A pathological intelligent review method and system based on case history comparison

Abstract

This invention discloses a pathological intelligent review method and system based on historical case comparison. The method includes multi-timepoint structured case archiving, pathological indicator evolution modeling, review anomaly detection, pathological deviation analysis, and intelligent pathological review. This invention relates to the field of medical big data analysis technology. It employs a multi-stage evolutionary modeling method with temporal residual sensitivity to perform trend learning and anomaly assessment on key pathological indicators. Combined with a causal-aware spatiotemporal pathological deviation analysis network, it achieves graph structure modeling of the deviation relationships and risk factors between indicators, outputting deviation feature patterns. Finally, it generates review anomaly correction suggestions with causal attribution capabilities, improving the accuracy, interpretability, and clinical decision support capabilities of pathological review, and has good clinical adaptability and promotional value.

Description

Technical Field

[0001] This invention relates to the field of medical big data analysis technology, specifically to a pathological intelligent review method and system based on historical case comparison. Background Technology

[0002] The pathology intelligent review method based on case history comparison refers to the intelligent medical assistance method that collects and organizes patients' pathological examination data and medical records at different time points, uses a computer system to perform structured processing and comparative analysis of this historical information, identifies the changing trends and deviations between the current pathological state and the previous state, thereby assisting doctors in discovering potential abnormalities, verifying initial diagnosis conclusions, or providing review suggestions. Its main function is to improve the accuracy of pathological judgment, reduce human omissions, and provide data support and risk warnings for clinical diagnosis and subsequent treatment.

[0003] However, existing intelligent pathology review methods based on historical case comparisons suffer from technical problems such as a lack of systematic modeling, reliance on static index comparisons, and difficulty in identifying dynamic evolution anomalies and related causes of deviations. Existing pathology index evolution modeling methods suffer from technical problems such as a lack of dynamic response capabilities, insufficient sensitivity to fluctuations, and difficulty in adapting to nonlinear temporal behaviors such as index mutations, lags, or abnormal rebounds. Existing pathology deviation analysis methods suffer from technical problems such as relying solely on static anomaly determinations, failing to reconstruct the interaction structure between indices, and ignoring the driving role of abnormal propagation paths and inducing variables on deviation patterns. Summary of the Invention

[0004] In view of the above situation and to overcome the shortcomings of the prior art, the technical solution adopted by the present invention is as follows: The present invention provides a pathological intelligent review method based on case history comparison, the method comprising the following steps:

[0005] Step S1: Structured archiving of multi-timepoint cases;

[0006] Step S2: Modeling the evolution of pathological indicators;

[0007] Step S3: Verify anomaly detection;

[0008] Step S4: Pathological deviation analysis;

[0009] Step S5: Intelligent Pathology Review.

[0010] Furthermore, in step S1, the multi-time point case structured archiving is used to structure and archive case information. Specifically, it involves extracting pathological case information from the medical system to obtain raw case information data, and then constructing structured time series archive data through structured processing and time series archiving.

[0011] The pathological case information includes pathological test reports and pathological image labeling results at historical time points.

[0012] Further, in step S2, the pathological indicator evolution modeling, used to extract key pathological indicators and establish a time evolution model, specifically involves using the structured time series archive data and a multi-stage pathological evolution modeling method with temporal residual sensitivity to perform pathological indicator evolution modeling, obtaining indicator evolution trend data, including the following steps:

[0013] Step S21: Key indicator screening, used to highlight indicators that are strongly related to disease evolution, specifically using time-varying importance scores to screen pathological key indicators and obtain pathological key indicator data;

[0014] Step S22: Sliding time window modeling is used to capture the short-term evolution pattern of indicators within a local time period and enhance the model's adaptability to mutations and lag phenomena. Specifically, a three-channel time window time series model is constructed, and time series features are extracted based on the pathological key indicator data to obtain three-channel time series feature data.

[0015] The three-channel time window timing model specifically includes a short-term window, a medium-term window, and a long-term window.

[0016] The short-term window uses a one-dimensional convolutional kernel for short-term feature extraction.

[0017] The intermediate window uses a long short-term neural network to extract periodic patterns.

[0018] The long-term window uses linear trend fitting calculation to calculate the trend slope feature.

[0019] Step S23: Temporal attention trend modeling, used to perform regression modeling of the indicator trend within a local window. Specifically, based on the three-channel time series feature data, attention trend modeling is performed by designing temporal position encoding and constructing a temporal attention layer, and error optimization is performed by introducing a dynamic residual sensitive regularization term to obtain trend modeling feature data.

[0020] Step S24: Global trend fusion, used to fuse the prediction results in all windows, specifically by using the trend modeling feature data and exponentially weighted moving average to perform global trend fusion and smooth reconstruction of fusion features to obtain global trend fusion feature data;

[0021] Step S25: Trend anomaly assessment, used to establish a time evolution model and predict trend reliability. Specifically, based on the global trend fusion feature data, the model is trained to obtain a pathological indicator evolution model, and the pathological indicator evolution trend is predicted by using the pathological indicator evolution model to obtain indicator evolution trend data.

[0022] Further, in step S3, the verification anomaly detection is used to identify pathological indicators that deviate from the normal evolutionary pattern. Specifically, based on the indicator evolution trend data, residual calculation is performed using the standard error analysis method to obtain indicator prediction residual data. Anomaly judgment threshold is constructed, and based on the indicator prediction residual data, anomaly residual judgment is performed to obtain abnormal evolution detection data. Adjacent time periods of abnormal evolution detection data that meet the conditions of continuous deviation or indicator mutation are merged and divided into abnormal evolution detection intervals, labeled, and abnormal evolution detection interval data are obtained.

[0023] Further, in step S4, the pathological deviation analysis is used to identify potential risk factors and deviation patterns behind indicator fluctuations. Specifically, based on the abnormal evolution detection interval data, a causal-aware spatiotemporal pathological deviation analysis network is used to perform pathological deviation analysis to obtain deviation feature pattern data, including the following steps:

[0024] Step S41: Deviation interval differential temporal coding, specifically, introducing a temporal differential operator and combining it with cross-attention coding to perform cross-branch coding on the abnormal evolution detection interval data to obtain deviation interval differential temporal coding vector data;

[0025] Step S42: Construction of pathological offset causal graph, specifically, based on the deviation interval differential temporal coding vector data, constructing a pathological offset causal graph to obtain pathological offset causal graph data;

[0026] Each node of the pathological offset causal graph data is used to represent a key pathological indicator; each edge of the pathological offset causal graph data is used to represent the offset propagation relationship between indicators detected in the abnormal interval; the set of edge weights of the pathological offset causal graph data is used to represent the intensity quantification value of offset propagation.

[0027] Step S43: Offset path subgraph classification extraction, specifically, based on the pathological offset causal graph data, index offset paths are extracted, and the risk types of index offsets are clustered into abnormal types to obtain offset path classification subgraph data;

[0028] Step S44: Offset risk analysis, specifically, based on the offset path classification sub-map data, the pathological offset pattern of each abnormal interval is analyzed, and offset risk analysis is performed through statistical correlation methods to obtain offset risk tracing data;

[0029] Step S45: Offset feature pattern data generation, specifically, based on the offset risk tracing data, offset feature pattern data is integrated to obtain deviation feature pattern data.

[0030] Furthermore, in step S5, the pathological intelligent review is used to generate review suggestions by combining deviation features with the comparison results of historical cases. Specifically, based on the deviation feature pattern data, the review suggestions are output in a structured manner to obtain review abnormality correction reference data.

[0031] The present invention provides a pathological intelligent review system based on historical case comparison, including a case archiving module, a pathological modeling module, an abnormality initial detection module, a deviation analysis module, and a pathological review module;

[0032] The case archiving module is used for structured archiving of cases at multiple time points. Through structured archiving of cases at multiple time points, structured time series archive data is obtained, and the structured time series archive data is sent to the pathology modeling module.

[0033] The pathological modeling module is used for pathological indicator evolution modeling. Through pathological indicator evolution modeling, indicator evolution trend data is obtained, and the indicator evolution trend data is sent to the anomaly initial detection module.

[0034] The anomaly initial detection module is used to verify the anomaly detection, obtain the anomaly evolution detection interval data by verifying the anomaly detection, and send the anomaly evolution detection interval data to the deviation analysis module;

[0035] The deviation analysis module is used for pathological deviation analysis. Through pathological deviation analysis, deviation feature pattern data is obtained, and the deviation feature pattern data is sent to the pathological review module.

[0036] The pathology review module is used for intelligent pathology review, and through intelligent pathology review, reference data for correcting review abnormalities is obtained.

[0037] The beneficial effects achieved by the present invention using the above solution are as follows:

[0038] (1) In view of the technical problems in the existing pathological intelligent review methods based on case history comparison, which lack systematic modeling, only perform static index comparison, and are difficult to identify dynamic evolution abnormalities and related deviation causes, this solution creatively adopts a comprehensive intelligent method that combines pathological index evolution modeling with review abnormality detection and pathological deviation analysis. It realizes the whole-chain review process from index trend modeling to abnormality identification and deviation attribution, effectively improving the accuracy, intelligence and clinical decision reference value of pathological review.

[0039] (2) In view of the technical problems in existing pathological index evolution modeling methods, such as lack of dynamic response capability, insufficient sensitivity to fluctuations, and difficulty in adapting to nonlinear temporal behaviors such as index mutation, lag or abnormal rebound, this scheme creatively adopts a multi-stage pathological evolution modeling method with temporal residual sensitive perception to model pathological index evolution, realizes high robustness fitting of abnormal trends and error-driven modeling optimization, and enhances the model's ability to identify nonlinear evolution patterns and clinical sensitivity.

[0040] (3) In view of the technical problems in existing pathological deviation analysis methods, which rely solely on static abnormality judgment, cannot restore the interaction structure between indicators, and ignore the driving role of abnormal propagation paths and causal variables on deviation patterns, this scheme creatively adopts a causal perception spatiotemporal pathological deviation analysis network to perform pathological deviation analysis. It realizes the unified modeling of causal mapping of deviation relationships between indicators, abnormal path identification and risk factor causal association, thereby improving the resolution of pathological abnormality types and the pertinence of clinical review. Attached Figure Description

[0041] Figure 1 A flowchart illustrating a pathological intelligent review method based on case history comparison provided by the present invention;

[0042] Figure 2 A schematic diagram of a pathological intelligent review system based on case history comparison provided by the present invention;

[0043] Figure 3 A flowchart illustrating the evolution of pathological indicators in step S2;

[0044] Figure 4 This is a flowchart illustrating the pathological deviation analysis in step S4.

[0045] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0046] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0047] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0048] Example 1, see Figure 1 This invention provides a pathological intelligent review method based on case history comparison, which includes the following steps:

[0049] Step S1: Structured archiving of multi-timepoint cases;

[0050] Step S2: Modeling the evolution of pathological indicators;

[0051] Step S3: Verify anomaly detection;

[0052] Step S4: Pathological deviation analysis;

[0053] Step S5: Intelligent Pathology Review.

[0054] By performing the above operations, this solution addresses the technical problems of existing intelligent pathology review methods based on historical case comparison, which lack systematic modeling, only perform static indicator comparison, and are difficult to identify dynamic evolution anomalies and related deviation causes. This solution creatively adopts a comprehensive intelligent method that combines pathological indicator evolution modeling with review anomaly detection and pathological deviation analysis. It realizes a full-chain review process from indicator trend modeling to anomaly identification and deviation attribution, effectively improving the accuracy, intelligence, and clinical decision reference value of pathology review.

[0055] Specifically, most current intelligent review solutions based on case history, such as those relying solely on threshold comparisons or classification model scoring, cannot reveal the changing trends of indicators over a continuous period of time, nor can they explain "why the deviation occurred" or "which indicators showed abnormal linkages." For example, although the creatinine value of a patient with chronic kidney disease is slowly rising within the normal range, traditional systems struggle to identify "trend deterioration." This solution, by constructing a trend model, dynamic residual detection, and causal offset mapping, can capture this type of slowly changing but dangerous evolutionary behavior.

[0056] Example 2, see Figure 1 and Figure 2 In step S1, the multi-time point case structured archiving is used to archive case information in a structured manner. Specifically, it involves extracting pathological case information from the medical system to obtain raw case information data, and then constructing structured time series archive data through structured processing and time series archiving.

[0057] The pathological case information includes pathological test reports and pathological image labeling results at historical time points;

[0058] The structured processing specifically refers to standardizing fields, unifying indicators, and normalizing units in the original case information data to obtain a structured indicator vector for the case.

[0059] The aforementioned time-series archiving construction specifically refers to integrating the structured indicator vectors of cases at multiple time points into multi-time-point structured time-series archive data in chronological order.

[0060] Example 3, see Figure 1 , Figure 2 and Figure 3 This embodiment is based on the above embodiment. In step S2, the pathological indicator evolution modeling is used to extract key pathological indicators and establish a time evolution model. Specifically, based on the structured time series archive data, a multi-stage pathological evolution modeling method with temporal residual sensitive perception is used to perform pathological indicator evolution modeling to obtain indicator evolution trend data, including the following steps:

[0061] Step S21: Key indicator screening, used to highlight indicators that are strongly related to disease evolution, specifically using event importance score to screen pathological key indicators and obtain pathological key indicator data.

[0062] Step S22: Sliding time window modeling is used to capture the short-term evolution pattern of indicators within a local time period and enhance the model's adaptability to mutations and lag phenomena. Specifically, a three-channel time window time series model is constructed, and time series features are extracted based on the pathological key indicator data to obtain three-channel time series feature data.

[0063] The three-channel time window timing model specifically includes a short-term window, a medium-term window, and a long-term window.

[0064] The short-term window uses a one-dimensional convolutional kernel for short-term feature extraction.

[0065] The intermediate window uses a long short-term neural network to extract periodic patterns.

[0066] The long-term window uses linear trend fitting calculation to calculate the trend slope feature.

[0067] Preferably, Table 1 is a structural example table of the three-channel time window timing model. As shown in the table, the time window length of the short-term window is 3 to 5 days, the time window length of the medium-term window is 7 to 14 days, and the time window length of the long-term window is more than 30 days.

[0068] Table 1. Structural Examples of Three-Channel Time Window Timing Models

[0069]

[0070] Step S23: Temporal attention trend modeling, used to perform regression modeling of the indicator trend within a local window. Specifically, based on the three-channel time series feature data, attention trend modeling is performed by designing temporal position encoding and constructing a temporal attention layer, and error optimization is performed by introducing a dynamic residual sensitive regularization term to obtain trend modeling feature data.

[0071] The calculation formula for designing temporal position encoding and constructing a temporal attention layer to model attention trends is as follows:

[0072] ;

[0073] In the formula, z t This is the temporal attention output feature vector, where Attention is the attention layer function, and Q is the query vector, used to represent the input feature vector f of the three-channel temporal feature data. t Temporal position encoding p t The sum of vectors, where K is the key vector and V is the value vector;

[0074] The formula for calculating error optimization by introducing a dynamic residual sensitive regularization term is as follows:

[0075] ;

[0076] In the formula, It is a loss function that introduces a dynamic residual sensitive regularization term, where t is the time index. It is a predicted value, x t It is the actual value. This is the regularization strength parameter, with a default value of 0.1. It is the standard deviation of the residuals;

[0077] Step S24: Global trend fusion, used to fuse the prediction results in all windows, specifically by using the trend modeling feature data and exponentially weighted moving average to perform global trend fusion and smooth reconstruction of fusion features to obtain global trend fusion feature data;

[0078] The calculation formula for the global trend fusion is as follows:

[0079] ;

[0080] In the formula, It is the trend prediction value after weighted fusion. This is the short-term window weight, with a default value of 0.3. This is a short-term window forecast value. This is the weight of the intermediate window, with a default value of 0.4. This is a medium-term window forecast value. This is the long-term window weight, with a default value of 0.3. It is a long-term window forecast value;

[0081] The calculation formula for the smooth reconstruction of the fused features is as follows:

[0082] ;

[0083] In the formula, It is global trend fusion feature data. This is the smoothing weight for fused features, with a default value of 0.2. It is the global trend fusion feature data from the previous moment;

[0084] Step S25: Trend anomaly assessment, used to establish a time evolution model and predict trend reliability. Specifically, based on the global trend fusion feature data, the model is trained to obtain a pathological indicator evolution model, and the pathological indicator evolution trend is predicted by using the pathological indicator evolution model to obtain indicator evolution trend data.

[0085] By performing the above operations, this solution addresses the technical problems of existing pathological indicator evolution modeling methods, which lack dynamic response capabilities, are insufficiently sensitive to fluctuations, and are difficult to adapt to nonlinear temporal behaviors such as indicator mutations, lags, or abnormal rebounds. It creatively adopts a multi-stage pathological evolution modeling method with temporal residual sensitivity to perform pathological indicator evolution modeling, achieving highly robust fitting to abnormal trends and error-driven modeling optimization, thereby enhancing the model's ability to identify nonlinear evolution patterns and its clinical sensitivity.

[0086] Specifically, traditional trend modeling often uses simple methods such as moving averages and linear regression, which are difficult to accurately fit the response pattern when indicators change drastically. For example, during acute infection, the C-reactive protein (CRP) value rises rapidly and then falls back quickly, and linear trends often overestimate the true trend. This invention introduces sliding window modeling, residual regularization adjustment and sensitivity factor weighting mechanism to achieve reinforcement learning for abnormal time points, making the modeling process more "explanatory of disease behavior".

[0087] Example 4, see Figure 1 , Figure 2This embodiment is based on the above embodiment. In step S3, the verification anomaly detection is used to identify pathological indicators that deviate from the normal evolution pattern. Specifically, based on the indicator evolution trend data, residual calculation is performed using the standard error analysis method to obtain indicator prediction residual data. Anomaly judgment threshold is constructed, and anomaly residual judgment is performed based on the indicator prediction residual data to obtain abnormal evolution detection data. Adjacent time periods of abnormal evolution detection data that meet the conditions of continuous deviation or indicator mutation are merged and divided into abnormal evolution detection intervals, labeled, and abnormal evolution detection interval data are obtained.

[0088] Preferably, the anomaly detection threshold is calculated based on the standard deviation and skewness of the indicator evolution trend data, using local statistical features. Then, based on these local statistical features, the anomaly detection threshold is constructed using the Z-score dynamic standardization method. The calculation formula is as follows:

[0089] ;

[0090] In the formula, S t It is the anomaly detection threshold, e t It is the residual value, specifically used to represent the difference between the actual value and the predicted value in the indicator evolution trend data. It is the standard deviation of the residuals. It is to prevent the zero constant. This is the skewness weighting coefficient, with a default value of 0.3. (Skew) t It is a local statistical feature.

[0091] Example 5, see Figure 1 , Figure 2 and Figure 4 This embodiment is based on the above embodiment. In step S4, the pathological deviation analysis is used to identify potential risk factors and deviation patterns behind the fluctuations in indicators. Specifically, based on the abnormal evolution detection interval data, a causal-aware spatiotemporal pathological deviation analysis network is used to perform pathological deviation analysis to obtain deviation feature pattern data, including the following steps:

[0092] Step S41: Deviation interval differential temporal coding, specifically, introducing a temporal differential operator and combining it with cross-attention coding to perform differential coding on the abnormal evolution detection interval data to obtain deviation interval differential temporal coding vector data;

[0093] The time-series difference operator specifically refers to the first-order difference operator;

[0094] Step S42: Construction of pathological offset causal graph, specifically, based on the deviation interval differential temporal coding vector data, constructing a pathological offset causal graph to obtain pathological offset causal graph data;

[0095] Each node of the pathological offset causal graph data is used to represent a key pathological indicator; each edge of the pathological offset causal graph data is used to represent the offset propagation relationship between indicators detected in the abnormal interval; the set of edge weights of the pathological offset causal graph data is used to represent the intensity quantification value of offset propagation.

[0096] Step S43: Offset path subgraph classification extraction, specifically, based on the pathological offset causal graph data, index offset paths are extracted, and the risk types of index offsets are clustered into abnormal types to obtain offset path classification subgraph data;

[0097] The anomaly clustering specifically employs an improved spectral clustering method for clustering calculation, and the calculation formula is as follows:

[0098] ;

[0099] In the formula, Cluster(P) i () represents the offset path classification subgraph data corresponding to the i-th pathological offset causal propagation path, where i is the pathological offset causal propagation path index, k is the cluster number index, and KMeans k It is the clustering result of the eigenvector matrix corresponding to the first k smallest eigenvalues, Eig k Let I be the eigenvector matrix corresponding to the first k smallest eigenvalues, where I is the identity matrix and D is the degree matrix. These are the pathological offset causal propagation paths corresponding to i pathological offset causal propagation paths. It is the pathological offset causal propagation path corresponding to the j-th neighboring pathological offset causal propagation path. This is the Gaussian kernel width, with a default value of 0.5. This is the risk weighting adjustment coefficient, with a default value of 0.7. It is a risk structure similarity factor;

[0100] The formula for calculating the pathological deviation causal propagation path is as follows:

[0101] ;

[0102] In the formula, L i S is the path length, used to represent the total number of nodes in the path. i It is the propagation intensity, C i It is the average propagation centrality, R i It is a path risk index;

[0103] Step S44: Offset risk analysis, specifically, based on the offset path classification sub-map data, the pathological offset pattern of each abnormal interval is analyzed, and offset risk analysis is performed through statistical correlation methods to obtain offset risk tracing data;

[0104] Step S45: Offset feature pattern data generation, specifically, based on the offset risk tracing data, offset feature pattern data is integrated to obtain deviation feature pattern data.

[0105] By performing the above operations, this solution addresses the technical problems in existing pathological deviation analysis methods, which rely solely on static anomaly determination, fail to reconstruct the interaction structure between indicators, and ignore the driving role of abnormal propagation paths and causal variables on deviation patterns. This solution creatively adopts a causal-aware spatiotemporal pathological deviation analysis network to perform pathological deviation analysis, achieving unified modeling of causal mapping of deviation relationships between indicators, abnormal path identification, and risk factor causal association, thereby improving the resolution of pathological abnormality types and the pertinence of clinical review.

[0106] Specifically, for example, in hepatitis cases, elevated alanine aminotransferase (ALT) levels may be directly caused by active viral replication or indirectly by structural damage to the liver. Without causal structural modeling, judging solely from outliers will fail to distinguish between "symptom indicators" and "causative indicators." This invention constructs a weighted directed graph to identify off-pathways and, combined with background factor analysis, can deduce the pathological chain of "elevated viral load; elevated ALT; decreased liver function," improving the explanatory power of etiology and the personalization of countermeasures.

[0107] Example 6, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S5, the pathological intelligent review is used to generate review suggestions by combining the deviation features with the comparison results of historical cases. Specifically, it outputs review suggestions in a structured manner based on the deviation feature pattern data to obtain review abnormality correction reference data.

[0108] Example 7, see Figure 1 and Figure 2 Based on the above embodiments, this embodiment provides a pathological intelligent review system based on case history comparison, including a case archiving module, a pathological modeling module, an abnormality initial detection module, a deviation analysis module, and a pathological review module;

[0109] The case archiving module is used for structured archiving of cases at multiple time points. Through structured archiving of cases at multiple time points, structured time series archive data is obtained, and the structured time series archive data is sent to the pathology modeling module.

[0110] The pathological modeling module is used for pathological indicator evolution modeling. Through pathological indicator evolution modeling, indicator evolution trend data is obtained, and the indicator evolution trend data is sent to the anomaly initial detection module.

[0111] The anomaly initial detection module is used to verify the anomaly detection, obtain the anomaly evolution detection interval data by verifying the anomaly detection, and send the anomaly evolution detection interval data to the deviation analysis module;

[0112] The deviation analysis module is used for pathological deviation analysis. Through pathological deviation analysis, deviation feature pattern data is obtained, and the deviation feature pattern data is sent to the pathological review module.

[0113] The pathology review module is used for intelligent pathology review, and through intelligent pathology review, reference data for correcting review abnormalities is obtained.

[0114] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, 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 process, method, article, or apparatus.

[0115] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0116] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A pathological intelligent review method based on historical case comparison, characterized in that: The method includes the following steps: Step S1: Multi-timepoint case structured archiving to obtain structured time series archive data; Step S2: Pathological indicator evolution modeling. Based on the structured time series archive data, a multi-stage pathological evolution modeling method with temporal residual sensitivity is used to model the evolution of pathological indicators, obtaining indicator evolution trend data. The pathological indicator evolution modeling includes: Screening of key pathological indicators yielded key pathological indicator data. A three-channel time window time series model including a short-term window, a medium-term window, and a long-term window is constructed, and time series features are extracted based on the pathological key indicator data to obtain three-channel time series feature data; the short-term window uses a one-dimensional convolutional kernel to extract short-term features, the medium-term window uses a long short-term neural network to extract periodic patterns, and the long-term window uses linear trend fitting calculation to calculate trend slope features. By designing temporal position encoding and constructing a temporal attention layer, attention trend modeling is performed, and error optimization is carried out by introducing a dynamic residual sensitive regularization term to obtain trend modeling feature data. Global trend fusion feature data is obtained by performing global trend fusion and smooth reconstruction of fusion features through exponential weighted moving average; Based on the global trend fusion feature data, the model is trained to obtain the pathological index evolution model, and the pathological index evolution trend is predicted by using the pathological index evolution model to obtain index evolution trend data. Step S3: Review anomaly detection. Based on the indicator evolution trend data, perform residual calculation using the standard error analysis method to obtain indicator prediction residual data. Then, by constructing anomaly judgment threshold, perform anomaly judgment based on the indicator prediction residual data to obtain anomaly evolution detection data. Next time periods of anomaly evolution detection data that meet the conditions of continuous deviation or indicator abrupt change are merged and divided into anomaly evolution detection intervals, labeled, and anomaly evolution detection interval data are obtained. Step S4: Pathological deviation analysis. Based on the abnormal evolution detection interval data, pathological deviation analysis is performed to obtain deviation feature pattern data; the pathological deviation analysis includes: By introducing a temporal difference operator and combining it with cross-attention coding, differential coding is performed on the abnormal evolution detection interval data to obtain deviation interval differential temporal coding vector data; Based on the aforementioned deviation interval differential temporal coding vector data, pathological offset causal graph data is constructed. Based on the pathological offset causal graph data, the index offset path is extracted and the risk type of the index offset is clustered into anomaly type to obtain the offset path classification subgraph data. Based on the offset path classification subgraph data, offset risk analysis is performed to obtain offset risk tracing data; Based on the aforementioned deviation risk attribution data, deviation feature pattern data is integrated to obtain deviation feature pattern data; Step S5: Intelligent pathological review. Based on the deviation feature pattern data, a structured review suggestion is output to obtain review abnormality correction reference data.

2. The pathological intelligent review method based on case history comparison according to claim 1, characterized in that: The formula for calculating error optimization by introducing a dynamic residual sensitive regularization term is as follows: ； In the formula, It is a loss function that introduces a dynamic residual sensitive regularization term, where t is the time index. It is a predicted value, x t It is the actual value. It is a canonical strength parameter. It is the standard deviation of the residuals.

3. The pathological intelligent review method based on case history comparison according to claim 2, characterized in that: In step S1, the multi-time point case structured archiving is used to archive case information in a structured manner. Specifically, it involves extracting pathological case information from the medical system to obtain raw case information data, and then constructing structured time series archive data through structured processing and time series archiving. The pathological case information includes pathological test reports and pathological image labeling results at historical time points.

4. The pathological intelligent review method based on case history comparison according to claim 3, characterized in that: The anomaly detection threshold is calculated based on the standard deviation and skewness of the indicator evolution trend data, using local statistical characteristics. Then, based on these local statistical characteristics, the anomaly detection threshold is constructed using the Z-score dynamic standardization method. The calculation formula is as follows: ； In the formula, S t It is the anomaly detection threshold, e t It is the residual value, specifically used to represent the difference between the actual value and the predicted value in the indicator evolution trend data. It is the standard deviation of the residuals. It is to prevent the zero constant. These are skewed weighting coefficients, Skew t It is a local statistical feature.

5. The pathological intelligent review method based on case history comparison according to claim 4, characterized in that: Each node of the pathological offset causal graph data is used to represent a key pathological indicator; each edge of the pathological offset causal graph data is used to represent the offset propagation relationship between indicators detected in the abnormal interval; the set of edge weights of the pathological offset causal graph data is used to represent the intensity quantification value of offset propagation.

6. A pathological intelligent review system based on case history comparison, used to implement the pathological intelligent review method based on case history comparison as described in any one of claims 1-5, characterized in that: It includes a case archiving module, a pathology modeling module, an abnormality initial detection module, a deviation analysis module, and a pathology review module; The case archiving module is used for structured archiving of cases at multiple time points. Through structured archiving of cases at multiple time points, structured time series archive data is obtained, and the structured time series archive data is sent to the pathology modeling module. The pathological modeling module is used for pathological indicator evolution modeling. Based on the structured time series archive data, it adopts a multi-stage pathological evolution modeling method with temporal residual sensitive perception to perform pathological indicator evolution modeling and obtain indicator evolution trend data. The pathological indicator evolution modeling includes: Screening of key pathological indicators yielded key pathological indicator data. A three-channel time window time series model including a short-term window, a medium-term window, and a long-term window is constructed, and time series features are extracted based on the pathological key indicator data to obtain three-channel time series feature data; the short-term window uses a one-dimensional convolutional kernel to extract short-term features, the medium-term window uses a long short-term neural network to extract periodic patterns, and the long-term window uses linear trend fitting calculation to calculate trend slope features. By designing temporal position encoding and constructing a temporal attention layer, attention trend modeling is performed, and error optimization is carried out by introducing a dynamic residual sensitive regularization term to obtain trend modeling feature data. Global trend fusion feature data is obtained by performing global trend fusion and smooth reconstruction of fusion features through exponential weighted moving average; Based on the global trend fusion feature data, the model is trained to obtain a pathological indicator evolution model. The pathological indicator evolution trend is predicted by using the pathological indicator evolution model to obtain indicator evolution trend data, and the indicator evolution trend data is sent to the anomaly initial detection module. The anomaly initial detection module is used to verify anomaly detection. Based on the indicator evolution trend data, it performs residual calculation using the standard error analysis method to obtain indicator prediction residual data. It also constructs anomaly judgment thresholds and performs anomaly residual judgment based on the indicator prediction residual data to obtain anomaly evolution detection data. It merges adjacent time periods of anomaly evolution detection data that meet the conditions of continuous deviation or indicator abrupt change into anomaly evolution detection intervals, marks them, and obtains anomaly evolution detection interval data. Finally, it sends the anomaly evolution detection interval data to the deviation analysis module. The deviation analysis module is used for pathological deviation analysis. Based on the abnormal evolution detection interval data, it performs pathological deviation analysis to obtain deviation feature pattern data. The pathological deviation analysis includes: By introducing a temporal difference operator and combining it with cross-attention coding, differential coding is performed on the abnormal evolution detection interval data to obtain deviation interval differential temporal coding vector data; Based on the aforementioned deviation interval differential temporal coding vector data, pathological offset causal graph data is constructed. Based on the pathological offset causal graph data, the index offset path is extracted and the risk type of the index offset is clustered into anomaly type to obtain the offset path classification subgraph data. Based on the offset path classification subgraph data, offset risk analysis is performed to obtain offset risk tracing data; Based on the deviation risk tracing data, the deviation feature pattern data is integrated to obtain the deviation feature pattern data, and the deviation feature pattern data is sent to the pathology review module; The pathology review module is used for intelligent pathology review, and through intelligent pathology review, reference data for correcting review abnormalities is obtained.

Images