated by each modality, For real data, represents the mean square error, is the adversarial loss (such as the binary cross entropy loss of the discriminator output), is the perceptual loss (e.g., distance based on feature space). represents the weight coefficient of the reconstruction error term, represents the weight coefficient of the adversarial loss term, Represents the weight coefficient of the perceptual loss term. For structured data or time series signal modalities, The reconstruction error alone can be used or combined with regularization terms such as KL divergence.

[0074] Step 1.2: Construct a modality-related knowledge graph;

[0075] By analyzing large-scale medical datasets, we quantify the mutual predictive power and dependencies between different medical data modalities and construct a modality-dependent knowledge graph. The nodes in this knowledge graph represent different medical data modalities (such as CT images, MRI images, blood test results, and pathology reports), while the edges represent the predictive relationships between modalities. The weights of the edges indicate the strength of the predictive power.

[0076] Edge weight Calculated by the following formula:

[0077] ;

[0078] in, Indicates in the dataset Using modal Prediction mode The accuracy of is the total number of data sets, and represent the source mode and target mode respectively, where represents the source modality (input modality) used for prediction, represents the target mode to be predicted (output mode), Indicates that the modal To Modal The strength of the predictive ability, Indicates the Medical datasets.

[0079] It should be understood that based on the modality correlation knowledge graph, when certain modality data is missing, the system can determine the optimal completion path. For example, when multiple modalities need to be completed at the same time and there are multiple possible paths, the completion path with the largest path weight product is selected.

[0080] The knowledge graph can be further integrated with expert domain knowledge for fine-tuning. For example, although from a statistical point of view the accuracy of predicting blood test results from X-rays may be lower than that from symptoms, for specific diseases such as tuberculosis, X-ray images have strong predictive value for indicators such as erythrocyte sedimentation rate and C-reactive protein. In this case, the expert scoring coefficient can be used to Adjust edge weights:

[0081] ;

[0082] in represents the final adjusted edge weight, represents the initial edge weights calculated based on the data, represents the expert scoring coefficient, It is a coefficient that balances data-driven and expert knowledge, which can be adjusted in different application scenarios. For scenarios with sufficient data but limited expert resources, a smaller For diseases with abundant expert experience but limited data, a larger value can be set. value (such as 0.6).

[0083] Step 1.3: Perform quality assessment on the generated modal data;

[0084] The quality of medical data generated by the modality completion network is evaluated to quantify its reliability and uncertainty bounds. The quality assessment function is defined as:

[0085] ;

[0086] in, For the generated data, is the observed data, Evaluate the rationality of the content, Assess consistency with observed data, Assess the degree of compliance with medical laws, 、 、 They represent the weight coefficients for evaluating the rationality of the content, the consistency with observed data, and the degree of compliance with medical laws, respectively, and meet the , represents the quality assessment function of the generated data conditioned on the observed data.

[0087] In different application scenarios, the weight coefficient can be optionally adjusted. For example, for the diagnosis of rare diseases, due to limited clinical data, the conformity of medical laws can be increased. The weight (if =0.5), which relies more on medical knowledge; for common diseases, consistency assessment can be added The weight (if =0.5), which relies more on actual observation data.

[0088] Content rationality Evaluate whether the generated data is medically reasonable. For example, for medical images, check whether the anatomical structure of the image is correct; for laboratory test results, check whether the values ​​are within a reasonable range. For medical images, this can be achieved through a pre-trained medical image quality assessment network. This network is trained on a large number of real medical images and can evaluate the anatomical accuracy, contrast, and clarity of the generated images. For laboratory test data, this is achieved by checking whether the generated values ​​are within the normal reference range or the typical range of a specific disease.

[0089] Conformity Assessment Checking for medical discrepancies between generated data and observed data, such as whether the generated images match the patient's age, gender, and symptom description. Optionally, this evaluation can be performed using a cross-modal consistency verification algorithm, which learns the inherent correlations between different modalities by analyzing large amounts of paired multimodal medical data.

[0090] Compliance with medical laws Verify that the generated data matches typical presentations of a specific disease by comparing it to standard cases in a medical knowledge base. This evaluation can be based on a pre-trained medical knowledge reasoning system that integrates knowledge from disease guidelines and textbooks to determine whether the generated data matches the presentation of the disease described in medical literature.

[0091] Furthermore, based on the evaluation results, the system assigns a quality score to each generated medical data set and sets a confidence interval for subsequent decision adjustments during the reasoning process. Quality scores are categorized as high, medium, and low, corresponding to different processing strategies in subsequent reasoning. The scoring also considers the timeliness and source reliability of the data. Data predicted using highly reliable modalities or with a shorter prediction time span receive higher scores, while data that relies on multiple intermediate predictions or has a longer prediction time span receives lower scores.

[0092] Step 2: Based on the completed medical data, a medical feature cross-modal mapping network is constructed using contrastive learning to map different modal medical data into a shared feature space, achieving multimodal medical data fusion and self-calibration.

[0093] It includes the following sub-steps:

[0094] Step 2.1: Construct a medical feature cross-modal mapping network;

[0095] Based on the contrastive learning method, a medical feature cross-modal mapping network is constructed, which maps medical data of different modalities into a shared feature space, so that semantically related data are closer in the feature space, while unrelated data are farther away.

[0096] It should be understood that the network consists of feature extractors and mappers for each modality. , feature extractor Extracting modality-specific feature representations , mapper Mapping feature representations to a shared feature space ;in, Indicates the Medical data of various modalities, Indicates the The feature extractor of the modality, represents the extracted modality-specific feature representation, Indicates the feature mappers of different modalities, Represents the feature vector mapped to the shared feature space.

[0097] In some embodiments, different feature extractor structures can be used for different types of medical data. For example, for medical imaging data, a convolutional neural network (CNN) architecture suitable for processing visual information, such as ResNet or DenseNet, can be used; for medical text records, the Transformer architecture used in natural language processing can be used; for time-series physiological signals, a long short-term memory network (LSTM) or a one-dimensional convolutional network can be used. All of these domain-specific feature extractors are trained together to ensure that features from different modalities are mapped into the same semantic space.

[0098] This application uses the contrast loss function Optimize network parameters so that mapping results for different modalities of data from the same medical case are similar, while mapping results for data from different cases vary greatly:

[0099] ;

[0100] in, represents the contrast loss function, Representing different modal features in the same medical case and The similarity of represents negative sample features from other medical cases, represents the number of negative samples, Indicates that the temperature parameter is used to adjust the similarity distribution, represents the exponential operation, Represents logarithmic operation;

[0101] ;

[0102] in, represents the total loss function, represents the contrast loss function, represents the loss of medical feature fidelity, represents the disease classification task loss, represents the loss for the anatomical structure segmentation task. 、 、 and Respectively represent the importance of contrastive learning loss, medical feature fidelity loss, disease classification loss, and anatomical structure segmentation loss in the total loss function;

[0103] Medical feature fidelity loss The mean square error (MSE) or L1 loss can be used to measure the difference between the generated features and the true features. The formula is as follows:

[0104] ;

[0105] in, Indicates the Medical features generated by samples, represents the corresponding real medical features, is the number of samples, represents the Euclidean norm (L2 norm).

[0106] Disease classification loss The cross entropy loss function is usually used to calculate the difference between the predicted category probability and the true category label. The formula is as follows:

[0107] ;

[0108] in, Indicates the The sample in The true label of the class (one-hot encoding), Indicates that the model predicts the The probability of the class, is the number of categories, is the number of samples, Represents the natural logarithm.

[0109] Anatomy segmentation loss Dice loss or cross entropy loss can be used. The commonly used Dice loss formula is as follows:

[0110] ;

[0111] in, Indicates the The segmentation region predicted by samples, represents the corresponding true segmentation area, Representing a collection The number of elements of Represents the intersection of sets.

[0112] This multi-task learning approach ensures that the shared feature space not only captures the correspondence between modalities, but also preserves medically critical information such as disease diagnosis and anatomical structure recognition.

[0113] Step 2.2, construct the modal reliability assessment network;

[0114] A modal reliability assessment network is constructed to automatically assess the reliability of each modal data by analyzing historical accuracy and current information completeness. The network receives the quality indicators of each modal data as input and outputs a modal reliability score.

[0115] It should be noted that the modal reliability score Calculated by the following function:

[0116] ;

[0117] in, is modal The historical accuracy of is the completeness of the current data, is the noise level, 、 and They represent the weight coefficients of historical accuracy, data completeness, and noise level, respectively. It is the sigmoid function that limits the scores to the range [0, 1].

[0118] Historical accuracy By recording the modal Contribution to past diagnosis; completeness Indicates the information range covered by the data; the degree of noise Evaluated through statistical properties of data and anomaly detection algorithms.

[0119] Modality reliability can also be dynamically adjusted based on specific disease types. For example, for fracture diagnosis, the reliability of X-ray images should be higher than laboratory test results; and for diabetes, the reliability of blood glucose testing should be higher than that of ordinary X-ray images. To achieve this, disease-specific score adjustments can be introduced:

[0120] ;

[0121] in, It is for disease Adjusted modality Reliability rating, represents the modal reliability score, is modal For disease The specificity coefficient was obtained through large-scale clinical data analysis.

[0122] Optionally, the system can further adjust the reliability score based on factors such as device type and the qualifications of the operating technician. For example, data acquired by advanced equipment in a high-level hospital could be assigned a higher initial reliability score, while data acquired by equipment operated by less-trained personnel could receive a correspondingly lower score. This multi-factor assessment mechanism makes the reliability evaluation more comprehensive and accurate.

[0123] Step 2.3: Implement inter-modality consistency scoring and self-calibration mechanism;

[0124] Construct an inter-modal consistency scoring mechanism to detect conflicts between different modal information and consistency scoring The calculation is as follows:

[0125] ;

[0126] in, is the mode in the shared feature space and modal The distance of the feature, is the normalization factor. Indicates modality and modal The consistency score between the two modalities ranges from [0 to 1]. The larger the value, the higher the consistency between the two modalities. and Represents the mode and modal Feature representation in a shared feature space.

[0127] Optionally, a more sophisticated consistency score calculation method can be used to account for the semantic connection between modalities. For example, a stricter consistency threshold can be used for modality pairs with close semantic connections, such as medical images and radiology reports, while a less stringent threshold can be used for modality pairs with looser semantic connections, such as symptom descriptions and laboratory test results. This differentiated approach can more accurately identify true modality conflicts.

[0128] When the inter-modality consistency score is detected to be lower than the threshold When , the self-calibration mechanism is triggered, and the weights of each modality are adjusted or additional information is requested through Bayesian uncertainty quantification. The self-calibration process mainly includes:

[0129] Calculate the reliability score for each modality ,in Indicates the modalities;

[0130] Adjust modal fusion weights according to reliability scores:

[0131] ;

[0132] in, represents the total number of modes, Indicates the The fusion weight of each modality, Indicates the modal, Indicates the Reliability scores for each modality;

[0133] If the reliability scores of all modalities are below the threshold , the system will request additional information, including represents the minimum threshold for reliability score;

[0134] Recalculate the fusion features based on the adjusted weights:

[0135] ;

[0136] in, represents the fused features, Indicates the The characteristics of a mode, represents the total number of modes, Indicates the The fusion weight of each modality.

[0137] The system can implement adaptive calibration strategy selection, adopting different calibration methods based on the nature and severity of the conflict. For example, when a minor conflict is detected, the system only adjusts the modal weights; when a serious conflict is found, the system may need to re-collect data or introduce additional diagnostic rules for arbitration. The severity can be assessed by combining the importance of the conflicting modal and the degree of conflict:

[0138] ;

[0139] in, Indicates modality and modal The comprehensive importance of the diagnosis task can be dynamically determined according to the criticality of the diagnosis task. Indicates the severity of the modal conflict, Indicates modality and modal The degree of inconsistency between Score the intermodality agreement defined above.

[0140] Optionally, the system also implements a conflict resolution memory mechanism, recording historical conflict cases and their resolutions to form a case library. When encountering new conflict situations, the system can retrieve similar historical cases and refer to their resolutions, implementing experience-driven conflict resolution, thereby continuously improving the system's ability to handle modal conflicts.

[0141] Therefore, through this self-calibration mechanism, the system can dynamically adjust its reliance on different modal data, reduce misdiagnosis caused by modal conflicts, and continuously accumulate experience during use to improve conflict handling capabilities.

[0142] Step 3: For the fused multimodal data, based on the uncertainty of the supplemented data, an uncertainty quantification model is constructed and the diagnostic confidence threshold is automatically adjusted to achieve adaptive decision threshold adjustment;

[0143] It includes the following sub-steps:

[0144] Step 3.1: Construct uncertainty quantification model;

[0145] We build an uncertainty quantification model to accurately assess the uncertainty of the completed data. This model, based on the Monte Carlo Dropout method, quantifies the uncertainty during the inference process by keeping the Dropout layer activated during the inference phase and performing multiple forward propagations and statistical analysis.

[0146] It should be understood that uncertainty can be divided into epistemic uncertainty and aleatoric uncertainty. Reflecting the lack of model knowledge, the variance can be calculated by inferring multiple samplings:

[0147] ;

[0148] in, represents epistemic uncertainty, It is The output of the forward propagation, yes The average of the forward propagation times, Indicates the number of forward propagation times, which can be set to 30. Represents input data, Indicates that the neural network model using random dropout is Function mapping in the forward propagation.

[0149] Optionally, different sampling strategies can be used to enhance uncertainty estimation. For example, for complex medical imaging diagnosis tasks, a higher sampling rate (e.g. ) to obtain a more accurate uncertainty estimate; for clinical decision-making scenarios with high real-time requirements, a lower sampling number (such as ) to improve response speed, while combining confidence interval correction technology to ensure assessment accuracy.

[0150] Random uncertainty Reflect the noise and variability of the data itself, and directly predict the variance of the data distribution through the model get:

[0151] ;

[0152] in, represents random uncertainty, The variance of the data distribution reflects the inherent noise level and variability of the data. This variance is usually directly predicted by the model as part of the model output and is used to quantify the inherent uncertainty of the data.

[0153] For specific types of medical data, such as medical imaging, the system can further assess random uncertainty at the data acquisition level, taking into account factors such as the resolution of the imaging device, scanning conditions, and image quality. For example, for low-dose CT scans, the system can automatically estimate the image noise level based on the recorded radiation dose and reconstruction algorithm; for ultrasound examinations, the system can assess the imaging quality based on factors such as acoustic window conditions and examination time.

[0154] Total uncertainty is the sum of the two:

[0155] ;

[0156] in, represents the total uncertainty, represents epistemic uncertainty, Represents random uncertainty.

[0157] The two types of uncertainty can be combined in a weighted manner, assigning different weights to different types of uncertainty:

[0158] ;

[0159] in, represents the weight coefficient of epistemic uncertainty, The weight coefficient representing random uncertainty can be dynamically adjusted according to specific medical scenarios and data characteristics. represents epistemic uncertainty, reflecting the lack of model knowledge; Represents random uncertainty, reflecting the noise and variability of the data itself; Represents total uncertainty. For example, in scenarios with sufficient medical resources but lack of experience with specific cases, the weight of epistemic uncertainty can be increased; whereas for grassroots hospitals with limited medical equipment, the weight of random uncertainty can be appropriately increased.

[0160] It's important to note that in medical scenarios, epistemic uncertainty primarily corresponds to the model's inadequate understanding of rare diseases or atypical presentations, while stochastic uncertainty primarily reflects inherent noise in the medical data and individual differences. Distinguishing these two types of uncertainty helps the system adopt different response strategies: for high epistemic uncertainty, the system can offer alternative diagnoses and recommend seeking expert advice; for high stochastic uncertainty, the system can recommend additional tests or try alternative testing methods to reduce data noise.

[0161] Step 3.2: Adaptively adjust the decision threshold based on the quality of the completed data;

[0162] Based on the quality score of the completed data in step 1 and uncertainty quantification model outputs , adaptively adjust the diagnostic decision threshold. Decision threshold The calculation is as follows:

[0163] ;

[0164] in, is the base decision threshold (set to 0.7), is the adjustment coefficient (optimized by the validation set and set to 0.5), Indicates the quality score of the completed data (range is 0 to 1, 1 indicates the highest quality), represents the total uncertainty (composed of epistemic uncertainty and aleatoric uncertainty), Indicates the final diagnostic decision threshold. Low or uncertain When it is higher, the decision threshold The confidence level will increase accordingly, requiring a higher level to make diagnostic decisions.

[0165] Basic decision threshold The threshold can be adjusted based on factors such as disease severity, medical institution level, and medical resources. For example, a tertiary hospital might use a lower baseline threshold (e.g., 0.65) to improve diagnostic efficiency, while a primary care institution might use a higher threshold (e.g., 0.75) to enhance diagnostic safety. Alternatively, the same medical institution could set different baseline thresholds for different departments or disease types, such as a lower threshold for minor dermatological conditions and a higher threshold for severe neurological conditions.

[0166] In addition, risk adjustment factors are introduced according to the severity of different disease types. , further adjust the decision threshold:

[0167] ;

[0168] in, is the final decision threshold, is the decision threshold adjusted based on the quality of the completed data, is the risk adjustment factor, Is the severity of the disease, for high-risk diseases (such as cancer, heart disease, etc.), Take a higher value (such as 0.8-1.0), for low-risk diseases (such as the common cold, etc.), Taking a lower value (e.g. 0.1-0.3) leads to a further increase in the decision threshold and reduces the risk of missed diagnosis. It is usually set to 0.2-0.5, and is preferably 0.3 in this embodiment.

[0169] Risk adjustment not only considers disease severity but also the asymmetric consequences of misdiagnosis. For example, for certain conditions that are not serious but where misdiagnosis could lead to delayed treatment (such as early-stage diabetes), the system will raise the decision threshold. On the other hand, for serious conditions where a false-positive diagnosis could lead to unnecessary invasive testing (such as certain invasive tests), the system will appropriately balance the decision threshold. This risk balancing mechanism is achieved through a comprehensive assessment of the risks of missed and misdiagnosis.

[0170] In actual application, the system also provides a threshold adjustment interface based on the risk preferences and resource constraints of each medical institution, allowing medical institutions to customize basic thresholds according to their own circumstances. and adjustment factor For example, a healthcare network with a well-established referral system can adopt relatively loose threshold settings, while an independently operated healthcare institution may require a more conservative threshold configuration.

[0171] Step 3.3: Implement a progressive reasoning strategy based on completion quality.

[0172] Based on the completion data quality score and the adjusted decision threshold, a progressive reasoning strategy is implemented. The completion data quality is divided into three categories:

[0173] High-quality completion area ( ,in =0.8): Using conventional reasoning strategies, the diagnosis decision is made directly based on the fusion features, and the diagnosis result and its confidence are output;

[0174] Medium quality completion area ( ,in =0.5): adopt a conservative reasoning strategy, increase the decision threshold, generate multiple possible diagnostic hypotheses and attach probability scores, and sort and output the top three most likely diagnostic results;

[0175] Low-quality completion areas ( ): The system refuses to make a definitive diagnosis, but instead provides preliminary analysis results, clearly indicating the items and reasons for which additional examinations are required, and generates a mark for further confirmation.

[0176] in, Indicates the quality score of the completed data, represents the threshold for high-quality completion, Threshold indicating low quality completions.

[0177] Optionally, in different clinical application scenarios, the quality threshold and It can be adjusted according to specific needs. For example, in an emergency environment, it can be appropriately lowered value (such as 0.75) to improve the system response speed; in complex disease screening, it can improve value (such as 0.85) to ensure diagnostic accuracy.

[0178] The system dynamically adjusts its reasoning strategy based on individual patient circumstances. For example, for patients with a clear family history of disease, the system may adopt a more proactive reasoning strategy even with moderate data quality; whereas for complex cases with multiple potential complications, the system may adopt a more conservative reasoning strategy even with high data quality. This personalized reasoning strategy is achieved by comprehensively considering the patient's risk factors, medical history, and current symptoms.

[0179] It should be understood that for critical decisions involving life safety, the system sets minimum data quality requirements (set to 0.3), when the quality of all available data is below this threshold, the system will refuse to make a diagnosis and immediately recommend that the patient be referred to a higher-level medical institution.

[0180] The system generates intelligent referral recommendations, not only identifying the reason for referral but also recommending the most appropriate hospital and department based on the patient's condition and local medical resources. For example, the system considers factors such as distance, specialist capacity, wait times, and accessibility of referral channels to provide personalized referral solutions.

[0181] In addition, the system automatically generates graded medical recommendations based on the results of progressive reasoning: for high-confidence diagnoses, a complete treatment plan is provided; for medium-confidence diagnoses, preliminary treatment recommendations are provided and close follow-up is recommended; for low-confidence diagnoses, only symptom management recommendations are provided and further examinations are recommended.

[0182] Optionally, the system can also generate time-based medical advice, providing phased guidance for dynamically evolving conditions. For example, for patients diagnosed with chronic diseases, the system will generate short-term, mid-term, and long-term health management plans. For suspected cases requiring observation, the system will clearly indicate the time for follow-up visits and symptom changes that require attention.

[0183] This progressive reasoning strategy effectively balances diagnostic accuracy and reliability within the context of limited medical resources, avoiding overconfident misdiagnoses based solely on incomplete data while maximizing the use of existing data to provide valuable medical guidance. Through differentiated decision thresholds and personalized reasoning strategies, the system can adapt to different medical scenarios and patient needs, improving overall medical service quality.

[0184] Step 4: Combined with the adjusted confidence threshold, for different disease types and symptom combinations, a disease modality association matrix and a modality reliability evaluation network are constructed to dynamically adjust the weight of each modality in the inference process;

[0185] It includes the following sub-steps:

[0186] Step 4.1, construct the disease-modality association matrix;

[0187] By analyzing medical literature and large-scale clinical data (more than 1 million medical records), we construct a disease-modality association matrix to quantify the importance and contribution of different medical data modalities to the diagnosis of various diseases. Indicates modality For disease The importance of the diagnosis is calculated as follows:

[0188] ;

[0189] in, Indicates disease The diagnostic process uses modal frequency, Indicates the use of modal Diagnosing diseases The accuracy of Indicates disease The diagnostic process uses modal frequency, Indicates the use of modal Diagnosing diseases The accuracy of is the total number of modes, Indicates modality For disease The importance of diagnosis ranges from 0 to 1. A larger value indicates a greater contribution of the modality to the diagnosis of the disease.

[0190] Optionally, a more complex correlation strength calculation method can be used to consider the time efficiency and cost effectiveness of different modalities for disease diagnosis. For example, the time efficiency coefficient and cost coefficient , corrected correlation strength calculation:

[0191] ;

[0192] in, Indicates the use of modal Diagnosing diseases The average time, represents the corresponding inspection cost, Represents diagnosed diseases in all modalities The longest average time required, Represents diagnosed diseases in all modalities The lowest inspection cost, Indicates modality For disease The importance of diagnosis ranges from 0 to 1. The larger the value, the greater the contribution of the modality to the diagnosis of the disease. It represents the modified modal importance index after considering time efficiency and cost factors, which is the original importance index This correction method prioritizes time-efficient and low-cost modalities while ensuring diagnostic accuracy, and is particularly suitable for environments with limited medical resources.

[0193] It should be noted that, when constructing the symptom-modality association matrix, the elements Indicates modality Symptoms The correlation is calculated based on the statistical co-occurrence frequency and information gain of symptoms and modalities:

[0194] ;

[0195] in, It's a symptom and modal The point mutual information of is modal Symptoms The point mutual information measures the logarithmic ratio of the frequency of two events occurring together relative to their independent occurrences, and is calculated as:

[0196] ;

[0197] in, Symptoms and modal The probability of co-occurrence, Symptoms The marginal probability of occurrence, Indicates modality The marginal probability of occurrence. When the symptoms and modality are independent, the PMI value is 0; when they are positively correlated, the PMI value is positive; when they are negatively correlated, the PMI value is negative.

[0198] Optionally, when processing symptom-modality associations, the system can add prior constraints based on medical expertise. For example, for certain specific symptoms (such as chest pain), a high correlation with cardiac examination modalities such as electrocardiogram and echocardiography can be pre-set, even if this correlation may not be obvious in the initial training data. This integration of prior knowledge can be achieved through the following methods:

[0199] ;

[0200] in, represents the final symptom-modality association strength, It is the prior association strength based on medical knowledge. is the correlation strength calculated from the data, is a coefficient that balances data-driven and prior knowledge. For rare but clinically important symptom-modality associations, A higher value (such as 0.7-0.8) can be set; for common symptoms, a lower value (such as 0.2-0.3) can be set.

[0201] In this way, the system can capture the complex correlation between different modalities and diseases and symptoms, providing a data basis for subsequent dynamic weight optimization.

[0202] Step 4.2, dynamically adjust the modal weight;

[0203] Based on the disease-modality association matrix, symptom-modality association matrix and the current patient's symptom combination, the weight of each modality in the reasoning process is dynamically adjusted. , first calculate the disease probability distribution:

[0204] ;

[0205] in, It is a disease Symptoms under these conditions The probability of It is a disease The prior probability of is obtained through medical statistics. Indicates the diseases, Indicates the patient's symptom combination, 、 、 Respectively represent different symptoms presented by patients (such as fever, cough, headache, etc.), Indicates the total number of symptoms, Indicates a given symptom combination Conditions where the patient has a disease The posterior probability of .

[0206] Prior probability of disease We can further consider the patient's demographic characteristics and regional epidemiological factors. For example, for elderly patients, the prior probability of cardiovascular disease will be increased accordingly; for patients in specific regions, the prior probability of local high-incidence diseases will also be adjusted. This personalized prior probability calculation can be expressed as:

[0207] ;

[0208] in, Indicates the disease after considering age, gender and region factors The conditional probability of Indicates disease The basic prior probability of 、 and are the adjustment functions of age, gender and region on the prior probability of disease, which are obtained through learning from epidemiological data. Indicates the diseases, Indicates the patient's age, Indicates the patient's gender, Represents the region where the patient is located. These factors jointly affect the prior probability distribution of the disease.

[0209] It should be understood that in order to avoid numerical underflow problems caused by continuous multiplication of probabilities, the actual calculation adopts the logarithmic form:

[0210] ;

[0211] in, Indicates the diseases, represents the patient's symptom set, Indicates the Symptoms, Represents a disease under a given set of symptoms The posterior probability of Indicates disease Symptoms under these conditions The probability of Indicates disease The prior probability of Indicates the total number of symptoms, Proportional to means that the left side of the equation is proportional to the right side, that is, the difference between the two is a constant factor (in Bayesian inference, this constant is usually a normalization factor);

[0212] Then, based on the disease probability distribution and the disease-modality association matrix, the initial weight of each modality is calculated:

[0213] ;

[0214] in, is the total number of disease types, is modal The initial weight of Is a given combination of symptoms Conditions and diseases The probability of is the modality represented in the disease-modality association matrix For disease The importance of diagnosis.

[0215] The system can also consider the temporal evolution of symptoms, using different adjustment strategies for acute symptoms and chronic long-term symptoms. For example, for sudden acute symptoms, the system will be more inclined to choose a modality that can quickly provide diagnostic information; while for long-term chronic symptoms, it will give more consideration to modalities that can provide information on the cause and development trend. In this case, symptom adjustment can be expressed as:

[0216] ;

[0217] in, Indicates the adjusted mode The weight of Indicates the mode of the initial calculation The weight of is the symptom adjustment factor, which controls the influence of symptom information on weight adjustment. Symptoms With modal The strength of association, Indicates the total number of symptoms, It is an adjustment function based on the temporal characteristics of symptoms, giving higher weights to acute symptoms (e.g., 1.2-1.5) and baseline weights to chronic symptoms (e.g., 0.8-1.0).

[0218] Step 4.3: Weight normalization and application;

[0219] Normalize the adjusted weights to ensure that the sum of all weights is 1 to obtain the final modal weights:

[0220] ;

[0221] in, Indicates modality The normalized weight of Indicates the adjusted mode The weight of Indicates the adjusted mode The weight of represents the total number of available diagnostic modalities, is the index variable of the mode, from 1 to Iterate over all modals.

[0222] Based on the calculated modality weights, the features or inference results of different modalities are weightedly fused:

[0223] ;

[0224] in, is modal characteristics or inference results, is the result of fusion. is modal The normalized weight of Indicates the total number of available diagnostic modalities.

[0225] In addition, the system dynamically allocates computing resources based on modal weights through the following strategies:

[0226] For weights above the threshold (set to 0.3) mode, enabling all analysis modules for deep feature extraction and fine-grained analysis;

[0227] For weights between (set to 0.1) and The modes between them enable the basic analysis module to perform standard feature extraction;

[0228] For weights lower than The modality is only simplified and key features are extracted for auxiliary verification.

[0229] Computing resource allocation also considers hardware limitations and response time requirements. For example, on edge devices with limited computing power, the system can adopt a more aggressive resource optimization strategy, such as processing only the two most heavily weighted modalities. In time-sensitive scenarios (such as emergency room visits), the modality with the shortest analysis time can be prioritized while processing other modalities in parallel. This flexible resource scheduling strategy is implemented through the following decision function:

[0230] ;

[0231] in, is modal The processing priority, is modal The normalized weight of Is the processing mode The average time required, is the amount of computing resources required. High value mode.

[0232] It should be understood that if the modal weight changes significantly (relative change exceeds 50%), the system will also trigger a re-evaluation mechanism to check the rationality of the weight change through the following steps:

[0233] Check the input data for outliers or noise interference;

[0234] Analyze whether changes in symptom combinations lead to significant changes in disease probability distribution;

[0235] Verify that weight changes are consistent with medical knowledge.

[0236] Optionally, the reassessment mechanism can be validated using historical diagnostic data. The system searches for historical cases with similar symptom patterns and compares the consistency of the current weighting with historical successes. If significant deviations are found, the system flags potential risks and offers additional validation options.

[0237] If an unreasonable weight change is detected, the system will revert to the last stable weight configuration and record the anomaly for subsequent analysis. This reassessment mechanism effectively prevents unreasonable weight allocation caused by noisy data and enhances the robustness of the system.

[0238] The system also implements a progressive weight adjustment strategy to avoid drastic fluctuations in weight. When the newly calculated weight differs significantly from the current weight, the system uses a smooth transition:

[0239] ;

[0240] in, is the smoothing coefficient, usually set to 0.3-0.7, which can be adjusted dynamically according to the magnitude of the weight change; Is the updated modal The weight of Is the current mode The weight of is the newly calculated mode This smoothing strategy ensures that the system is sufficiently responsive to new information while avoiding overreaction due to temporary fluctuations.

[0241] Step 5: Based on dynamically adjusted modality weights, through weighted voting, evidence aggregation, and expert authority assessment, cross-validation of multi-specialty knowledge and collaborative decision-making are achieved to form the final diagnostic recommendation;

[0242] It includes the following sub-steps:

[0243] Step 5.1: Construct a specialized knowledge representation model;

[0244] Dedicated knowledge representation models are built for different medical specialties, such as internal medicine, surgery, and radiology. Each model contains the specific knowledge and diagnostic ideas for that specialty. Knowledge representation uses a combination of knowledge graphs and probabilistic graphical models to simultaneously express deterministic knowledge and uncertain reasoning.

[0245] It should be noted that each specialized knowledge model It contains three core components:

[0246] Specialty-specific conceptual system , contains the commonly used medical terms, disease classifications and diagnostic criteria of the specialty, organized by ontology structure;

[0247] Inference rule set , which includes the diagnostic ideas and decision rules unique to this specialty, and is expressed using rules that include confidence levels:

[0248] ;

[0249] Indicates when the condition When satisfied, Confidence inference Established; among them, Indicates the rule identifier, 、 、 They represent the first prerequisite (such as a specific symptom, such as fever), the second prerequisite (such as a specific sign, such as pharyngeal congestion), and the nth prerequisite (such as a specific test result, such as an elevated white blood cell count). These conditions can be symptoms, signs, test results, laboratory indicators, or other medical findings. represents the total number of preconditions, Indicates the conclusion of the rule (such as diagnosis or treatment recommendations for diseases such as upper respiratory tract infection), Indicates the confidence of the rule, reflecting the reliability of the inference rule;

[0250] Weight of evidence matrix , define the importance weight of different evidence types in the diagnosis of this specialty, elements Express evidence For disease support strength.

[0251] The specialist knowledge model can optionally introduce the time evolution dimension to capture the dynamic characteristics of disease development. For example, for the management of chronic diseases in internal medicine, a disease progression network can be constructed. , representing the characteristic changes and transition probabilities of different stages of a disease from early to late stages. This temporal knowledge representation is particularly suitable for chronic diseases that require long-term follow-up, such as diabetes and hypertension, as well as diseases with a clear course of development, such as tumors. The disease progression network can be expressed as:

[0252] ;

[0253] in, and Representing disease stages a and b, which are two consecutive stages of the disease. From the stage Progress to stage The probability of is the mean progression time.

[0254] It should be understood that these specialized knowledge models are built based on specialized guidelines, textbooks, and expert knowledge, and are fine-tuned and validated using large-scale medical data. The knowledge models are modular in design, allowing each component to be independently updated, facilitating iterative upgrades based on the latest medical research findings.

[0255] Optionally, knowledge model updates can be performed semi-automatically. The system regularly analyzes updates to the medical literature and clinical practice guidelines, automatically extracts new findings and the latest recommendations, generates candidate knowledge updates, and integrates them into the existing knowledge model after expert review. This continuous learning mechanism ensures that the knowledge model always reflects the latest medical advances.

[0256] The system also establishes concept mappings between specialties to address the problem of different specialties using different terminology to describe the same or related medical concepts, thereby improving knowledge sharing and collaboration between specialties. Concept mapping connects the conceptual systems of different specialties through three types of relationships: equivalence, inclusion, and correlation.

[0257] Concept mapping can be further refined into context-sensitive mapping, which takes into account the semantic differences of terms in different diagnostic contexts. For example, the severity of cardiac insufficiency may be classified differently in cardiology and geriatrics. This context-sensitive mapping is represented by the following triple:

[0258] ;

[0259] in, represents a set of context-related concept mappings, The medical concept representing the first specialty (e.g., heart failure in cardiology), Represents a related medical concept in the second specialty (such as cardiac insufficiency in geriatrics), is the application context (such as a specific disease scenario), It is the similarity or mapping strength in this context, indicating the degree of correspondence between two concepts in a specific context. The value is usually between 0 and 1. The larger the value, the stronger the mapping relationship.

[0260] Step 5.2: Implement specialist cross-validation;

[0261] Based on the multimodal medical data processed in steps 1-4, each specialty knowledge model is activated for independent reasoning to generate specialty-specific diagnostic hypotheses and confidence levels.

[0262] Specialist The diagnostic results are expressed as:

[0263] ;

[0264] in, Indicates the specialty number, Represents the total number of diagnostic hypotheses proposed by this specialty. 、 、 Respectively represent specialist The first, second, and K1 diagnostic hypotheses proposed, 、 、 Respectively represent specialist The confidence level of the first, second, and K1 diagnostic hypotheses. Is a variable that represents a specific specialty The number of diagnostic hypotheses proposed may vary between specialties.

[0265] Optionally, you can enable the intra-specialty grouping reasoning mechanism to simulate the diagnostic thinking of doctors of different qualifications within the specialty. For example, the internal medicine specialty model can be divided into three reasoning groups: junior, intermediate, and senior, which respectively simulate the diagnostic process of residents, attending physicians, and chief physicians, generating multi-level diagnostic results:

[0266] , ,

[0267] in, represents the set of diagnostic results of simulated resident doctors (junior doctors), represents the set of diagnostic results of the simulated attending physician (intermediate doctor), Represents a set of diagnostic results from a simulated chief physician (senior doctor). Each set contains possible diagnostic hypotheses and their confidence levels that doctors with a corresponding level of experience might propose.

[0268] This hierarchical reasoning not only enhances diagnostic diversity, but also identifies different diagnostic clues that different levels of experience may focus on, improving the system's ability to handle complex cases.

[0269] It should be understood that, then, specialist cross-validation is performed to check the consistency and complementarity between the diagnostic results of different specialists. Cross-validation includes:

[0270] Calculation of diagnostic concordance matrix ,element Indicates a specialist and specialist The degree of consistency of the diagnostic results is calculated as follows:

[0271] ;

[0272] in, Indicates a specialist and specialist The degree of consistency of diagnostic results, Indicates a specialist and specialist The intersection of the diagnostic results, represents their union, and Respectively represent specialist and specialist Diagnosis confidence level; Indicates taking a specialist degree and specialist Diagnosis The smaller value of the confidence level; Indicates taking a specialist degree and specialist Diagnosis The larger value of the confidence level; Represents a sum operation;

[0273] Optionally, when calculating inter-specialty consistency, the semantic similarity of the diagnosis in the medical ontology can be considered, rather than just strict term matching. For example, type 2 diabetes and adult-onset diabetes, although expressed differently, refer to the same disease. This semantically aware consistency calculation can be expressed as:

[0274] ;

[0275] in, Indicates a specialist and specialist Semantic consistency matrix of diagnosis results, which considers the semantic similarity of diagnosis terms instead of just strict term matching, It is the semantic similarity between diagnoses calculated based on medical ontology. Indicates a specialist One of the diagnosis results, Indicates a specialist One of the diagnosis results, Indicates a specialist The set of all diagnostic results, Indicates a specialist The set of all diagnostic results, Indicates a specialist Diagnosis The confidence level, Indicates a specialist Diagnosis confidence level.

[0276] Identify diagnostic conflicts between specialties when the difference in confidence between two specialties for the same diagnosis exceeds a threshold (set to 0.4), it is determined that there is a conflict, and the system will resolve the conflict based on the expertise of the specialist and the reliability of the relevant data;

[0277] Conflict resolution strategies are dynamically selected based on the nature of the conflict. For example, in diagnostic conflicts (i.e., different diagnoses for the same symptom), the system tends to favor the opinion of a specialist with greater expertise in that symptom. In contrast, in severity assessment conflicts (i.e., different severity assessments for the same disease), the system considers the opinions of various specialists and favors a more conservative assessment to ensure patient safety.

[0278] Discover the complementary information of each specialty's diagnosis and identify diagnostic hypotheses that are only proposed by a single specialty but have high confidence as potential diagnostic clues that are easily overlooked.

[0279] Optionally, the system also implements a proactive diagnostic blind spot mining mechanism, automatically searching medical literature for similar cases for specific cases and identifying commonly overlooked related diseases. For example, for patients presenting with abdominal pain, the system will proactively check for evidence of cardiovascular disease in addition to common digestive system diseases, as some myocardial infarctions can present as upper abdominal pain, yet are often missed clinically.

[0280] In addition, the system also constructs a diagnostic path diagram to record the reasoning path from symptoms to diagnosis in each specialty, so as to facilitate logic verification and explanation generation. The diagnostic path is represented as a directed graph ,in, Represents the diagnostic path graph, vertex Includes symptoms, test results, and diagnosis, Represents reasoning steps and evidence support relationships.

[0281] Diagnostic pathway diagrams can also label reasoning types, distinguishing between rule-based deterministic reasoning and statistical-based probabilistic reasoning, enhancing the transparency of interpretation. Deterministic reasoning edges are accompanied by clear references to medical rules.

[0282] Step 5.3: Realize collaborative decision-making based on specialized knowledge;

[0283] Based on the results of specialist cross-validation, the final collaborative diagnosis results are formed through weighted voting, evidence aggregation and specialist authority evaluation.

[0284] First, calculate the authority score of each specialty in the current case :

[0285] ;

[0286] in, It is a specialist The authoritative rating of It is a specialist The historical accuracy of It is a specialist Relevance to the current symptom combination, It is a specialist Internal consistency of diagnostic results, 、 and is the weight coefficient, where represents the weight of historical accuracy, represents the weight of symptom relevance, Represents the weight of internal consistency, satisfying , in this embodiment, it is set to , , .

[0287] Optionally, in scenarios where there are significant differences in medical resources, the authority score calculation can take into account the differences in the strength of medical institutions. For example, the imaging department of a grassroots hospital may have different equipment and expert levels than that of a tertiary hospital, so an institution-level adjustment factor is introduced into the calculation:

[0288] ;

[0289] in, It is a specialist The authoritative rating of It is a specialist The historical accuracy of It is a specialist Relevance to the current symptom combination, It is a specialist Internal consistency of diagnostic results, It is the institution's grade for the specialty. 、 、 is the weight coefficient, represents the weight of historical accuracy, represents the weight of symptom relevance, represents the weight of internal consistency, is the weight coefficient of the institutional level adjustment factor.

[0290] in, It is the institution's grade for the specialty. is the corresponding weight coefficient. This institution-sensitive authority score ensures that the system can make reasonable specialty weight allocations in different medical environments.

[0291] It should be noted that the historical accuracy Based on specialist Evaluation of diagnostic performance in previous similar cases; symptom relevance Calculated as current symptoms and specialist Focus on symptom overlap; internal consistency Measuring Specialties The internal logical consistency of the diagnosis was assessed by examining the relationships between the diagnosis, evidence, and inference rules.

[0292] The historical accuracy evaluation can introduce a time decay factor to make the system focus more on recent performance:

[0293] ;

[0294] in, It is a specialist The historical accuracy of It is a specialist In historical cases The accuracy in It's a case timestamp, is the current time, is the time decay coefficient, is the total number of historical cases, is the base of natural logarithms, is the index of the historical case, indicating the This time-aware accuracy assessment can reflect the dynamic changes in specialist capabilities, especially in areas where medical knowledge and technology are rapidly evolving.

[0295] Then, using a weighted voting mechanism, each diagnostic hypothesis is calculated Overall rating:

[0296] ;

[0297] in, is the total number of specialties, It is a specialist The basic weight of the drug (reflecting its importance in the medical system), It is a specialist The authoritative rating of It is a specialist Diagnosis The confidence level, Indicates diagnostic hypothesis The comprehensive rating of . The diagnostic hypothesis of the position is selected as the candidate diagnosis, where It is usually set to 3-5 and adjusted dynamically according to the complexity of the disease.

[0298] Alternatively, in complex multi-system disease scenarios, the system can adopt a hierarchical voting strategy, first forming a consensus within the relevant specialist groups, and then integrating between the groups. For example, for symptoms that may involve both the cardiovascular and nervous systems, a preliminary consensus is first formed within the cardiovascular specialist groups such as cardiology and cardiac surgery, while another consensus is simultaneously formed within the nervous system specialist groups such as neurology and neurosurgery, and finally integrated between the two specialist groups. This hierarchical voting strategy is implemented using the following formula:

[0299] ;

[0300] Among them, A is the number of specialist groups, It is a specialist group The weight of It is a specialist In group The relative weight within It is a specialist The authoritative rating of It is a specialist Diagnosis The confidence level, Indicates a specialist Belong to the specialist group , It is a diagnosis 's comprehensive rating.

[0301] Finally, evidence is aggregated for diagnostic hypotheses with higher scores, integrating all evidence supporting the diagnosis to form a structured diagnostic report, including: the main diagnostic conclusion and its confidence level, supporting evidence and its source (distinguishing between original observational evidence and systematic inference evidence), diagnostic opinions of various specialties and their weights, potential differential diagnoses and their likelihood, recommended further examination items and their purpose.

[0302] Diagnostic reports can also generate differentiated content based on the needs of different users. For example, for physicians, the report may include detailed technical information and medical literature citations; while for patients, it may provide easy-to-understand explanations and lifestyle recommendations. This audience-sensitive report generation is achieved through a natural language generation model that is specifically trained for different audience types.

[0303] It should be understood that the system also provides diagnostic explanations, visually displaying the reasoning path from symptoms to diagnosis. This enhances transparency in medical decision-making and facilitates physicians' evaluation and verification of the system's diagnostic recommendations. This explanation includes quantification of the contribution of key evidence, comparisons of different specialist perspectives, and relevant references from clinical guidelines.

[0304] Optionally, the system's explanation function supports interactive exploration, allowing users to evaluate diagnostic changes under different conditions through what-if analysis. For example, a doctor can ask how the diagnosis would change if the patient's blood pressure dropped by 20 mmHg. The system will recalculate and display the change in the diagnosis based on the existing model, helping doctors understand the weight of each factor influencing the diagnosis.

[0305] In addition, the system integrates an active learning mechanism to record the user's (doctor's) acceptance, modification or rejection of the system's diagnostic recommendations, and adjusts the specialty weights and reasoning rules based on this feedback, continuously improving the accuracy and practicality of collaborative decision-making based on multi-specialty knowledge.

[0306] Active learning is not limited to weight adjustment but also includes rule reconstruction and new knowledge discovery. The system analyzes patterns in physician modification behavior, identifies possible new rules or knowledge gaps, and generates recommendations for knowledge updates. For example, if a physician frequently modifies the system's diagnosis for a certain symptom combination, the system automatically extracts these cases and analyzes commonalities to identify potential new rule candidates. These are then added to the knowledge base after expert confirmation. This swarm intelligence mechanism enables the system to continuously learn from clinical practice, forming a knowledge evolution cycle of experience-verification-refinement-application.

[0307] Application examples:

[0308] This application example demonstrates the practical application of this implementation method in a grassroots hospital in a mountainous area. This county-level hospital has limited equipment, including basic X-ray equipment and routine laboratory testing capabilities. It lacks advanced equipment such as CT and MRI, as well as specialist physicians. Patient Zhang, a 65-year-old male, complained of recurrent chest pain for three days, accompanied by difficulty breathing. The local hospital could only provide basic X-rays, blood tests, electrocardiograms, and basic vital signs tests, and was unable to perform more advanced tests such as echocardiography and coronary CT.

[0309] System implementation process:

[0310] Modality Completion and Quality Assessment: The system first collects the patient's basic vital signs (blood pressure 165 / 95 mmHg, heart rate 98 beats / min, respiratory rate 23 breaths / min, temperature 36.8°C), a description of basic symptoms, X-ray results (showing enlarged cardiac shadow), electrocardiogram (showing ST segment changes), and routine blood test results (white blood cell count 10.5×10^9 / L). The system identifies coronary CT and echocardiography as key missing modalities and initiates the modality completion network for prediction.

[0311] Based on the available data, the system generated predictions for echocardiography and coronary CT scans using a hierarchical modality completion network. For echocardiography, the system predicted a left ventricular ejection fraction of approximately 45% ± 5%, suggesting the presence of moderate heart failure. For coronary CT scans, the system predicted a possible stenosis of the anterior descending artery of approximately 70% ± 15%.

[0312] The system performed a quality assessment on the completed results, calculating the following scores: the echocardiogram data quality score was 0.72 (moderate quality), and the coronary CT prediction quality score was 0.58 (moderately low). The quality assessment function comprehensively considers the rationality of the content (consistency with the patient's age, gender, and symptoms), consistency with observed data (consistency with electrocardiogram and X-ray results), and adherence to medical principles.

[0313] Multimodal medical data fusion and self-calibration: The system mapped the original acquired data and the supplemented data into a shared feature space and detected a slight discrepancy between ST segment changes in the electrocardiogram (ECG) and predicted coronary artery stenosis. The ECG indicated possible inferior myocardial ischemia, while the predicted coronary CT scan indicated stenosis of the anterior descending artery (usually corresponding to anterior wall ischemia). The system calculated an intermodality consistency score of 0.65, below the threshold of 0.8, triggering the self-calibration mechanism.

[0314] After self-calibration, the system adjusted the weights of each modality: the weight of the ECG was increased from 0.25 to 0.35 (due to its high reliability as raw data), and the weight of the predicted coronary CT was decreased from 0.20 to 0.12 (due to its low quality rating as supplemental data). The feature fusion results after self-calibration favored the ECG information, revising the estimated location of coronary artery lesions.

[0315] Adaptive decision threshold adjustment: The system calculates a decision threshold based on the quality of the supplemented data and the uncertainty assessment results. For this patient, the baseline decision threshold was 0.7. Taking into account the severity of the heart disease (severity coefficient s = 0.9) and the moderate quality of the supplemented data, the system calculated a final decision threshold of 0.85.

[0316] Since the overall quality score of the completed data is at a medium level (Q=0.65, between θ_Qlow and θ_Qhigh), the system adopts a conservative reasoning strategy to generate multiple diagnostic hypotheses and their confidence levels:

[0317] Acute coronary heart attack: confidence level 0.87 (above the adjusted decision threshold of 0.85)

[0318] Cardiac insufficiency: confidence level 0.79 (below the adjusted decision threshold)

[0319] Aortic dissection: confidence level 0.45 (significantly below the threshold)

[0320] Dynamic modal weight optimization: The system calculates disease probability distribution based on the patient's symptom combination (chest pain, dyspnea, elevated blood pressure), significantly improving the probability of cardiovascular disease. Based on the disease-modality association matrix, the system dynamically adjusts the weight of each modality: the weight of the electrocardiogram is increased to 0.4, the weight of the X-ray is 0.25, the weight of the blood routine is reduced to 0.1, the weight of the supplementary cardiac ultrasound data is 0.15, and the weight of the supplementary coronary CT is 0.1.

[0321] This dynamic weight adjustment enables the system to pay more attention to modal data such as electrocardiograms and X-rays that contribute more to the diagnosis of cardiovascular diseases during the reasoning process, while correspondingly reducing the weights of less relevant modalities such as blood routine tests, thereby improving the targeted nature of the diagnosis.

[0322] Multi-specialty knowledge collaboration: The system activated three specialized knowledge models for internal medicine (cardiology), imaging, and emergency medicine for cross-validation. The diagnosis results of each specialty are as follows:

[0323] Cardiology model: The primary diagnosis is acute coronary heart disease (confidence 0.90), followed by heart failure (confidence 0.82)

[0324] Imaging model: The main diagnosis is cardiac enlargement (confidence level 0.95), supporting the diagnosis of heart failure (confidence level 0.85)

[0325] Emergency Department Model: The primary diagnosis is acute coronary heart disease (confidence 0.88), and it also suggests the need to be vigilant about the possibility of aortic dissection (confidence 0.42).

[0326] The system calculated the specialist authority scores: 0.85 for cardiology, 0.75 for imaging, and 0.80 for emergency medicine. Using a weighted voting mechanism, the final diagnosis was: acute coronary heart disease (overall confidence 0.89) with associated heart failure (overall confidence 0.78). The system also indicated the need to rule out the possibility of aortic dissection and recommended prompt referral to a higher-level hospital for further confirmation.

[0327] Technical effect verification:

[0328] This example demonstrates the effectiveness of the system in an environment with limited medical resources, mainly in the following two aspects:

[0329] Improved diagnostic accuracy:

[0330] As shown in Table 1, the system's diagnostic accuracy evaluation of 100 similar cases in the county-level hospital showed that under the condition of limited medical resources, the system's final diagnosis accuracy reached 90.5%, while the traditional fixed model that does not use modality completion and dynamic weight optimization had an accuracy of only 72.3%, an improvement of 18.2 percentage points.

[0331] In 50 difficult and complex cases, the system achieved a diagnostic accuracy of 86.4% through multidisciplinary collaboration, compared to 65.8% for a single-specialty model, an improvement of 20.6 percentage points. In particular, the system significantly enhanced its ability to identify multi-system diseases, identifying 24.5% more potential disease risks than a single-specialty model.

[0332] Table 1: Comparison of system diagnostic accuracy under different case types

[0333]

[0334] Optimizing the efficiency of medical resource utilization:

[0335] As shown in Table 2, the system's progressive reasoning strategy and adaptive referral recommendation mechanism significantly improved the efficiency of medical resource utilization. In county-level hospitals where this system was deployed, the unnecessary referral rate dropped from 65.3% to 30.5%, a decrease of 34.8 percentage points. Simultaneously, the missed diagnosis rate (where patients should have been referred but were not) dropped from 12.8% to 2.6%, a decrease of 10.2 percentage points.

[0336] The patient experience process has also been streamlined, with the average time from initial consultation to initial diagnosis shortened by 54.6%, from 3.2 hours to 1.45 hours. Doctors' diagnostic efficiency has increased by 48.7%, with the average number of cases handled per doctor per day increasing from 15 to 22.

[0337] Table 2: Comparison of medical resource utilization efficiency before and after system deployment

[0338]

[0339] Overall, this application example fully demonstrates the value of this implementation in resource-constrained medical environments. Through the integrated application of technologies such as modality completion and quality assessment, multimodal fusion and self-calibration, adaptive decision threshold adjustment, dynamic modality weight optimization, and multi-specialty knowledge collaboration, the system successfully addresses the limited diagnostic capabilities of primary healthcare institutions due to incomplete data, improves diagnostic accuracy, optimizes the efficiency of medical resource utilization, and provides intelligent medical guidance services in resource-constrained areas, approaching the level of tertiary hospitals.

[0340] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A method for constructing a reasoning agent for intelligent medical guidance, characterized in that: The following steps are involved: Construct a hierarchical modality completion network and modality correlation knowledge graph to achieve intelligent completion and quality assessment of missing medical data modalities in a resource-constrained environment, including: Construct a modality completion subnetwork with an encoder-decoder structure to complete different types of missing medical data; By analyzing large-scale medical data sets, we quantify the mutual predictive power and dependencies between different medical data modalities and construct a modality correlation knowledge graph. Perform quality assessment on the generated modal data and quantify its reliability and uncertainty bounds through the quality assessment function: ； in, For the generated data, is the observed data, Evaluate the rationality of the content, Assess consistency with observed data, Assess the degree of compliance with medical laws, 、 、 They represent the weight coefficients for evaluating the rationality of the content, the consistency with observed data, and the degree of compliance with medical laws, respectively, and meet the , represents the quality assessment function of the generated data under the condition of the observed data; Based on the completed medical data, contrastive learning is used to build a medical feature cross-modal mapping network, mapping different modal medical data into a shared feature space to achieve multimodal medical data fusion and self-calibration; For the fused multimodal data, an uncertainty quantification model is constructed based on the uncertainty of the supplemented data, and the diagnostic confidence threshold is automatically adjusted to achieve adaptive decision threshold adjustment; Combined with the adjusted confidence threshold, a disease modality association matrix and a modality reliability assessment network are constructed for different disease types and symptom combinations, dynamically adjusting the weight of each modality in the inference process; Based on dynamically adjusted modal weights, through weighted voting, evidence aggregation and expert authority assessment, cross-validation and collaborative decision-making of multi-specialty knowledge are achieved to form final diagnostic recommendations.

2. The method for constructing a reasoning agent for intelligent medical guidance according to claim 1, characterized in that: Based on contrastive learning, a medical feature cross-modal mapping network is constructed to map medical data of different modalities into a shared feature space, thus achieving multimodal medical data fusion and self-calibration. Construct a medical feature cross-modal mapping network to map medical data of different modalities into a shared feature space; By optimizing the network parameters through contrast loss function and medical feature fidelity loss, the mapping results of different modal data from the same medical case are similar. Construct a modal reliability assessment network to automatically assess the reliability of each modal data by analyzing historical accuracy and current information completeness; Implement inter-modal consistency scoring and self-calibration mechanism. When the inter-modal consistency score is detected to be lower than the threshold, the self-calibration mechanism is triggered to adjust the weight of each modality or request additional information.

3. The method for constructing a reasoning agent for intelligent medical guidance according to claim 1, characterized in that: The steps for automatically adjusting the diagnostic confidence threshold based on the uncertainty of the completed data include: Construct an uncertainty quantification model based on the Monte Carlo random dropout method to quantify epistemic uncertainty and random uncertainty; Adaptively adjust the diagnostic decision threshold based on the completed data quality score and the output of the uncertainty quantification model; Implement a progressive reasoning strategy based on completion quality, divide the completion data into three categories according to their quality: high quality, medium quality, and low quality, and adopt different reasoning strategies.

4. The method for constructing a reasoning agent for intelligent medical guidance according to claim 1, characterized in that: The steps to achieve dynamic modal weight optimization for different disease types and symptom combinations include: Construct disease modality association matrices and symptom modality association matrices to quantify the importance of different medical data modalities for the diagnosis of various diseases; Based on the disease probability distribution and association matrix, the weight of each modality in the reasoning process is dynamically adjusted; Normalize the adjusted weights to ensure that the sum of all weights is 1, and apply them to multimodal feature fusion; Computing resources are dynamically allocated according to modal weights, high-weight modes are deeply analyzed, and low-weight modes are simplified.

5. The method for constructing a reasoning agent for intelligent medical guidance according to claim 1, characterized in that: The steps to achieve multi-specialty knowledge collaboration through weighted voting, evidence aggregation, and specialist authority assessment include: Build dedicated knowledge representation models for different medical specialties, including specialty-specific concept systems, reasoning rule sets, and evidence weight matrices; Based on multimodal medical data, each specialty knowledge model is activated to perform independent reasoning and generate specialty-specific diagnostic hypotheses and confidence levels; Conduct specialist cross-validation, calculate the diagnostic consistency matrix, and identify diagnostic conflicts and complementary information between specialties; The authoritative score of each specialty in the current case is calculated, and the final collaborative diagnosis result is formed through a weighted voting mechanism.

6. The method for constructing a reasoning agent for intelligent medical guidance according to claim 3, characterized in that: The progressive reasoning strategy is specifically as follows: For high-quality completed regions, conventional reasoning strategies are used to make diagnostic decisions directly based on fused features; For medium-quality completed regions, a conservative reasoning strategy is adopted to increase the decision threshold and generate multiple possible diagnostic hypotheses simultaneously; For low-quality completed areas, the system refuses to make a definitive diagnosis and instead provides preliminary analysis results and recommendations for additional examinations; For critical decisions involving life safety, set minimum data quality requirements. When the quality of all available data is below this threshold, it is recommended that the patient be referred to a higher-level medical institution.

7. The method for constructing a reasoning agent for intelligent medical guidance according to claim 5, characterized in that: The specialist knowledge model includes: Specialty-specific conceptual frameworks, including commonly used medical terminology, disease classifications, and diagnostic criteria; A set of inference rules, including unique diagnostic ideas and decision rules; The weight of evidence matrix defines the importance of different types of evidence in the diagnosis of this specialty; The concept mapping relationship between specialties connects the concept systems of different specialties through equivalence relationships, inclusion relationships and correlation relationships.

8. The method for constructing a reasoning agent for intelligent medical guidance according to claim 1, characterized in that: It also includes steps for automatically generating graded medical recommendations: For high-confidence diagnoses, a complete treatment plan is provided; For moderate-confidence diagnoses, preliminary treatment recommendations are provided and close follow-up is recommended; For low-confidence diagnoses, provide symptom management advice and recommend further testing; Intelligently adjust the level of detail and urgency of treatment recommendations based on different disease types and severity.

9. A system for constructing a reasoning agent for intelligent medical guidance, configured to execute a method for constructing a reasoning agent for intelligent medical guidance as described in any one of claims 1 to 8, characterized in that: include: A hierarchical modality completion module, which is used to intelligently complete and assess the quality of missing medical data modalities in resource-constrained environments; Medical feature cross-modal mapping module, used to map different modal medical data into a shared feature space, achieving multimodal medical data fusion and self-calibration; Adaptive decision threshold adjustment module, used to automatically adjust the diagnosis confidence threshold according to the uncertainty of the completed data; Dynamic modality weight optimization module, which is used to dynamically adjust the weight of each modality in the inference process according to different disease types and symptom combinations; The multi-disciplinary knowledge collaboration module is used to achieve cross-validation and collaborative decision-making of multi-disciplinary knowledge through weighted voting, evidence aggregation and expert authority evaluation.

Images