Knowledge graph-based remote health diagnosis result generation method and device, and medium

Abstract

The application discloses a knowledge graph-based remote health diagnosis result generation method and device and a medium, and relates to the technical field of computers. The application standardizes home data uploaded by manufacturers through a hospital server, discriminates abnormal labels and corrects soft labels for hospital diagnosis data, realizes heterogeneous data fusion diagnosis by using a double-channel network, and finally feeds back diagnosis results after medical school examination and confidence calibration to manufacturers, effectively filling the mode blank that hospitals cannot open data and diagnosis services to home monitoring manufacturers, while standardizing home data and efficiently utilizing noisy hospital data, realizing reliable joint diagnosis and standardized result output, and directly supporting manufacturers to generate compliance test reports.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, device and medium for generating remote health diagnosis results based on knowledge graphs. Background Technology

[0002] Current medical data exchange scenarios are limited to hospital-to-hospital interactions. The closest existing technology is the inter-hospital case information exchange method, which only achieves data sharing between medical institutions through anonymized summaries, target hospital screening, and privacy-secure channels. It completely fails to consider opening up data services to non-medical entities such as home monitoring device manufacturers. It cannot provide manufacturers with clean datasets that have undergone misdiagnosis correction and soft label weighting and can be used for model training, nor can it receive and process home time-series monitoring data uploaded by manufacturers. Furthermore, it cannot return standardized diagnostic conclusions to manufacturers that can be directly used to generate test reports. It is not suitable for the business model of hospitals providing data support and diagnostic feedback to external home monitoring manufacturers. At the same time, home monitoring data has a large number of non-standard and abnormal issues, hospital diagnosis and treatment data has a high proportion of misdiagnosis label noise, hospital structured data and home time-series data are difficult to integrate efficiently, and diagnostic results lack medical logic constraints and confidence calibration. This further prevents hospitals from providing safe, compliant, and implementable medical diagnostic support services to home monitoring manufacturers.

[0003] Furthermore, wearable devices and home self-testing instruments generate massive amounts of physiological monitoring time-series data. However, the home data collection scenario lacks professional standards and constraints, commonly resulting in various anomalies such as inappropriate monitoring timing, improper operation, device data drift, outliers in the time series, arbitrary data entry, random noise, and conflicts with physiological logic. If home monitoring device manufacturers generate test reports based on the generated physiological monitoring time-series data, the results may be biased. If hospital medical data is used to assist in analysis, hospitals do not open their diagnostic service interfaces to the public, and the historical medical data accumulated by hospitals already contains a certain proportion of misdiagnosed samples. If these misdiagnosed samples are used directly for health prediction without identifying and correcting suspicious samples, the final diagnostic results may contradict medical logic and cannot be used directly. Summary of the Invention

[0004] This application provides a method, device, and medium for generating remote health diagnosis results based on knowledge graphs, aiming to overcome the shortcomings of existing medical data interaction, which only supports sharing between hospitals and cannot enable hospitals to open up data and diagnostic feedback services to home monitoring manufacturers. At the same time, home monitoring data cannot be directly used by manufacturers to generate test reports.

[0005] This application provides a method for generating remote health diagnosis results based on knowledge graphs, including: In response to receiving out-of-hospital physiological time-series data uploaded by the home monitoring device manufacturer, the system performs anomaly identification on the out-of-hospital physiological time-series data according to a preset general medical time-series compliance rule library, sets anomaly type labels for abnormal out-of-hospital physiological time-series data, and obtains out-of-hospital classification data. The outpatient classification data is standardized, time-series noise is suppressed, and outliers are removed to generate outpatient feature vectors with aligned feature dimensions and standardized distribution. Based on the outpatient classification data, the corresponding structured diagnosis and treatment history data in the hospital database is read. The structured diagnosis and treatment history data is used to identify misdiagnosed samples and set abnormal labels through medical knowledge graph rule verification, density cluster discrete point detection, and self-supervised feature reconstruction error verification. The suspiciousness score of the misdiagnosed samples is determined and the abnormal labels are calibrated to obtain the historical correction dataset. A dual-channel feature adaptive fusion diagnostic network is constructed, with the historical corrected dataset and the outpatient feature vector as dual inputs. A noise-robust loss function composed of a focusing loss function and label softening is adopted. Two-stage step-by-step training and progressive optimization of sample weights are performed to obtain a converged medical auxiliary diagnostic model. The out-of-hospital feature vector is input into the medical auxiliary diagnostic model to perform forward inference and classification prediction, and outputs multi-class preliminary diagnostic confidence results; Based on a preset medical logic post-verification rule base, the multi-category preliminary diagnostic confidence results are subjected to logic conflict detection and correction to obtain a medical auxiliary diagnostic conclusion that meets medical logic constraints and is fed back to the home monitoring device manufacturer.

[0006] Furthermore, the outpatient physiological time-series data is anomaly identified based on a preset general medical time-series compliance rule library. Anomaly type labels are set for the abnormal outpatient physiological time-series data to obtain outpatient classification data, including: A general medical time series compliance rule library is pre-built on the hospital server. In the general medical time series compliance rule library, standard monitoring constraints, physiological normal range boundaries, resting / fasting / postprandial physiological state adaptation specifications, and reasonable ranges of time series physiological changes are defined for various monitoring items. The hospital server receives out-of-hospital physiological time-series data uploaded by the home monitoring device manufacturer; Based on the general medical time series compliance rule library, the outpatient physiological time series data are anomaly identified from the dimensions of compliance of monitoring timing, compliance of operating procedures, rationality of numerical physiological intervals, rationality of short-term time series mutations, and rationality of equipment acquisition drift. Based on the screening results, abnormal out-of-hospital physiological time-series data with abnormalities are labeled with abnormality type tags; Each sample in the outpatient physiological time series data is bound to a corresponding abnormality type label and an abnormality level score to obtain outpatient classification data with abnormality type labels.

[0007] Furthermore, the step of standardizing, suppressing temporal noise, and removing outliers from the outpatient classification data to generate outpatient feature vectors with aligned feature dimensions and standardized distribution includes: In the hospital server, a multiple linear regression mapping model is pre-constructed based on the home non-standard original monitoring feature vector, feature mapping weight matrix, feature mapping weight matrix and the transformed equivalent standard benchmark feature vector; The multivariate linear regression mapping model is trained by using normal standard medical samples and non-standard home samples to form training pairs, establishing a regression mapping relationship from the non-standard monitoring feature space to the standard benchmark physiological feature space, and obtaining a cross-time series benchmark mapping model. Substitute the non-standard home samples labeled with the abnormality type in the outpatient classification data into the cross-time series benchmark mapping model to perform equivalent standard feature conversion; A multi-mode joint filtering strategy is formed by combining median filtering, Kalman filtering, and sliding window mean filtering to remove single-point impulse outliers, smooth temporal drift, and suppress noise for each sample in the outpatient classification data. For each sample data, normalization and time series length alignment are performed to unify the feature dimensions and input format, thereby obtaining an out-of-hospital feature vector with aligned feature dimensions and standardized distribution.

[0008] Furthermore, based on the outpatient classification data, corresponding structured medical history data is read from the hospital database. The structured medical history data is then used to identify misdiagnosed samples and assign anomaly labels through medical knowledge graph rule verification, density clustering discrete point detection, and self-supervised feature reconstruction error verification. A suspicion score is determined for each misdiagnosed sample, and the anomaly labels are calibrated to obtain a historical correction dataset, including: A medical knowledge graph is built in the hospital server. The nodes of the medical knowledge graph include disease types, physiological test indicators, age groups, genders, complications, and clinical diagnostic threshold rules. The edges of the medical knowledge graph are associated with the fixed relationships between indicators and disease types, the constraints between population and disease incidence, and the mutually exclusive logical relationships between indicator combinations. Read the structured medical history data corresponding to the outpatient classification data in the hospital database, extract patient attribute features, test index features, and original diagnostic label features, and construct a high-dimensional sample feature set; Based on the medical knowledge graph, the high-dimensional sample feature set is subjected to rule logic verification, and samples that violate the clinical diagnosis threshold, population adaptation, and indicator combination logic are marked as first-level suspected misdiagnosis samples. The high-dimensional sample feature set is divided into clusters using a density clustering algorithm. The neighborhood radius and minimum number of cluster samples are set to identify outliers that are far from the main distribution clusters. These outliers are marked as secondary suspected misdiagnosed samples. A multi-layer autoencoder is built in the hospital server to perform unsupervised feature learning and reconstruction. It learns the inherent feature distribution of normal standard diagnosis and treatment samples, calculates the reconstruction loss of each sample in the high-dimensional sample feature set, and marks samples with reconstruction errors greater than the error threshold as level three suspected misdiagnosis samples. If a target sample is simultaneously marked as any two or more of the following: Level 1 suspected misdiagnosed sample, Level 2 suspected misdiagnosed sample, and Level 3 suspected misdiagnosed sample, then the target sample is determined to be a misdiagnosed sample. The misdiagnosed samples are assigned a suspicion score based on the number of suspected misdiagnosed samples marked, and an anomaly label is set to obtain an anomaly labeled sample set with suspicion scores; Based on the set of anomaly-labeled samples, a soft label correction strategy that combines knowledge-driven and data-driven approaches is used to calibrate the anomaly labels. Differentiated weights are assigned to misdiagnosed samples with different suspicion scores. Anomaly labels are calibrated and weights are bound to all misdiagnosed samples. The set of anomaly-labeled samples in the high-dimensional sample feature set is then updated to form a historical correction dataset.

[0009] Furthermore, based on the set of anomaly-labeled samples, a soft label correction strategy that integrates knowledge-driven and data-driven approaches is used to calibrate the anomaly labels, and differentiated weight allocation is applied to misdiagnosed samples with different suspicion scores. Anomaly label calibration and weight binding are performed on all misdiagnosed samples, and the set of anomaly-labeled samples in the high-dimensional sample feature set is updated to form a historical correction dataset, including: For each misdiagnosed sample in the set of anomaly-labeled samples, a rule-guided diagnostic probability distribution P_rule is obtained through reasoning using the medical knowledge graph; and for each misdiagnosed sample in the set of anomaly-labeled samples, a model-guided diagnostic probability distribution P_model is obtained through prediction using a pre-trained basic classification model. Substituting the rule-guided diagnostic probability distribution P_rule and the model-guided diagnostic probability distribution P_model into the soft label fusion formula P_new=λP_model+(1-λ)P_rule, we obtain the soft label probability distribution P_new; where λ is the fusion balance coefficient. The suspicion level of each misdiagnosed sample is determined based on the suspicion score corresponding to each misdiagnosed sample. The suspicion level includes credible normal sample, sample with ambiguous boundary, slightly suspected misdiagnosed sample and seriously suspected misdiagnosed sample. Training weights are assigned to misdiagnosed samples of each suspicion level. For each misdiagnosed sample, bind the corrected soft label and corresponding training weights according to the soft label probability distribution P_new; The set of anomaly-labeled samples in the high-dimensional sample feature set is updated to form a historical correction dataset.

[0010] Furthermore, the construction of the dual-channel feature adaptive fusion diagnostic network uses the historical corrected dataset and the outpatient feature vector as dual inputs. It employs a noise-robust loss function composed of a focusing loss function and label softening processing, and performs two-stage step-by-step training and progressive optimization of sample weights to obtain a converged medical auxiliary diagnostic model, including: A dual-channel feature adaptive fusion diagnostic network is built in the hospital server. The first channel uses a structured feature processing channel to process the structured diagnosis and treatment history data, and the second channel uses a temporal feature parsing channel to process the out-of-hospital physiological temporal data. The first and second channels are connected to an adaptive attention fusion layer at the back end. Input the historical correction dataset into the first channel and input the outpatient feature vector into the second channel; During the training of the dual-channel feature adaptive fusion diagnostic network, a noise-robust loss function is constructed by combining a focusing loss function with label softening processing to reduce gradient interference from misdiagnosed samples and abnormal data. The network backbone and feature extraction layer of the dual-channel feature adaptive fusion diagnostic network are pre-trained using the historical correction dataset; After pre-training is completed, the bottom feature layer of the dual-channel feature adaptive fusion diagnostic network is frozen, and the outpatient feature vector is used for cross-scenario fine-tuning. During the fine-tuning process, a progressive weight adjustment strategy is adopted to gradually increase the training weights of mild and severe suspected misdiagnosis samples. After the dual-channel feature adaptive fusion diagnostic network training converges, a medical auxiliary diagnostic model is obtained.

[0011] Furthermore, the step of inputting the outpatient feature vector into the medical auxiliary diagnostic model to perform forward inference and classification prediction, and outputting multi-class preliminary diagnostic confidence results, includes: Read the outpatient feature vector and perform dimensional normalization according to the input format of the medical auxiliary diagnosis model; The dimension-normalized outpatient feature vector is input into the medical auxiliary diagnosis model, and forward propagation inference is performed. After normalization processing by the maximum activation function at the end of the medical auxiliary diagnostic model, the predicted probability scores corresponding to each category of disease are output, and the confidence scores of multi-category preliminary diagnosis are obtained.

[0012] Furthermore, the step of performing logical conflict detection and correction on the multi-category preliminary diagnostic confidence results according to a preset medical logic post-verification rule base, obtaining a medical auxiliary diagnostic conclusion that satisfies medical logic constraints, and feeding it back to the home monitoring device manufacturer includes: A medical logic post-validation rule base is pre-built in the hospital server to detect and filter medical auxiliary diagnostic conclusions that violate medical logic; Read the anomaly type and anomaly level from the outpatient classification data, and determine the confidence decay coefficient η for each anomaly level according to the severity of the anomaly; Substitute the multi-category preliminary diagnostic confidence results into the confidence decay formula to obtain the calibrated diagnostic confidence; wherein, the confidence decay formula is P_final=P_init·η, P_init is the multi-category preliminary diagnostic confidence, η is the confidence decay coefficient for each abnormality level, and P_final is the calibrated diagnostic confidence; Based on the post-calibration diagnostic confidence level, the medical auxiliary diagnostic conclusions with logical conflicts in the multi-category preliminary diagnostic confidence level results are identified through the medical logic post-verification rule base. The medical auxiliary diagnostic conclusions with logical conflicts are filtered out, and medical auxiliary diagnostic conclusions that meet medical logic constraints are formed and fed back to the home monitoring device manufacturer for the home monitoring device manufacturer to generate remote monitoring health reports.

[0013] This application also provides an electronic device, including: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of any of the above-described knowledge graph-based remote health diagnosis result generation methods.

[0014] This application also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the steps of any of the above-described knowledge graph-based remote health diagnosis result generation methods.

[0015] This application first identifies various anomalies in out-of-hospital physiological time-series data based on a general medical time-series compliance rule library and completes feature standardization transformation. Then, it combines medical knowledge graphs, density clustering discrete point detection, and multi-layer autoencoders for triple screening to locate misdiagnosed samples. Labels are optimized through dual-path probabilistic fusion to generate a historical corrected dataset. A dual-channel feature adaptive fusion diagnostic network is built, and a diagnostic model is obtained through two-stage step-by-step training with a dedicated loss function. After the model outputs a preliminary diagnostic confidence score, it undergoes post-verification and correction using medical logic, and finally, the compliance conclusion is pushed to the home monitoring device manufacturer. This solution fills the gap in existing technologies that only support inter-hospital data interaction and cannot provide complete diagnostic services to manufacturers. It achieves joint diagnosis of heterogeneous data from inside and outside hospitals while ensuring the privacy and security of in-hospital medical records. The output conclusions comply with medical standards and can be directly used by manufacturers to generate remote health reports. Attached Figure Description

[0016] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 is a flowchart illustrating a method for generating remote health diagnosis results based on knowledge graphs in one embodiment of this application; Figure 2 is a schematic diagram of the post-verification of diagnostic results and confidence decay process in one embodiment of this application; Figure 3 is a structural block diagram of a remote health diagnosis result generation device based on knowledge graph in one embodiment of this application; Figure 4 is an internal structural diagram of a computer device in one embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0019] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., in this application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0020] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] It is understood that in the specific implementation of this application, data related to personnel, such as physiological data, are involved. When the embodiments of this application are applied to specific products or technologies, it is necessary to obtain separate permission or consent from the personnel. Furthermore, the collection, use, and processing of the relevant data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0022] like Figure 1 As shown, embodiments of this application provide a method for generating remote health diagnosis results based on knowledge graphs, including the following steps: S1. In response to receiving out-of-hospital physiological time-series data uploaded by the home monitoring device manufacturer, the out-of-hospital physiological time-series data is anomaly identified according to a preset general medical time-series compliance rule library, and anomaly type labels are set for the abnormal out-of-hospital physiological time-series data to obtain out-of-hospital classification data. S2. Standardize the outpatient classification data, suppress temporal noise and remove outliers to generate outpatient feature vectors with aligned feature dimensions and standardized distribution; S3. Based on the outpatient classification data, read the corresponding structured diagnosis and treatment history data in the hospital database. Identify misdiagnosed samples and set abnormal labels for the structured diagnosis and treatment history data through medical knowledge graph rule verification, density cluster discrete point detection, and self-supervised feature reconstruction error verification. Determine the suspiciousness score for the misdiagnosed samples and calibrate the abnormal labels to obtain the historical correction dataset. S4. Construct a dual-channel feature adaptive fusion diagnostic network. The historical corrected dataset and the outpatient feature vector are used as dual inputs. A noise-robust loss function consisting of a focusing loss function and label softening is adopted. Two-stage step-by-step training and progressive optimization of sample weights are performed to obtain a converged medical auxiliary diagnostic model. S5. Input the out-of-hospital feature vector into the medical auxiliary diagnosis model to perform forward inference and classification prediction, and output the multi-class preliminary diagnosis confidence results; S6. Based on the preset medical logic post-verification rule library, perform logical conflict detection and correction on the multi-category preliminary diagnosis confidence results, obtain medical auxiliary diagnosis conclusions that meet medical logic constraints, and feed them back to the home monitoring device manufacturer.

[0023] In this embodiment, the step of identifying anomalies in the outpatient physiological time-series data according to a preset general medical time-series compliance rule library, and setting anomaly type labels for abnormal outpatient physiological time-series data to obtain outpatient classification data includes: A general medical time series compliance rule library is pre-built on the hospital server. In the general medical time series compliance rule library, standard monitoring constraints, physiological normal range boundaries, resting / fasting / postprandial physiological state adaptation specifications, and reasonable ranges of time series physiological changes are defined for various monitoring items. The hospital server receives out-of-hospital physiological time-series data uploaded by the home monitoring device manufacturer. The out-of-hospital physiological time-series data includes at least one of structured indicator data, continuous time-series monitoring data, manually entered data, and data automatically collected by the device. Based on the general medical time series compliance rule library, the outpatient physiological time series data are anomaly identified from the dimensions of compliance of monitoring timing, compliance of operating procedures, rationality of numerical physiological intervals, rationality of short-term time series mutations, and rationality of equipment acquisition drift. Based on the screening results, abnormal out-of-hospital physiological time series data are labeled with abnormality type tags. The abnormality type includes at least one of the following: mismatched monitoring timing, non-standard operation, equipment drift, time series outlier, human entry, random noise, and physiological logic conflict. Each sample in the outpatient physiological time series data is bound to a corresponding abnormality type label and an abnormality level score to obtain outpatient classification data with abnormality type labels.

[0024] Among them, by pre-building a general medical time series compliance rule library containing multi-dimensional physiological norms, a unified judgment standard is established. This comprehensively covers time series data uploaded by various types of manufacturers and conducts batch compliance screening from five core dimensions. Simultaneously, seven typical abnormalities are distinguished and corresponding abnormality level scores are bound to form complete outpatient classification data. This enables automated identification of data defects caused by various non-standard monitoring behaviors in home collection scenarios without human intervention. It accurately marks the abnormality type and severity of each sample, providing a clear basis for subsequent standardized and regularized processing, and avoiding the distortion of results caused by unmarked abnormal data directly entering the diagnostic process.

[0025] In this embodiment, the step of standardizing, suppressing temporal noise, and removing outliers from the outpatient classification data to generate outpatient feature vectors with aligned feature dimensions and standardized distribution includes: In the hospital server, a multiple linear regression mapping model is pre-constructed based on the home non-standard original monitoring feature vector, the feature mapping weight matrix, the feature mapping weight matrix, and the transformed equivalent standard benchmark feature vector; wherein the multiple linear regression mapping model is Y=W·X+b; X is the home non-standard original monitoring feature vector, Y is the transformed equivalent standard benchmark feature vector, W is the feature mapping weight matrix, and b is the bias vector. The multivariate linear regression mapping model is trained by using normal standard medical samples and non-standard home samples to form training pairs, establishing a regression mapping relationship from the non-standard monitoring feature space to the standard benchmark physiological feature space, and obtaining a cross-time series benchmark mapping model. Substitute the non-standard home samples labeled with the abnormality type in the outpatient classification data into the cross-time series benchmark mapping model to perform equivalent standard feature conversion; A multi-mode joint filtering strategy is formed by combining median filtering, Kalman filtering, and sliding window mean filtering to remove single-point impulse outliers, smooth temporal drift, and suppress noise for each sample in the outpatient classification data. For each sample data, normalization and time series length alignment are performed to unify the feature dimensions and input format, thereby obtaining an out-of-hospital feature vector with aligned feature dimensions and standardized distribution.

[0026] Among them, the non-standard physiological data is transformed into a standard feature space by relying on a pre-trained multivariate linear regression cross-time-series benchmark mapping model. Then, a multi-level joint filtering strategy combining median, Kalman, and sliding window mean is used to remove pulse outliers, smooth time-series drift, and suppress random noise layer by layer. Finally, the normalization and time-series length are unified, which can eliminate the data distribution deviation caused by inconsistent home equipment collection standards, equipment drift, and environmental interference. It unifies the heterogeneous time-series data output by different manufacturers and different equipment into out-of-hospital feature vectors with consistent dimensions and stable distribution, ensuring the standardization of subsequent network input data and reducing model training and inference errors.

[0027] In this embodiment, based on the outpatient classification data, corresponding structured medical history data is read from the hospital database. The structured medical history data is then used to identify misdiagnosed samples and assign anomaly labels through medical knowledge graph rule verification, density clustering discrete point detection, and self-supervised feature reconstruction error verification. A suspicion score is determined for each misdiagnosed sample, and the anomaly labels are calibrated to obtain a historical correction dataset, including: A medical knowledge graph is built on the hospital server. The nodes of the medical knowledge graph include disease types, physiological test indicators, age groups, genders, complications, and clinical diagnostic threshold rules. The edges of the medical knowledge graph are associated with the fixed relationships between indicators and disease types, the constraints between population and disease incidence, and the mutually exclusive logical relationships between indicator combinations. Read the structured medical history data corresponding to the outpatient classification data in the hospital database, extract patient attribute features, test index features, and original diagnostic label features, and construct a high-dimensional sample feature set; Based on the medical knowledge graph, the high-dimensional sample feature set is subjected to rule logic verification, and samples that violate the clinical diagnosis threshold, population adaptation, and indicator combination logic are marked as first-level suspected misdiagnosis samples. The high-dimensional sample feature set is divided into clusters using the density clustering (DBSCAN) algorithm. The neighborhood radius (Eps) and the minimum number of cluster samples (MinPts) are set to identify outlier samples that are far from the main distribution clusters. These outlier samples are marked as secondary suspected misdiagnosed samples. A multi-layer autoencoder is built in the hospital server to perform unsupervised feature learning and reconstruction. It learns the inherent feature distribution of normal standard diagnosis and treatment samples, calculates the reconstruction loss (Loss) of each sample in the high-dimensional sample feature set, and marks samples with reconstruction errors greater than the error threshold as level three suspected misdiagnosis samples. If a target sample is simultaneously marked as any two or more of the following: Level 1 suspected misdiagnosed sample, Level 2 suspected misdiagnosed sample, and Level 3 suspected misdiagnosed sample, then the target sample is determined to be a misdiagnosed sample. The misdiagnosed samples are assigned a suspicion score based on the number of suspected misdiagnosed samples marked, and an anomaly label is set to obtain an anomaly labeled sample set with suspicion scores; Based on the set of anomaly-labeled samples, a soft label correction strategy that combines knowledge-driven and data-driven approaches is used to calibrate the anomaly labels. Differentiated weights are assigned to misdiagnosed samples with different suspicion scores. Anomaly labels are calibrated and weights are bound to all misdiagnosed samples. The set of anomaly-labeled samples in the high-dimensional sample feature set is then updated to form a historical correction dataset.

[0028] Among them, clinical judgment benchmarks are established based on a pre-built medical knowledge graph. Combined with DBSCAN density clustering, multi-layer autoencoder reconstruction loss, and three independent screening paths, three levels of suspected misdiagnosed samples are hierarchically labeled. High-risk misdiagnosed samples are accurately identified and the suspicion score is quantified through multi-condition joint judgment. All original diagnosis and treatment samples are completely retained without deletion. It can comprehensively mine mislabeled labels in medical records from three dimensions: medical rules, data distribution, and feature fitting. Rare diseases and marginal cases missed by a single screening method can be detected by complementary mechanisms, providing a complete and quantitative basis for sample suspicion for subsequent soft label calibration.

[0029] In this embodiment, based on the set of anomaly-labeled samples, a soft label correction strategy that integrates knowledge-driven and data-driven approaches is used to calibrate the anomaly labels, and differentiated weight allocation is applied to misdiagnosed samples with different suspicion scores. Anomaly label calibration and weight binding are performed on all misdiagnosed samples, and the set of anomaly-labeled samples in the high-dimensional sample feature set is updated to form a historical correction dataset, including: For each misdiagnosed sample in the set of anomaly-labeled samples, a rule-guided diagnostic probability distribution P_rule is obtained through reasoning using the medical knowledge graph; and for each misdiagnosed sample in the set of anomaly-labeled samples, a model-guided diagnostic probability distribution P_model is obtained through prediction using a pre-trained basic classification model. Substituting the rule-guided diagnostic probability distribution P_rule and the model-guided diagnostic probability distribution P_model into the soft label fusion formula P_new=λP_model+(1-λ)P_rule, we obtain the soft label probability distribution P_new; where λ is the fusion balance coefficient, λ∈[0.5,0.7]; The suspicion level of each misdiagnosed sample is determined based on the suspicion score corresponding to each misdiagnosed sample. The suspicion level includes credible normal sample, sample with ambiguous boundary, slightly suspected misdiagnosed sample and seriously suspected misdiagnosed sample. Training weights are assigned to misdiagnosed samples of each suspicion level. For each misdiagnosed sample, bind the corrected soft label and corresponding training weights according to the soft label probability distribution P_new; The set of anomaly-labeled samples in the high-dimensional sample feature set is updated to form a historical correction dataset.

[0030] The label optimization process, which generates a historical corrected dataset based on anomaly-labeled sample sets, simultaneously incorporates rule-guided diagnostic probabilities obtained from medical knowledge graph reasoning and model-guided diagnostic probabilities output by the basic classification model. Soft label weighting is performed using a fixed interval fusion balancing coefficient, and differentiated training weights are assigned to four levels of sample categories based on sample suspicion scores. All original diagnostic and treatment sample information is retained throughout the process. This approach balances clinical medical logic constraints with the inherent characteristics of the data, while reducing the training weight of highly suspicious samples to mitigate the interference of label noise on the model. The result is a historical corrected dataset that combines clinical credibility with data fitting capabilities, which can be directly used for the pre-training stage of a dual-channel network.

[0031] In this embodiment, the construction of a dual-channel feature adaptive fusion diagnostic network uses the historical corrected dataset and the outpatient feature vector as dual inputs. It employs a noise-robust loss function composed of a focusing loss function and label softening processing, and performs two-stage step-by-step training and progressive optimization of sample weights to obtain a converged medical auxiliary diagnostic model, including: A dual-channel feature adaptive fusion diagnostic network (Attention Fusion Neural Network) is built in the hospital server. The first channel uses a structured feature processing channel to process the structured diagnosis and treatment history data, and the second channel uses a temporal feature parsing channel to process the out-of-hospital physiological time-series data. The first and second channels are connected to an adaptive attention fusion layer at the back end. Input the historical correction dataset into the first channel and input the outpatient feature vector into the second channel; During the training of the dual-channel feature adaptive fusion diagnostic network, a noise-robust loss function is constructed by combining the FocalLoss function and the LabelSmoothing function to reduce gradient interference from misdiagnosed samples and abnormal data. The network backbone and feature extraction layer of the dual-channel feature adaptive fusion diagnostic network are pre-trained using the historical correction dataset; After pre-training is completed, the bottom feature layer of the dual-channel feature adaptive fusion diagnostic network is frozen, and the outpatient feature vector is used for cross-scenario fine-tuning. During the fine-tuning process, a progressive weight adjustment strategy is adopted to gradually increase the training weights of mild and severe suspected misdiagnosis samples. After the dual-channel feature adaptive fusion diagnostic network training converges, a medical auxiliary diagnostic model is obtained.

[0032] The system sets up separate channels for structured feature processing and temporal feature parsing to adapt to two types of heterogeneous medical data. The backend uses an adaptive attention fusion layer to dynamically allocate feature weights. It combines a focus loss and a label softening combined with a noise-robust loss function to reduce gradient interference from misdiagnosed samples. The network backbone is first pre-trained based on a historical corrected dataset. Then, the bottom layer is frozen and only the fusion and output layers are fine-tuned, and the training weights of suspicious samples are gradually increased. This fully explores the complementary features of standardized medical records in hospitals and temporal monitoring data outside hospitals, avoids model deviation during cross-scenario training, and the medical auxiliary diagnostic model obtained after training convergence has stronger anti-interference ability and cross-scenario generalization performance.

[0033] In this embodiment, the step of inputting the outpatient feature vector into the medical auxiliary diagnostic model to perform forward inference and classification prediction, and outputting multi-class preliminary diagnostic confidence results, includes: Read the outpatient feature vector and perform dimensional normalization according to the input format of the medical auxiliary diagnosis model; The dimension-normalized outpatient feature vector is input into the medical auxiliary diagnostic model, and forward propagation inference is performed. After normalization processing by the maximum activation function (Softmax) at the end of the medical auxiliary diagnostic model, the predicted probability scores corresponding to each category of disease are output to obtain the multi-category preliminary diagnostic confidence.

[0034] After the outpatient feature vectors are dimensionally normalized according to the model input specifications, they are fed into the network for forward propagation. The predicted scores for various diseases are output through the Softmax function. The fused unified feature vectors can be automatically mapped to quantitative confidence values ​​of various health risks, intuitively presenting the probability of different diseases. This provides standardized numerical input for subsequent medical logic verification and adaptive confidence correction, ensuring the uniformity of the calculation benchmark in the post-verification process.

[0035] like Figure 2 As shown, in this embodiment, the step of performing logical conflict detection and correction on the multi-category preliminary diagnostic confidence results according to a preset medical logic post-verification rule base, obtaining a medical auxiliary diagnostic conclusion that satisfies medical logic constraints, and feeding it back to the home monitoring device manufacturer includes: A medical logic post-validation rule base is pre-built in the hospital server to detect and filter medical auxiliary diagnostic conclusions that violate medical logic; Read the anomaly type and anomaly level from the outpatient classification data, and determine the confidence decay coefficient η for each anomaly level according to the severity of the anomaly; Substitute the multi-category preliminary diagnostic confidence results into the confidence decay formula to obtain the calibrated diagnostic confidence; wherein, the confidence decay formula is P_final=P_init·η, P_init is the multi-category preliminary diagnostic confidence, η is the confidence decay coefficient for each abnormality level, and P_final is the calibrated diagnostic confidence; Based on the post-calibration diagnostic confidence level, the medical auxiliary diagnostic conclusions with logical conflicts in the multi-category preliminary diagnostic confidence level results are identified through the medical logic post-verification rule base. The medical auxiliary diagnostic conclusions with logical conflicts are filtered out, and medical auxiliary diagnostic conclusions that meet medical logic constraints are formed and fed back to the home monitoring device manufacturer for the home monitoring device manufacturer to generate remote monitoring health reports.

[0036] Specifically, by pre-building a medical logic post-verification rule base, conflicting diagnostic conclusions that violate clinical common sense are intercepted. Based on the abnormality level of outpatient classification data, corresponding confidence attenuation coefficients are matched to complete adaptive probability correction. After filtering invalid conflicting conclusions, compliant medical auxiliary diagnostic conclusions are pushed to home monitoring device manufacturers. This can automatically offset the inflated diagnostic confidence caused by abnormal home data, eliminate erroneous outputs that do not conform to clinical common sense, and deliver rigorous and compliant diagnostic conclusions to manufacturers. They can be directly used for remote monitoring health report generation without additional medical review by manufacturers.

[0037] In one embodiment, such as Figure 3 As shown, a remote health diagnosis result generation device 10 based on knowledge graph is provided, including: a data identification module 1, a feature normalization module 2, a hospital data optimization module 3, a model training module 4, an inference and prediction module 5, and a diagnosis output module 6.

[0038] The data identification module 1 is used to respond to receiving out-of-hospital physiological time series data uploaded by the home monitoring device manufacturer, and to identify anomalies in the out-of-hospital physiological time series data according to a preset general medical time series compliance rule library, and to set anomaly type labels for the abnormal out-of-hospital physiological time series data to obtain out-of-hospital classification data.

[0039] The feature normalization module 2 is used to standardize the outpatient classification data, suppress temporal noise and remove outliers, and generate outpatient feature vectors with aligned feature dimensions and standardized distribution.

[0040] The hospital data optimization module 3 is used to read the corresponding structured diagnosis and treatment history data in the hospital database based on the outpatient classification data. The structured diagnosis and treatment history data is used to identify misdiagnosed samples and set abnormal labels through medical knowledge graph rule verification, density clustering discrete point detection, and self-supervised feature reconstruction error verification. The suspiciousness score of the misdiagnosed samples is determined and the abnormal labels are calibrated to obtain the historical correction dataset.

[0041] The model training module 4 is used to construct a dual-channel feature adaptive fusion diagnostic network. It takes the historical corrected dataset and the outpatient feature vector as dual inputs, and adopts a noise robust loss function composed of a focusing loss function and label softening processing. It performs two-stage step-by-step training and progressive optimization of sample weights to obtain a converged medical auxiliary diagnostic model.

[0042] The reasoning and prediction module 5 is used to input the outpatient feature vector into the medical auxiliary diagnosis model to perform forward reasoning and classification prediction, and output multi-category preliminary diagnosis confidence results.

[0043] The diagnostic output module 6 is used to perform logical conflict detection and correction on the multi-category preliminary diagnostic confidence results according to the preset medical logic post-verification rule library, obtain medical auxiliary diagnostic conclusions that meet medical logic constraints, and feed them back to the home monitoring device manufacturer.

[0044] The aforementioned knowledge graph-based remote health diagnosis result generation device first identifies various anomalies in out-of-hospital physiological time-series data and completes feature standardization transformation based on a general medical time-series compliance rule library. Then, it combines medical knowledge graphs, density clustering discrete point detection, and multi-layer autoencoder triple screening to locate misdiagnosed samples. Labels are optimized through dual-path probabilistic fusion to generate a historical corrected dataset. A dual-channel feature adaptive fusion diagnostic network is built, and a dedicated loss function is used to complete two-stage step-by-step training to obtain a diagnostic model. After the model inferences and outputs a preliminary diagnostic confidence score, it undergoes post-verification and correction using medical logic, and finally, the compliant conclusion is pushed to the home monitoring device manufacturer. This solution fills the gap in existing technologies that only support inter-hospital data interaction and cannot provide complete diagnostic services to manufacturers. It achieves joint diagnosis of heterogeneous data from inside and outside hospitals while ensuring the privacy and security of in-hospital medical records. The output conclusions comply with medical standards and can be directly used by manufacturers to generate remote health reports.

[0045] For a description of the features in the embodiment of the knowledge graph-based remote health diagnosis result generation device, please refer to the relevant description of the embodiment of the knowledge graph-based remote health diagnosis result generation method, which will not be repeated here.

[0046] Embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above embodiments of the knowledge graph-based remote health diagnosis result generation method.

[0047] In one embodiment, the electronic device may be a server, and its internal structure diagram may be as follows: Figure 4As shown, the electronic device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores remote health diagnosis result generation data based on a knowledge graph. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for generating remote health diagnosis results based on a knowledge graph.

[0048] Embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above embodiments of the knowledge graph-based remote health diagnosis result generation method.

[0049] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0050] The embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the knowledge graph-based remote health diagnosis result generation method.

[0051] Embodiments of this application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in any of the above embodiments of the knowledge graph-based remote health diagnosis result generation method.

[0052] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0053] The foregoing has provided a detailed description of a knowledge graph-based remote health diagnosis result generation method, device, and medium. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of these embodiments are only intended to aid in understanding the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the scope of protection of the claims.

Claims

1. A method for generating remote health diagnosis results based on knowledge graphs, used on a hospital server, characterized in that, include: In response to receiving out-of-hospital physiological time-series data uploaded by the home monitoring device manufacturer, the system performs anomaly identification on the out-of-hospital physiological time-series data according to a preset general medical time-series compliance rule library, sets anomaly type labels for abnormal out-of-hospital physiological time-series data, and obtains out-of-hospital classification data. The outpatient classification data is standardized, time-series noise is suppressed, and outliers are removed to generate outpatient feature vectors with aligned feature dimensions and standardized distribution. Based on the outpatient classification data, the corresponding structured diagnosis and treatment history data in the hospital database is read. The structured diagnosis and treatment history data is used to identify misdiagnosed samples and set abnormal labels through medical knowledge graph rule verification, density cluster discrete point detection, and self-supervised feature reconstruction error verification. The suspiciousness score of the misdiagnosed samples is determined and the abnormal labels are calibrated to obtain the historical correction dataset. A dual-channel feature adaptive fusion diagnostic network is constructed, with the historical corrected dataset and the outpatient feature vector as dual inputs. A noise-robust loss function composed of a focusing loss function and label softening is adopted. Two-stage step-by-step training and progressive optimization of sample weights are performed to obtain a converged medical auxiliary diagnostic model. The out-of-hospital feature vector is input into the medical auxiliary diagnostic model to perform forward inference and classification prediction, and outputs multi-class preliminary diagnostic confidence results; Based on a preset medical logic post-verification rule base, the multi-category preliminary diagnostic confidence results are subjected to logic conflict detection and correction to obtain a medical auxiliary diagnostic conclusion that meets medical logic constraints and is fed back to the home monitoring device manufacturer.

2. The method for generating remote health diagnosis results based on knowledge graphs according to claim 1, characterized in that, The process involves identifying anomalies in the outpatient physiological time-series data based on a pre-defined general medical time-series compliance rule base, assigning anomaly type labels to the abnormal outpatient physiological time-series data, and obtaining outpatient classification data, including: A general medical time series compliance rule library is pre-built on the hospital server. In the general medical time series compliance rule library, standard monitoring constraints, physiological normal range boundaries, resting / fasting / postprandial physiological state adaptation specifications, and reasonable ranges of time series physiological changes are defined for various monitoring items. The hospital server receives out-of-hospital physiological time-series data uploaded by the home monitoring device manufacturer; Based on the general medical time series compliance rule library, the outpatient physiological time series data are anomaly identified from the dimensions of compliance of monitoring timing, compliance of operating procedures, rationality of numerical physiological intervals, rationality of short-term time series mutations, and rationality of equipment acquisition drift. Based on the screening results, abnormal out-of-hospital physiological time-series data with abnormalities are labeled with abnormality type tags; Each sample in the outpatient physiological time series data is bound to a corresponding abnormality type label and an abnormality level score to obtain outpatient classification data with abnormality type labels.

3. The method for generating remote health diagnosis results based on knowledge graphs according to claim 1, characterized in that, The process of standardizing, suppressing temporal noise, and removing outliers from the outpatient classification data to generate outpatient feature vectors with aligned feature dimensions and standardized distribution includes: In the hospital server, a multiple linear regression mapping model is pre-constructed based on the home non-standard original monitoring feature vector, feature mapping weight matrix, feature mapping weight matrix and the transformed equivalent standard benchmark feature vector; The multivariate linear regression mapping model is trained by using normal standard medical samples and non-standard home samples to form training pairs, establishing a regression mapping relationship from the non-standard monitoring feature space to the standard benchmark physiological feature space, and obtaining a cross-time series benchmark mapping model. Substitute the non-standard home samples labeled with the abnormality type in the outpatient classification data into the cross-time series benchmark mapping model to perform equivalent standard feature conversion. A multi-mode joint filtering strategy is formed by combining median filtering, Kalman filtering, and sliding window mean filtering to remove single-point impulse outliers, smooth temporal drift, and suppress noise for each sample in the outpatient classification data. For each sample data, normalization and time series length alignment are performed to unify the feature dimensions and input format, thereby obtaining an out-of-hospital feature vector with aligned feature dimensions and standardized distribution.

4. The method for generating remote health diagnosis results based on knowledge graphs according to claim 1, characterized in that, Based on the aforementioned outpatient classification data, corresponding structured medical history data is read from the hospital database. The structured medical history data is then processed using medical knowledge graph rule verification, density clustering discrete point detection, and self-supervised feature reconstruction error verification to identify misdiagnosed samples and assign anomaly labels. A suspicion score is determined for each misdiagnosed sample, and the anomaly labels are calibrated to obtain a historical correction dataset, including: A medical knowledge graph is built in the hospital server. The nodes of the medical knowledge graph include disease types, physiological test indicators, age groups, genders, complications, and clinical diagnostic threshold rules. The edges of the medical knowledge graph are associated with the fixed relationships between indicators and disease types, the constraints between population and disease incidence, and the mutually exclusive logical relationships between indicator combinations. Read the structured medical history data corresponding to the outpatient classification data in the hospital database, extract patient attribute features, test index features, and original diagnostic label features, and construct a high-dimensional sample feature set; Based on the medical knowledge graph, the high-dimensional sample feature set is subjected to rule logic verification, and samples that violate the clinical diagnosis threshold, population adaptation, and indicator combination logic are marked as first-level suspected misdiagnosis samples. The high-dimensional sample feature set is divided into clusters using a density clustering algorithm. The neighborhood radius and minimum number of cluster samples are set to identify outliers that are far from the main distribution clusters. These outliers are marked as secondary suspected misdiagnosed samples. A multi-layer autoencoder is built in the hospital server to perform unsupervised feature learning and reconstruction. It learns the inherent feature distribution of normal standard diagnosis and treatment samples, calculates the reconstruction loss of each sample in the high-dimensional sample feature set, and marks samples with reconstruction errors greater than the error threshold as level three suspected misdiagnosis samples. If a target sample is simultaneously marked as any two or more of the following: Level 1 suspected misdiagnosed sample, Level 2 suspected misdiagnosed sample, and Level 3 suspected misdiagnosed sample, then the target sample is determined to be a misdiagnosed sample. The misdiagnosed samples are assigned a suspicion score based on the number of suspected misdiagnosed samples and an anomaly label is set to obtain an anomaly labeled sample set with suspicion scores; Based on the set of anomaly-labeled samples, a soft label correction strategy that combines knowledge-driven and data-driven approaches is used to calibrate the anomaly labels. Differentiated weights are assigned to misdiagnosed samples with different suspicion scores. Anomaly labels are calibrated and weights are bound to all misdiagnosed samples. The set of anomaly-labeled samples in the high-dimensional sample feature set is then updated to form a historical correction dataset.

5. The method for generating remote health diagnosis results based on knowledge graphs according to claim 4, characterized in that, Based on the anomaly-annotated sample set, a soft label correction strategy that integrates knowledge-driven and data-driven approaches is used to calibrate the anomaly labels. Differential weight allocation is applied to misdiagnosed samples with different suspicion scores. Anomaly label calibration and weight binding are performed on all misdiagnosed samples. The anomaly-annotated sample set in the high-dimensional sample feature set is then updated to form a historical correction dataset, including: For each misdiagnosed sample in the set of anomaly-labeled samples, a rule-guided diagnostic probability distribution P_rule is obtained through reasoning using the medical knowledge graph; and for each misdiagnosed sample in the set of anomaly-labeled samples, a model-guided diagnostic probability distribution P_model is obtained through prediction using a pre-trained basic classification model. Substituting the rule-guided diagnostic probability distribution P_rule and the model-guided diagnostic probability distribution P_model into the soft label fusion formula P_new=λP_model+(1-λ)P_rule, we obtain the soft label probability distribution P_new; where λ is the fusion balance coefficient. The suspicion level of each misdiagnosed sample is determined based on the suspicion score corresponding to each misdiagnosed sample. The suspicion level includes credible normal sample, sample with ambiguous boundary, slightly suspected misdiagnosed sample and seriously suspected misdiagnosed sample. Training weights are assigned to misdiagnosed samples of each suspicion level. For each misdiagnosed sample, bind the corrected soft label and corresponding training weights according to the soft label probability distribution P_new; The set of anomaly-labeled samples in the high-dimensional sample feature set is updated to form a historical correction dataset.

6. The method for generating remote health diagnosis results based on knowledge graphs according to claim 1, characterized in that, The constructed dual-channel feature adaptive fusion diagnostic network uses the historical corrected dataset and the outpatient feature vector as dual inputs. It employs a noise-robust loss function composed of a focusing loss function and label softening, and performs two-stage step-by-step training and progressive optimization of sample weights to obtain a converged medical auxiliary diagnostic model, including: A dual-channel feature adaptive fusion diagnostic network is built in the hospital server. The first channel uses a structured feature processing channel to process the structured diagnosis and treatment history data, and the second channel uses a temporal feature parsing channel to process the out-of-hospital physiological temporal data. The first and second channels are connected to an adaptive attention fusion layer at the back end. Input the historical correction dataset into the first channel and input the outpatient feature vector into the second channel; During the training of the dual-channel feature adaptive fusion diagnostic network, a noise-robust loss function is constructed by combining a focusing loss function with label softening processing to reduce gradient interference from misdiagnosed samples and abnormal data. The network backbone and feature extraction layer of the dual-channel feature adaptive fusion diagnostic network are pre-trained using the historical correction dataset. After pre-training is completed, the bottom feature layer of the dual-channel feature adaptive fusion diagnostic network is frozen, and the outpatient feature vector is used for cross-scenario fine-tuning. During the fine-tuning process, a progressive weight adjustment strategy is adopted to gradually increase the training weights of mild and severe suspected misdiagnosis samples. After the dual-channel feature adaptive fusion diagnostic network training converges, a medical auxiliary diagnostic model is obtained.

7. The method for generating remote health diagnosis results based on knowledge graphs according to claim 1, characterized in that, The step involves inputting the outpatient feature vector into the medical auxiliary diagnostic model to perform forward inference and classification prediction, and outputting multi-class preliminary diagnostic confidence results, including: Read the outpatient feature vector and perform dimensional normalization according to the input format of the medical auxiliary diagnosis model; The dimension-normalized outpatient feature vector is input into the medical auxiliary diagnosis model, and forward propagation inference is performed. After normalization processing by the maximum activation function at the end of the medical auxiliary diagnostic model, the predicted probability scores corresponding to each category of disease are output, and the confidence scores of multi-category preliminary diagnosis are obtained.

8. The method for generating remote health diagnosis results based on knowledge graphs according to claim 7, characterized in that, The process of performing logical conflict detection and correction on the multi-category preliminary diagnostic confidence results according to a preset medical logic post-verification rule base, obtaining a medical auxiliary diagnostic conclusion that satisfies medical logic constraints, and feeding it back to the home monitoring device manufacturer includes: A medical logic post-validation rule base is pre-built in the hospital server to detect and filter medical auxiliary diagnostic conclusions that violate medical logic; Read the anomaly type and anomaly level from the outpatient classification data, and determine the confidence decay coefficient η for each anomaly level according to the severity of the anomaly; Substitute the multi-category preliminary diagnostic confidence results into the confidence decay formula to obtain the calibrated diagnostic confidence; wherein, the confidence decay formula is P_final=P_init·η, P_init is the multi-category preliminary diagnostic confidence, η is the confidence decay coefficient for each abnormality level, and P_final is the calibrated diagnostic confidence; Based on the post-calibration diagnostic confidence level, the medical auxiliary diagnostic conclusions with logical conflicts in the multi-category preliminary diagnostic confidence level results are identified through the medical logic post-verification rule base. The medical auxiliary diagnostic conclusions with logical conflicts are filtered out, and medical auxiliary diagnostic conclusions that meet medical logic constraints are formed and fed back to the home monitoring device manufacturer for the home monitoring device manufacturer to generate remote monitoring health reports.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the knowledge graph-based remote health diagnosis result generation method as described in any one of claims 1 to 8 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, it implements the steps of the knowledge graph-based remote health diagnosis result generation method as described in any one of claims 1 to 8.

Images