Medical digital case intelligent management method
Through customized disease templates and multimodal data preprocessing, refined case information management is achieved, solving the problem of inefficient patient screening in traditional clinical projects, improving the intelligence and adaptability of data management, and ensuring data accuracy and the system's rapid response capabilities.
Patent Information
- Application Number
- CN202510733806.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-04
- Publication Date
- 2025-09-19
AI Technical Summary
Patient screening in traditional clinical projects relies on doctors' experience and judgment, which is inefficient and prone to missing eligible patients. Existing technologies lack a systematic case information management mechanism, resulting in delays in clinical trial progress and limited patients' access to appropriate treatment.
Adopting the intelligent management method of medical digital cases, through customized disease templates, the key point data of patient cases are automatically extracted and structured and stored in the corresponding diagnosis and treatment nodes. Combined with multimodal data preprocessing and temporal logic verification, refined data management is achieved.
It improves patient screening efficiency, ensures data accuracy and system adaptability, enhances the intelligence level and overall effectiveness of medical management, and supports rapid response to new diseases or examination items.
Smart Images
Figure CN120673954A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical management technology, and more particularly to an intelligent management method for medical digital cases. Background Art
[0002] With the continuous development of modern medicine, clinical research plays an increasingly important role in new drug development, treatment optimization, and disease prevention. However, traditional clinical program recruitment often relies on physicians' experience and manual screening, which is inefficient and prone to missing eligible patients. This delays clinical trial progress and, to a certain extent, limits patients' access to cutting-edge treatments suitable for their conditions.
[0003] In recent years, the application of artificial intelligence and big data technologies in the medical field has steadily deepened, achieving remarkable results, particularly in electronic medical record management, intelligent assisted diagnosis, and personalized recommendations. By establishing a structured and standardized case information processing mechanism, data utilization efficiency can be effectively improved and provide a foundation for subsequent intelligent analysis and matching. However, existing technologies still have significant shortcomings in connecting information between patients and clinical medical projects, particularly the lack of a systematic, disease-specific intelligent management mechanism.
[0004] Therefore, how to find an efficient case information management method for matching clinical medical project information is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention
[0005] In light of this, the present invention provides an intelligent management method for digital medical records. This method customizes disease templates based on different disease types. Each template contains multiple diagnosis and treatment nodes, organized according to the medical diagnostic process. It automatically extracts key point data from patient records and stores it in a structured manner within the corresponding diagnosis and treatment nodes. This refined data management approach not only helps improve patient screening efficiency for clinical projects but also facilitates subsequent data tracking and analysis.
[0006] In order to achieve the above object, the present invention adopts the following technical solutions:
[0007] A method for intelligent management of digital medical cases, comprising the following steps:
[0008] S1: Receive the patient's multimodal original medical record data, preprocess the multimodal original medical record data, and output a structured data packet with a timestamp;
[0009] S2: extracting key point information based on the preprocessed structured data packet, performing similarity matching between the key point information and a predefined disease template library, and determining the disease template;
[0010] S3: Verify the temporal logic rules based on the key point information and historical key point information, assign the key point information to the corresponding diagnosis and treatment node in the disease template and store it.
[0011] Preferably, the preprocessing of the multimodal original medical record data in S1 includes the image processing sub-step:
[0012] Noise modeling is performed on the scanned image, and noise reduction is performed using a bilateral filter, wherein the parameters of the bilateral filter are dynamically adjusted to obtain a denoised image;
[0013] Use deep learning segmentation models to locate key areas in medical record images and segment them to obtain key area coordinate sets and segmented area image blocks;
[0014] The text content is extracted by OCR through multi-scale feature fusion, and the errors of the OCR extracted text content are corrected by combining with the medical knowledge base.
[0015] Preferably, the step of locating key areas in the medical record image using a deep learning segmentation model includes:
[0016] The denoised image is subjected to dimensionality reduction through multi-layer convolution operations, and the dimensionality reduction features undergo two layers of parallel linear transformation to generate projection matrices η1 and η2 respectively;
[0017] The projection matrix η1 is feature mapped using state space dual transformation and discretized;
[0018] The discretization result is subjected to depthwise separable convolution and fused with the dimensionality reduction feature to generate the intermediate feature η3;
[0019] Fuse the dimensionality reduction feature with the intermediate feature η3 to obtain the fused feature η that represents the global dependency relationship;
[0020] The dimensionality reduction features and fusion features are mixed separately and then summed up to obtain the detail features of the denoised image;
[0021] Perform feature extraction on detail features and output classification results.
[0022] Preferably, the key areas include: patient information, chief complaint, current medical history, physical examination, auxiliary examination, preliminary diagnosis or medical advice.
[0023] Preferably, the preprocessing of the multimodal original medical record data in S1 includes the speech processing sub-step:
[0024] Separate segments of patient oral recordings, use a speech recognition model fine-tuned in the medical field, use Mel spectrograms for speech feature extraction, and perform dialect phoneme embedding;
[0025] Based on the Whisper-large basic model architecture, a medical term conversion matrix is constructed;
[0026] The top-K candidate standard terms are screened through the transformation matrix M, and a multi-factor score is calculated based on the phoneme matching degree, symptom association degree and regional weight. Multiple candidate terms are screened according to the score to obtain semantic features.
[0027] Preferably, the step of matching disease templates according to key point information in S2 includes: calculating similarity between the key point information and a predefined disease template library, and selecting the template with the highest matching degree as the current disease template.
[0028] Preferably, the S2 further comprises the step of expanding the dynamic disease template:
[0029] If the data package contains new location information that does not exist in the disease template library, the template expansion process will be automatically triggered;
[0030] Extract the newly added point information, find the most similar key point information in the existing disease template, and establish the association relationship.
[0031] Generate a new disease template and output the remaining key point information bound to the disease template as the physical sign dataset in the current new disease template.
[0032] Preferably, the diagnosis and treatment nodes in S3 include: diagnosis nodes, inspection nodes, treatment nodes and inspection nodes, which respectively correspond to different diagnosis and treatment time stages of the patient and are used to store and record corresponding key point information.
[0033] Preferably, the step of sequential logic verification in S3 includes:
[0034] Obtain the current key point information C, historical key point sequence H and disease model T:
[0035] C={(t c ,E c ,P c )}
[0036] H=[(t1,E1,P1),...,(t n ,E n ,P n )]
[0037] T={N i →R ij}
[0038] Where t is the timestamp, E is the key point type, P is the key point medical parameter, i, j∈N, N is the number of diagnosis and treatment nodes, and R is the timing rule;
[0039] The timing constraint relationship between the current key point information C and the historical key point sequence H is extracted from the disease template. The target diagnosis and treatment node to which the current key point information should be assigned is determined based on the timing constraint relationship, and all key point information in the original medical record data is stored as the vital sign data of the target diagnosis and treatment node.
[0040] Preferably, the timing constraint relationship includes:
[0041] Strong order relation:
[0042]
[0043] Among them, Δ tol is the tolerance interval, E i is the current key point information, E j It is the historical key point information;
[0044] Mutually exclusive relationship:
[0045]
[0046] Among them, Δ safe For E i to E j Safety interval for timestamps.
[0047] It can be seen from the above technical solutions that, compared with the prior art, the present invention provides a method for intelligent management of digital medical cases, which has the following beneficial effects:
[0048] Improve patient screening efficiency: Through customized disease templates, each template contains multiple diagnosis and treatment nodes divided according to the medical diagnostic process. It can automatically extract key point data from patient records and store them in a structured manner in the corresponding diagnosis and treatment nodes. This refined data management method not only helps improve patient screening efficiency for clinical projects, but also facilitates subsequent data tracking and analysis.
[0049] The present invention preprocesses and structures multimodal original medical record data to make the data more standardized and systematized, thereby improving the basic support capabilities for subsequent intelligent analysis and matching.
[0050] When a data packet contains new location information that is not already in the disease template library, the present invention can automatically trigger the template expansion process to generate a new disease template. This approach ensures the system's rapid response to new diseases or examination items, enhancing the system's adaptability and flexibility.
[0051] This invention aligns multi-source data during the data preprocessing phase, automatically identifying and marking inconsistencies between data from different sources, ensuring data consistency and accuracy. Furthermore, during the allocation of diagnosis and treatment nodes, logical verification is performed using temporal constraints such as strong order relationships and mutual exclusion relationships, further ensuring the scientific nature and rationality of the diagnosis and treatment process.
[0052] In summary, the present invention not only effectively solves the inefficiency problem of traditional clinical project recruitment relying on doctors' experience and manual screening, but also greatly improves the intelligence level and overall efficiency of medical management by building a systematic and disease-specific intelligent management mechanism. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.
[0054] Figure 1 A flowchart of a medical digital case intelligent management method provided by the present invention;
[0055] Figure 2 A flowchart of the image processing sub-steps provided by the present invention;
[0056] Figure 3 A flowchart of the steps for locating key areas in medical record images using the deep learning segmentation model provided by the present invention;
[0057] Figure 4 This is a flowchart of the speech processing sub-steps provided by the present invention. DETAILED DESCRIPTION
[0058] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0059] The embodiment of the present invention discloses a method for intelligent management of medical digital cases. Figure 1 As shown, the process includes four main steps: multimodal original medical record data collection and preprocessing, case structuring and disease template matching, intelligent allocation and time sequence arrangement of diagnosis and treatment nodes, and data quality review and multidimensional storage. The following is the specific implementation process:
[0060] Multimodal original medical record data collection and preprocessing stage:
[0061] S1: Receive the original medical record data input by the user. The original medical record data format includes electronic medical record text, case scan images / test order images, and voice recording data, and pre-process the original medical record data to output structured field data with timestamps.
[0062] In one embodiment, S1 includes the image data preprocessing steps:
[0063] S111: Removes noise from scanned images and enhances the contrast of text areas using medical image processing algorithms.
[0064] For the original scanned medical record image I(x,y), the noise characteristics are first analyzed:
[0065] Calculate the local noise intensity by wavelet transform: D HH is the horizontal-high frequency subband coefficient.
[0066] Construct an improved bilateral filter to filter and obtain the denoised image I denoised (x,y):
[0067]
[0068] Where:
[0069] Spatial weight
[0070] Luminance Weight
[0071] Dynamic adjustment parameter σ s =1.5σ n ,σ r =2.0σ n .
[0072] S112: Use the deep learning segmentation model to automatically select core areas such as patient information, chief complaint, current medical history, and test results in the denoised image.
[0073] S1121: The denoised image is subjected to dimensionality reduction through multi-layer convolution operations. The dimensionality reduction features are subjected to two layers of parallel linear transformation to generate projection matrices η1 and η2 respectively.
[0074] S1122: Use state space dual transformation to perform feature mapping on the projection matrix η1 and perform discretization processing.
[0075] S1123: Perform depth-wise separable convolution on the discretization result and fuse it with the dimensionality reduction feature to generate the intermediate feature η3;
[0076] S1124: Fuse the dimension reduction feature with the intermediate feature η3 to obtain a fused feature η that represents the global dependency relationship.
[0077] S1125: Dimensionality reduction features and fusion features are mixed in channels respectively and then summed up to obtain detail features of the denoised image.
[0078] S1126: Extract the detail features and output the classification results.
[0079] It should be noted that the dataset labels used by the deep learning segmentation model provided in this embodiment during the training phase identify classification areas according to the standard case record format. For example, the standard paper medical record template mainly includes the following area contents:
[0080] [Patient information], [Chief complaint], [Current medical history], [Physical examination], [Auxiliary examination], [Preliminary diagnosis], [Doctor's instructions].
[0081] The identification areas and categories are shown in Table 1. The output of the trained deep learning segmentation model for segmentation is the key area coordinate set {B k |k∈C} and segmented region image blocks {I Bk}, C is the total number of categories.
[0082] Table 1 Examples of training data identification areas and categories
[0083]
[0084]
[0085] S113: Multi-scale feature fusion OCR extracts text content and automatically corrects recognition errors based on the medical knowledge base.
[0086] S1131: Obtain the original OCR output word ω using multi-scale feature fusion OCR extraction OCR .
[0087] S1132: OCR original output word ω OCR With dictionary word ω i The feature similarity of ∈D is defined as:
[0088] Sim(ω OCR ,ω i )=cos(ω OCR ,ω i )
[0089] By minimizing the feature similarity, it is determined whether the original OCR output word ω OCR Make corrections.
[0090] S114: Timestamp extraction: extract matching date formats (such as "2023-11-05") and associate test reports, medical orders, etc. with specific time nodes.
[0091] In one embodiment, S1 includes the following steps of pre-processing the speech data:
[0092] S121: Separate the effective segments in the patient's oral recording, filter the environmental noise, use the Mel spectrogram to extract speech features, and perform dialect phoneme embedding (select the dialect phoneme set based on regional information).
[0093] S122: Based on the architecture of the Whisper-large medical model, a two-layer fine-tuning strategy is adopted.
[0094] The first level fine-tunes general dialect recognition (covering seven major Chinese dialect areas), and the second level fine-tunes medical terminology by converting colloquial descriptions into standardized expressions (focusing on optimizing dialect expressions for symptom descriptions).
[0095] Construct a medical term conversion matrix:
[0096] M = [dialect expression × standard term × (α·symptom relevance + β·regional weight + γ·timeliness factor)]
[0097] Among them, the symptom correlation is calculated based on clinical guidelines (such as the pathological matching degree of "heart pain" and "chest pain"); the regional weight is based on the probability of using the dialect in the patient's permanent residence (such as the weight of region A = 0.93); the timeliness factor is determined by tracking the frequency of term updates (such as new pneumonia-related terms have a higher timeliness weight).
[0098] The top-K candidate standard terms are screened through the transformation matrix M, and multi-factor scores are calculated based on phoneme matching, symptom association and regional weight, and multiple candidate terms are screened based on the scores.
[0099] S123: Store the semantic features output by the medical version of Whisper-large basic model in text format.
[0100] In one embodiment, S1 includes the steps of aligning multi-source data:
[0101] S131: Map the examination time in the image, the doctor's order time in the voice record, and the diagnosis time in the text medical record to the same timeline.
[0102] S132: Automatically identify inconsistencies in data from different sources (e.g., discrepancies between imaging reports and textual diagnostic conclusions) and mark abnormal items that require manual review.
[0103] S133: Output a structured data packet with a timestamp and source tag (including text, numerical value, and time information).
[0104] Case structuring and disease template matching stage:
[0105] S2: Extract key point information from the preprocessed structured field data packet, perform disease template matching based on the key point information, and store the key point information in the corresponding disease template.
[0106] In one embodiment, the key point information in S2 includes the corrected text content extracted by multi-scale feature fusion OCR and the semantic features output by the medical version of Whisper-large basic model.
[0107] In one embodiment, the step of matching disease templates based on key point information includes: calculating similarity between the key point information and a predefined disease template library (such as diabetes and coronary heart disease templates), and selecting the template with the highest matching degree as the current disease template.
[0108] It should be noted that the predefined disease template library stores point information corresponding to each disease, and the disease template is determined by calculating the matching coverage between the key point information and the point information in the disease template library.
[0109] In one embodiment, the step of expanding the dynamic disease template is further included:
[0110] If the data package contains new site information that does not exist in the disease template library (such as new gene detection projects), the template expansion process will be automatically triggered:
[0111] Extract the newly added point information, find the most similar key point information (such as "conventional genetic testing") in the existing disease template, and establish an association relationship.
[0112] Generate a new disease template for expert review and then store it in the database, and output the remaining key point information bound to the disease template as the physical sign dataset in the current new disease template.
[0113] It should be noted that the key point information generated by the original medical record data uploaded at the same time node is bound to each other.
[0114] Intelligent allocation and timing arrangement of diagnosis and treatment nodes:
[0115] S3: Perform sequential logic verification based on the current key point information and the historical key point information of the current patient in the disease template, and assign them to the corresponding diagnosis and treatment nodes in the disease template based on the verification results.
[0116] In one embodiment, the diagnosis and treatment node types include: diagnosis nodes, examination nodes, treatment nodes and inspection nodes, which correspond to different diagnosis and treatment time stages of the patient, and are used to store and record corresponding key point information to assist in the overall observation of the patient's disease progression.
[0117] In one embodiment, the specific execution steps of S3 include:
[0118] S31: Obtain current key point information C, historical key point sequence H and disease model T:
[0119] C={(t c ,E c ,P c )}
[0120] H=[(t1,E1,P1),...,(t n ,E n ,P n )]
[0121] T={N i →R ij}
[0122] Where t is the timestamp, E is the key point type, P is the key point medical parameter, i, j∈N, N is the number of diagnosis and treatment nodes, and R is the timing rule.
[0123] S32: Extract three types of constraints from the disease template:
[0124] Strong order relation:
[0125]
[0126] Among them, Δ tol For example, "biopsy" must be performed before "chemotherapy plan formulation", and the tolerance interval Δ tol =24h). If the tolerance interval is met, the chemotherapy regimen will be recorded after the diagnosis and treatment node of the pathological biopsy.
[0127] Mutually exclusive relationship:
[0128]
[0129] Among them, Δ safe For E i to E j Safety interval of timestamps, for example, "enhanced CT" and "iodine contrast agent allergy test" need to be separated by Δ safe =48h. If the safety interval is not reached, an alarm will be issued to remind the entry of the current key point information.
[0130] According to the above constraint relationship, the target diagnosis and treatment node to which the current key point information should be assigned is determined, and all key point information in the original medical record data is stored as the vital sign data of the target diagnosis and treatment node.
[0131] In this embodiment, the priority of strong order relationship constraint is greater than that of weak order relationship constraint. It can be understood that strong order relationship is a must-satisfy item, weak order relationship is a recommended satisfaction item, and mutually exclusive relationship is a conditional satisfaction item.
[0132] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.
[0133] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for intelligent management of medical digital cases, characterized in that: The steps include: S1: Receive the patient's multimodal original medical record data, preprocess the multimodal original medical record data, and output a structured data packet with a timestamp; S2: extracting key point information based on the preprocessed structured data packet, performing similarity matching between the key point information and a predefined disease template library, and determining the disease template; S3: Verify the temporal logic rules based on the key point information and historical key point information, assign the key point information to the corresponding diagnosis and treatment node in the disease template and store it.
2. A medical digital case intelligent management method according to claim 1, characterized in that: The preprocessing of multimodal raw medical record data in S1 includes the following image processing substeps: Noise modeling is performed on the scanned image, and noise reduction is performed using a bilateral filter, wherein the parameters of the bilateral filter are dynamically adjusted to obtain a denoised image; Use deep learning segmentation models to locate key areas in medical record images and segment them to obtain key area coordinate sets and segmented area image blocks; The text content is extracted by OCR through multi-scale feature fusion, and the errors of the OCR extracted text content are corrected by combining with the medical knowledge base.
3. The intelligent management method for medical digital cases according to claim 2, characterized in that: The step of locating the key areas in the medical record image using the deep learning segmentation model includes: The denoised image is subjected to dimensionality reduction through multi-layer convolution operations, and the dimensionality reduction features undergo two layers of parallel linear transformation to generate projection matrices η1 and η2 respectively; The projection matrix η1 is feature mapped using state space dual transformation and discretized; The discretization result is subjected to depthwise separable convolution and fused with the dimensionality reduction feature to generate the intermediate feature η3; Fuse the dimensionality reduction feature with the intermediate feature η3 to obtain the fused feature η that represents the global dependency relationship; The dimensionality reduction features and fusion features are mixed separately and then summed up to obtain the detail features of the denoised image; Perform feature extraction on detail features and output classification results.
4. The intelligent management method for digital medical cases according to claim 2, characterized in that: The key areas include: patient information, chief complaint, current medical history, physical examination, auxiliary examination, preliminary diagnosis or medical advice.
5. The intelligent management method for digital medical cases according to claim 1, characterized in that: The preprocessing of multimodal raw medical record data in S1 includes the following sub-steps: Separate segments of patient oral recordings, use a speech recognition model fine-tuned in the medical field, use Mel spectrograms for speech feature extraction, and perform dialect phoneme embedding; Based on the Whisper-large basic model architecture, a medical term conversion matrix is constructed; The top-K candidate standard terms are screened through the transformation matrix M, and a multi-factor score is calculated based on the phoneme matching degree, symptom association degree and regional weight. Multiple candidate terms are screened according to the score to obtain semantic features.
6. The intelligent management method for digital medical cases according to claim 1, characterized in that: The step of matching disease templates according to key point information in S2 includes: calculating similarity between the key point information and a predefined disease template library, and selecting the template with the highest matching degree as the current disease template.
7. The intelligent management method for digital medical cases according to claim 1, characterized in that: The S2 also includes the following steps to expand the dynamic disease template: If the data package contains new location information that does not exist in the disease template library, the template expansion process will be automatically triggered; Extract the newly added point information, find the most similar key point information in the existing disease template, and establish the association relationship. Generate a new disease template and output the remaining key point information bound to the disease template as the physical sign dataset in the current new disease template.
8. The intelligent management method for digital medical cases according to claim 1, characterized in that: The diagnosis and treatment nodes in S3 include: diagnosis nodes, examination nodes, treatment nodes and inspection nodes, which correspond to different diagnosis and treatment time stages of patients respectively and are used to store and record corresponding key point information.
9. The intelligent management method for digital medical cases according to claim 1, characterized in that: The steps for sequential logic verification in S3 include: Obtain the current key point information C, historical key point sequence H and disease model T: C={(t c ,E c ,P c )} H=[(t1,E1,P1),...,(t n ,E n ,P n )] T={N i →R ij } Where t is the timestamp, E is the key point type, P is the key point medical parameter, i, j∈N, N is the number of diagnosis and treatment nodes, and R is the timing rule; The timing constraint relationship between the current key point information C and the historical key point sequence H is extracted from the disease template. The target diagnosis and treatment node to which the current key point information should be assigned is determined based on the timing constraint relationship, and all key point information in the original medical record data is stored as the vital sign data of the target diagnosis and treatment node.
10. The intelligent management method for digital medical cases according to claim 9, characterized in that: The timing constraint relationship includes: Strong order relation: Among them, Δ tol is the tolerance interval, E i is the current key point information, E j It is the historical key point information; Mutually exclusive relationship: Among them, Δ safe For E i to E j Safety interval for timestamps.