Epileptic information management method and system

Abstract

The invention provides an epileptic information management method and system, and relates to the technical field of medical information management. The method comprises the following steps: firstly, performing standardization processing on multi-source heterogeneous data of a patient, and constructing a multi-dimensional patient portrait containing static, dynamic and time sequence features; secondly, performing multi-level clustering on historical patients based on treatment response and disease course evolution similarity; generating a management suggestion according to the association degree of the to-be-managed patient and the clustering cluster and the current management scene; and finally, dynamically updating the portrait and the clustering affiliation based on treatment feedback. According to the invention, by mining time sequence features and establishing a closed loop feedback mechanism, precise classification and dynamic personalized management of epileptics are realized, and the accuracy of clinical treatment recommendation is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of medical information management technology, specifically to a method and system for managing information on epilepsy patients. Background Technology

[0002] The clinical management of epilepsy is characterized by its long-term nature, complexity, dynamism, and individualization: patients require long-term, even lifelong, medication and regular follow-up; there are numerous subtypes of epilepsy with heterogeneous etiologies, leading to significant differences in response to the same treatment regimen among different patients; the patient's condition evolves over time, and seizure patterns and medication responses may change; individual factors such as the patient's lifestyle, adherence, and cognitive abilities also significantly affect management outcomes. These characteristics place high technical demands on epilepsy patient information management systems.

[0003] Existing methods for managing epilepsy patient information have the following technical problems in clinical applications: Current methods typically recommend treatment plans based on the similarity of the patient's condition. This means that treatment plans from patients with similar histories of epilepsy, seizure patterns, and other characteristics are recommended to new patients. However, due to individual differences in drug metabolism, comorbidities, and lifestyle habits, patients with similar conditions may respond significantly differently to the same treatment plan. This results in insufficient accuracy for recommendation methods based on condition similarity, making it difficult to meet the needs of personalized clinical treatment.

[0004] Most existing methods rely on static features for patient analysis and classification, lacking effective utilization of temporal characteristics. Epilepsy exhibits significant temporal characteristics, including periodic seizure patterns, disease progression, and time-dependent treatment responses. This temporal information is crucial for seizure risk early warning and treatment optimization. Current methods fail to effectively mine and utilize this temporal information, limiting the system's predictive and early warning capabilities.

[0005] Information on epilepsy patients comes from diverse sources, including in-hospital electronic medical records, electroencephalogram (EEG) examinations, imaging studies, patient-reported seizure diaries, and wearable device data. Existing methods struggle to effectively integrate this multi-source, heterogeneous data, resulting in incomplete patient profiles and impacting the accuracy and comprehensiveness of management decisions.

[0006] Epilepsy, as a chronic disease, requires long-term management, and the patient's condition continues to change over time. Current methods often employ a one-off recommendation model, lacking a closed-loop feedback mechanism that dynamically updates patient classifications and optimizes recommendation strategies based on treatment feedback, thus failing to achieve continuous improvement and dynamic adjustment of management strategies.

[0007] Furthermore, effective epilepsy management requires the active participation of patients and their families, including regular medication, seizure recording, and avoidance of triggers. Existing systems lack sufficient support for patient self-management and struggle to provide personalized lifestyle guidance and adherence interventions tailored to different management scenarios, thus hindering the improvement of overall management effectiveness. Summary of the Invention

[0008] The purpose of this invention is to provide a method for managing epilepsy patients based on multi-dimensional temporal profiles and multi-level clustering of treatment response similarity. By dynamically matching historical patient data with the current management scenario, personalized treatment suggestions can be generated and the entire disease course can be dynamically optimized in a closed loop.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A method for managing epilepsy patient information includes the following steps: S1: Obtain multi-source heterogeneous data of epilepsy patients, standardize the multi-source heterogeneous data, and construct a unified patient data model; S2: Extract static features, dynamic features, and temporal features from the patient data model, and construct a multidimensional patient profile based on the static features, dynamic features, and temporal features; S3: Calculate the comprehensive similarity between historical patients based on the similarity of their treatment response and the similarity of their disease progression, and perform multi-level clustering of historical patients based on the comprehensive similarity to obtain multiple clusters; S4: Calculate the degree of association between the patient to be managed and each cluster based on the similarity between the multidimensional patient profile of the patient to be managed and the multidimensional patient profile of the historical patients in each cluster; S5: Based on the degree of correlation and the current management scenario, generate management suggestions for the patient to be managed; S6: Collect the treatment feedback data of the patients to be managed, and update the multidimensional patient profile and the cluster to which the patients to be managed belong based on the treatment feedback data.

[0010] Furthermore: the multi-level clustering includes: Based on static features, basic layer clustering of historical patients is performed to obtain multiple basic layer clusters; Within each basic layer cluster, response layer clustering is performed based on treatment response similarity to obtain multiple response layer clusters; Within each response layer cluster, state layer clustering is performed based on the similarity of disease progression to obtain multiple state layer clusters.

[0011] Furthermore: the calculation of the treatment response similarity includes: Segment the treatment records of historical patients to extract keywords related to treatment plans; Calculate the effect description feature value of each keyword in a preset neighborhood, where the effect description feature value is the average of the effect tendency scores of each word segment in the preset neighborhood; The similarity of treatment responses between any two historical patients is calculated based on the similarity of the keyword sets in their treatment records and the similarity of the effect description feature values ​​of the same keywords.

[0012] Furthermore: the calculation of the similarity in the course of disease includes: Obtain the control state sequences of historical patients, and use the dynamic time warping algorithm to calculate the similarity of control state sequences of any two historical patients. Determine whether the disease stages of any two historical patients are the same to obtain the disease stage matching degree; The similarity of disease progression is obtained by weighted summation of the control state sequence similarity and the disease stage matching degree.

[0013] Furthermore: the calculation of the association between the patients to be managed and each cluster includes: Calculate the distribution breadth of each feature dimension of the patients to be managed in the target cluster to obtain the overall distribution breadth; Calculate the cosine similarity between the multidimensional patient profiles of the patient to be managed and each historical patient in the target cluster to obtain the overall similarity. The degree of association between the patients to be managed and the target cluster is obtained by weighted summation of the overall distribution breadth and the overall similarity.

[0014] Furthermore, the generation of management suggestions based on the degree of correlation and the current management scenario includes: Identify the current management scenario, which includes at least one of the following: initial diagnosis scenario, follow-up scenario, medication adjustment scenario, attack scenario, and daily scenario; Select the corresponding recommendation strategy based on the current management scenario; based on the degree of correlation, statistically analyze the treatment plans and effect data of historical patients from the matched clusters, and generate the management suggestions according to the recommendation strategy.

[0015] Furthermore, the method also includes a risk warning step: Obtain current values ​​of multiple risk factors for patients under management, including at least two of the following: changes in seizure frequency, medication adherence, sleep quality, emotional stress, historical seizure cycles, and trigger exposure; Risk scores are calculated based on the current values ​​and preset weights of each risk factor. The corresponding early warning response is triggered based on the risk level to which the risk score belongs.

[0016] Furthermore: the timing features include: Seizure pattern characteristics, including at least one of seizure periodicity, seizure clustering, and seizure trend; Treatment response trajectory characteristics, including at least one of the following: number of medication adjustments, average duration of medication use, and treatment response pattern; The disease progression features include at least one of the following: disease stages, sequence of changes in control status, and exacerbation events.

[0017] Furthermore: the updating of the multidimensional patient profile and cluster of the patient to be managed based on treatment feedback data includes: Dynamic features are updated in real time based on treatment feedback data, and time-series features are updated periodically. The cluster affiliation of the state layer is reassessed after each follow-up visit, the cluster affiliation of the response layer is reassessed after changes in the medication regimen or at a preset cycle, and the cluster affiliation of the basic layer is reassessed when diagnostic information changes. The parameters of the clustering model are periodically optimized based on accumulated feedback data.

[0018] The present invention also provides an epilepsy patient information management system, comprising: The data acquisition module is used to acquire multi-source heterogeneous data from epilepsy patients; The data processing module is used to standardize the multi-source heterogeneous data and construct a unified patient data model; The feature extraction module is used to extract static features, dynamic features and time-series features from the patient data model to construct a multi-dimensional patient profile. The clustering analysis module is used to calculate a comprehensive similarity based on the similarity of treatment response and disease course evolution of historical patients, and to perform multi-level clustering of historical patients based on the comprehensive similarity to obtain multiple clusters; The association matching module is used to calculate the degree of association between the patient to be managed and each cluster based on the similarity between the multidimensional patient profile of the patient to be managed and the multidimensional patient profile of the historical patients in each cluster. The recommendation generation module is used to generate management suggestions for patients to be managed based on the degree of correlation and the current management scenario; The feedback update module is used to collect treatment feedback data and update the multidimensional patient profile and the cluster to which the patient to be managed belongs based on the treatment feedback data.

[0019] Compared with the prior art, the present invention has the following advantages: I. This invention clusters patients based on the similarity of treatment response and the similarity of disease progression, grouping patients with "similar treatment needs" into one category, rather than classifying them solely based on disease characteristics. This effectively eliminates the interference of individual differences on treatment plan recommendations and significantly improves the accuracy and clinical applicability of recommendations.

[0020] Second, this invention adopts a multi-level dynamic clustering mechanism, including a nested clustering structure of a basic layer, a response layer, and a state layer, which enables patients to migrate between clusters at different levels as the disease progresses, realizing dynamic adjustment of management strategies and adapting to the long-term management needs of epilepsy as a chronic disease.

[0021] Third, this invention constructs a multi-dimensional patient profile by extracting static features, dynamic features and temporal features, effectively integrating multi-source heterogeneous data such as in-hospital electronic medical records, examination and test data, and patient self-reported data, and fully mining temporal information such as attack patterns, treatment response trajectories, and disease progression, providing data support for attack risk warning and trend prediction.

[0022] Fourth, when calculating the correlation between patients to be managed and clusters, this invention comprehensively considers two dimensions: the similarity of patient profiles and the breadth of feature distribution in clusters, making the matching results more accurate and reliable, and avoiding the bias that may be caused by a single similarity index.

[0023] Fifth, this invention establishes a closed-loop optimization mechanism based on treatment feedback data. By continuously collecting treatment effect feedback, it dynamically updates patient profiles and cluster affiliations, and optimizes model parameters, forming a virtuous cycle of continuously improving management effectiveness, thus ensuring the long-term effectiveness and adaptability of the system. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating one embodiment of the epilepsy patient information management method of the present invention; Figure 2 This is a schematic diagram of the architecture of the epilepsy patient information management system of the present invention in another embodiment. Detailed Implementation

[0025] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] This invention provides a method for managing epilepsy patient information. This method achieves precise, personalized, and dynamic full-cycle management support for epilepsy patients through steps such as multi-source data collection and standardized processing, multi-dimensional feature extraction and patient profile construction, multi-level dynamic clustering based on treatment response patterns, dynamic association matching between patients to be managed and clusters, scenario-based intelligent recommendation and risk warning, and closed-loop feedback and dynamic model optimization. It includes the following steps: S1: Obtain multi-source heterogeneous data of epilepsy patients, standardize the multi-source heterogeneous data, and construct a unified patient data model.

[0027] Specifically, in the multi-source data acquisition and standardization processing steps, the first step is to acquire multi-source data from epilepsy patients. This multi-source data includes in-hospital electronic medical record data, examination and testing data, patient-reported data, and wearable device data. The in-hospital electronic medical record data includes the patient's basic information, diagnostic information, medical history, and consultation records. Basic information includes name, gender, age, and contact information; diagnostic information includes epilepsy type, epilepsy syndrome, etiological classification, and comorbidities; medical history information includes age of onset, course of disease, family history, and past medical history; and consultation records include outpatient records, inpatient records, and discharge summaries. The examination and testing data includes electroencephalogram (EEG) data, imaging data, and laboratory test data. EEG data includes structured reports and raw signals from routine EEG, video EEG, and ambulatory EEG; imaging data includes structured reports and imaging features from cranial MRI and CT scans; and laboratory test data includes blood drug concentration monitoring, liver and kidney function tests, and complete blood counts. Patient-reported data includes seizure diaries, medication records, lifestyle records, and self-assessment scales. Seizure diaries record seizure time, type, duration, triggers, and prodromal symptoms. Medication records record medication administration times, missed doses, and adverse reactions. Lifestyle records record sleep patterns, emotional state, and any special events. Self-assessment scales include quality of life scales and anxiety / depression scales. Wearable device data includes activity levels and physiological data. Activity data covers activity levels and sleep quality, while physiological data includes heart rate variability.

[0028] The collected multi-source data underwent standardization processing. Standardization included data cleaning, terminology standardization, and text structuring. Data cleaning included handling missing values, outliers, and duplicate data. Missing data was filled using the mean, mode, or marked as missing, depending on the data type. Clearly unreasonable data records were identified and processed, and duplicate records were identified and merged. Terminology standardization involved establishing a standard terminology database for epilepsy, mapping non-standard terms to standard terms. This database included terms for epileptic seizure types based on the ILAE 2017 classification, standard names of antiepileptic drugs, and standard terms for EEG descriptions. Text structuring used medical natural language processing techniques to extract structured information from unstructured medical record text, employing named entity recognition technology to extract key information such as drug names, dosages, and frequencies.

[0029] After standardization, a unified patient data model is constructed. The patient data model includes patient ID, basic information, diagnostic information, seizure information, medication information, examination information, and follow-up records. Basic information includes fields such as gender, date of birth, age of onset, and duration of illness (in months). Diagnostic information includes fields such as epilepsy type, epilepsy syndrome, etiological classification, and a list of comorbidities. The epilepsy type ranges from generalized to focal, generalized with focal, and unknown. The etiological classification ranges from hereditary to structural, metabolic, immune, infectious, and unknown. Seizure information is a list of seizure records, each containing fields such as seizure time, seizure type, duration, trigger, prodromal symptoms, and post-seizure status. Medication information is a list of medication records, each containing fields such as drug name, start date, end date, dosage, frequency, reason for adjustment, efficacy assessment, and adverse reactions. The efficacy assessment ranges from complete control to significant reduction, partial reduction, ineffectiveness, and exacerbation. Examination information includes a list of electroencephalogram (EEG) records and a list of imaging records. The follow-up record is a list of follow-up information. Each follow-up record includes fields such as follow-up date, frequency of attacks, control status, compliance assessment, and treatment recommendations. The control status ranges from complete control, basic control, partial control, to no control, and the compliance assessment ranges from good, fair, to poor.

[0030] S2: Extract static features, dynamic features, and temporal features from the patient data model, and construct a multidimensional patient profile based on the static features, dynamic features, and temporal features.

[0031] Specifically, static characteristics refer to patient characteristics that remain relatively stable over a relatively long period. Static characteristics include demographic characteristics, diagnostic characteristics, baseline examination characteristics, and medical history characteristics. Demographic characteristics include sex, age of onset, and current age, where sex is represented by binary codes, and age of onset and current age are represented numerically. Diagnostic characteristics include epilepsy type, etiological classification, and epilepsy syndrome, where epilepsy type and etiological classification are represented by unique thermal codes, and epilepsy syndrome is represented by embedded vectors. Baseline examination characteristics include the type of abnormality on the first EEG and imaging abnormalities, where the type of abnormality on the first EEG is represented by multi-label codes, and imaging abnormalities are represented by binary codes indicating the presence or absence of structural abnormalities. Medical history characteristics include family history, history of febrile seizures, and number of comorbidities, where family history and history of febrile seizures are represented by binary codes, and the number of comorbidities is represented numerically.

[0032] Dynamic characteristics refer to the patient's state characteristics that change over time and need to be updated regularly. Dynamic characteristics include seizure characteristics, medication characteristics, control status characteristics, and adherence characteristics. Seizure characteristics are calculated based on recent seizure records and include seizure frequency, seizure type distribution, primary triggers, and seizure severity. The formula for calculating seizure frequency is:

[0033] Seizure type distribution is represented by a vector showing the proportion of each seizure type. The primary triggers are a list of the most frequently occurring triggers. Seizure severity is a weighted score based on duration and type. Medication characteristics include the current medication regimen, dosage, and highest dose percentage. The current medication regimen is a coded representation of the drug combination, the dosage is the number of antiepileptic drugs currently used, and the highest dose percentage is the ratio of the current dose to the recommended highest dose for each drug. Control status characteristics include control level and seizure-free days. Control level is assigned values ​​of 4, 3, 2, and 1 for complete control, basic control, partial control, and no control, respectively. Seizure-free days are the number of days since the most recent seizure. Adherence characteristics include medication adherence score and follow-up adherence. The formula for calculating the medication adherence score is:

[0034] Follow-up compliance refers to the rate of timely follow-up visits.

[0035] Temporal characteristics are the results of time-series analysis of patients' historical data. These characteristics include seizure pattern characteristics, treatment response trajectory characteristics, and disease progression characteristics. Seizure pattern characteristics include seizure periodicity, seizure clustering, and seizure trend. Seizure periodicity is identified through autocorrelation analysis to determine the periodicity of seizures; seizure clustering indicates whether seizures tend to occur in clusters; and the formula for calculating seizure trend is:

[0036] in e The smoothing parameter is set to 0.1 to avoid a denominator of zero. Treatment response trajectory characteristics include the number of medication adjustments, average duration of medication use, and treatment response pattern. The number of medication adjustments is the number of times the historical medication regimen has changed; the average duration of medication use is the average number of months used for each medication; and the treatment response pattern is an encoding of the response to different medication categories. Disease progression characteristics include disease stages, control status change sequences, and exacerbation events. Disease stages are divided into new diagnosis, adjustment, stable, and refractory phases based on the duration of illness and control status. The specific division rules are: if the duration of illness is less than 12 months, it is a new diagnosis phase; if medication adjustments have been made in the past 6 months, it is an adjustment phase; if complete control has been achieved in the past 12 months, it is a stable phase; if two or more suitable medications have been used and control is still not achieved, it is a refractory phase; otherwise, it is an adjustment phase. The control status change sequence is the control status sequence of the most recent follow-up visits; exacerbation events indicate whether there has been a significant deterioration in control status.

[0037] A multidimensional patient profile is constructed for each patient based on the three types of features mentioned above. The patient profile vector is represented as follows:

[0038] in These are static feature vectors with dimension m; It is a dynamic feature vector with dimension n; This is a time-series feature vector with dimension k. The features of each dimension are normalized to unify their value range to the interval between 0 and 1.

[0039] S3: Calculate the comprehensive similarity between historical patients based on the similarity of their treatment response and the similarity of their disease progression, and perform multi-level clustering of historical patients based on the comprehensive similarity to obtain multiple clusters.

[0040] Specifically, in the multi-level dynamic clustering step based on treatment response patterns, the similarity of treatment responses for historical patients is first calculated. The treatment records of historical patients are then segmented using a word segmentation tool combined with a medical dictionary. These records include medication regimen descriptions, efficacy assessment records, and follow-up records. Keywords are extracted from the treatment records using an epilepsy-specific thesaurus. These keywords include drug keywords, efficacy keywords, and adverse reaction keywords. Drug keywords are the names of antiepileptic drugs, efficacy keywords are descriptive terms, and adverse reaction keywords are descriptive terms.

[0041] For each keyword, its effect description feature value in the treatment record is calculated. The average effect tendency score of all words in the preset neighborhood of each keyword in the target treatment record is used as the effect description feature value of that keyword. The preset neighborhood consists of 5 words before and after the keyword. The effect tendency score is obtained through a predefined effect dictionary, with positive effect words scoring positive, negative effect words scoring negative, and neutral words scoring 0. The formula for calculating the effect description feature value is:

[0042] Treatment response similarity is calculated based on keyword set similarity and effect description similarity. Keyword set similarity is calculated using the Jaccard similarity coefficient, and the formula is as follows:

[0043] in and Patients i and j The set of keywords for treatment records. The formula for calculating the similarity of effect descriptions is:

[0044] in and Patients i and j Chinese keywords k The effect describes the feature value.N This refers to the number of identical keywords. The formula for calculating treatment response similarity is:

[0045] in α and β For weight parameters, ,default α The value is 0.4. β The value is 0.6.

[0046] Then, the similarity of disease progression among historical patients is calculated. The control state sequences from the most recent follow-up visits for each patient are obtained, with the six most recent visits used by default. The control states are encoded as sequences. The dynamic time warping algorithm is used to calculate the similarity of the control state sequences between two patients. The formula for calculating the similarity of control state sequences is:

[0047] in and Patients i and j The control state sequence. The disease stage matching degree is determined based on the patient's disease stage; if two patients are at the same disease stage, the matching degree is 1; otherwise, it is 0.5. The formula for calculating the similarity of disease progression is:

[0048] in c and d For weight parameters, ,default c The value is 0.7. d The value is 0.3.

[0049] A comprehensive similarity score is calculated based on treatment response similarity and disease progression similarity. The formula for calculating the comprehensive similarity score is:

[0050]

[0051] Among them, the image similarity ( i,j The cosine similarity between the multidimensional patient profile vectors of patients i and j is given by ( ). , , For weight parameters, ,default The value is 0.5. The value is 0.3. The value is 0.2.

[0052] Multi-level clustering of historical patients is performed based on comprehensive similarity, resulting in multiple clusters. This multi-level clustering includes basic layer clustering, response layer clustering, and state layer clustering. Basic layer clustering is based on static features, forming relatively stable basic groups. The K-means algorithm is used, with a distance metric of 1 minus the static feature similarity. The number of clusters is determined using the silhouette coefficient method. Response layer clustering performs secondary clustering within each basic layer cluster based on treatment response similarity, using a hierarchical clustering algorithm. The distance metric is 1 minus the treatment response similarity, and the number of clusters is determined by dendrogram segmentation. State layer clustering performs tertiary clustering within each response layer cluster based on disease progression similarity, using the K-means algorithm. The distance metric is 1 minus the disease progression similarity, and the number of clusters is set to 2 to 3. The final clustering structure is a three-layer nested structure, with each basic layer cluster containing several response layer clusters, and each response layer cluster containing several state layer clusters.

[0053] S4: Calculate the degree of association between the patient to be managed and each cluster based on the similarity between the multidimensional patient profile of the patient to be managed and the multidimensional patient profile of the historical patients in each cluster.

[0054] Specifically, the cluster to which a patient belongs is dynamically updated as the disease progresses. The basic layer clusters are updated when there are significant changes in diagnostic information, the response layer clusters are reassessed quarterly or after changes in medication regimens, and the status layer clusters are updated after each follow-up based on the latest status.

[0055] In the dynamic association matching step between patients to be managed and their clusters, a hierarchical matching strategy is adopted. First, basic layer matching is performed, determining the basic layer cluster to which each patient belongs based on their static characteristics, and assigning each patient to the basic layer cluster with the highest similarity to their static characteristics. Next, response layer matching is performed, calculating the association degree between each patient and each response layer cluster within the determined basic layer clusters. Finally, state layer matching is performed, further matching state layer clusters within the response layer clusters with the highest association degree.

[0056] The calculation of correlation degree comprehensively considers two dimensions: feature distribution breadth and profile similarity. For each feature dimension in the profile of the patient to be managed, the distribution breadth of that feature value in the target cluster is calculated. The formula for calculating the distribution breadth is:

[0057] The similarity feature value is defined as a feature value difference less than a preset threshold θ. The formula for calculating the overall distribution breadth of patients to be managed is:

[0058] Calculate the multidimensional similarity between the patient to be managed and each historical patient in the target cluster. The image similarity is calculated using cosine similarity, and the formula is as follows:

[0059] The formula for calculating the overall similarity with the target cluster is:

[0060] The formula for calculating the degree of correlation is:

[0061] in m and n For weight parameters, ,default m The value is 0.4. n The value is 0.6. The normalize() function is a normalization function that maps the result to the interval between 0 and 1.

[0062] When calculating the degree of correlation, different weights were assigned to features across different dimensions. Epilepsy type, as a core diagnostic feature, had a weight factor of 1.5. Etiological classification, as an important pathological feature, had a weight factor of 1.3. Seizure characteristics, as key clinical manifestations, had a weight factor of 1.2. Medication characteristics, related to treatment decisions, had a weight factor of 1.4. Control status, as an indicator of current condition, had a weight factor of 1.3. Demographic characteristics, as basic reference information, had a weight factor of 0.8.

[0063] S5: Based on the degree of correlation and the current management scenario, generate management suggestions for the patient to be managed.

[0064] In the scenario-based intelligent recommendation and risk warning process, management suggestions are generated based on the current management scenario. These scenarios include initial diagnosis, follow-up, medication adjustment, seizure, and daily scenarios. The initial diagnosis scenario is triggered when a newly diagnosed epilepsy patient seeks medical attention. The recommendation content is an initial treatment plan suggestion. Based on the patient's baseline and response layer clustering matching results, the initial medication plans and their effectiveness rates of historical patients in the matched clusters are statistically analyzed, and the recommended plans are sorted from high to low effectiveness rates for reference. The follow-up scenario is triggered when the patient has a regular check-up. The recommendation content includes an assessment of the control status, suggestions for plan adjustment, and the next follow-up time. By comparing the patient's current status with the previous follow-up status, the effectiveness of the current plan is evaluated, and the subsequent evolution of patients with similar statuses is referenced based on the status layer clustering matching. The medication adjustment scenario is triggered when a medication plan adjustment is needed. The recommendation content includes alternative drug plans and dosage adjustment suggestions. By analyzing the patient's past medication history and response, patients with similar medication histories are screened in the response layer clusters. The subsequent medication adjustments and effects of these patients are statistically analyzed, and the adjustment plan with the highest success rate is recommended. The seizure scenario is triggered when a patient reports a new seizure. Recommended content includes seizure management guidance, advice on whether medical attention is needed, assessment of severity based on the reported seizure characteristics, and provision of corresponding management suggestions compared to the patient's past seizure patterns. The daily scenario is triggered during the patient's daily self-management. Recommended content includes lifestyle management suggestions, medication reminders, and health knowledge. Personalized lifestyle precautions are pushed based on the patient's trigger analysis, medication reminders are set according to the medication regimen, and relevant health knowledge is pushed based on the patient's concerns.

[0065] Seizure risk warning is based on the temporal characteristics of patients under management and pre-defined risk factors. Risk factors include increased recent seizure frequency, decreased medication adherence, decreased sleep quality, increased emotional stress, proximity to historical seizure cycles, and exposure to known triggers. The weights of each risk factor are 0.25, 0.20, 0.15, 0.10, 0.15, and 0.15, respectively. The risk score is calculated using the following formula:

[0066] Each risk factor has a score range of 0 to 100. The risk level is determined based on the risk score, triggering a corresponding warning response. A risk score between 0 and 30 indicates low risk, with no warning triggered; a risk score between 31 and 60 indicates medium risk, with a notification and advice sent to the patient; a risk score between 61 and 80 indicates high risk, with a warning sent to the patient and prevention recommendations pushed; a risk score between 81 and 100 indicates extremely high risk, with an emergency warning sent to the patient and family members notified, and medical attention recommended.

[0067] S6: Collect the treatment feedback data of the patients to be managed, and update the multidimensional patient profile and the cluster to which the patients to be managed belong based on the treatment feedback data.

[0068] Specifically, in the closed-loop feedback and model dynamic optimization steps, treatment feedback data is collected from patients under management. This data includes feedback on treatment effectiveness, recommendation adoption, and early warning accuracy. Treatment effectiveness feedback is collected at each follow-up visit or 1 to 3 months after medication adjustments, and includes information on seizure control, adverse reactions, and patient satisfaction. Recommendation adoption feedback is collected after the recommended treatment plan is presented, and includes whether the physician adopted the recommendation and the reasons for non-adoption. Early warning accuracy feedback is collected during the observation period after an early warning is issued, and includes whether a seizure actually occurred after the warning.

[0069] Patient profiles and cluster assignments for managed patients are updated based on treatment feedback data. Dynamic features are updated in real time whenever new data becomes available, including epidural characteristics, medication characteristics, disease control status, and adherence characteristics. Time-series features are updated in batches weekly, including epidural trends, treatment response trajectories, and disease progression characteristics. The cluster assignment update strategy is as follows: the status layer is reassessed after each follow-up visit; the response layer is reassessed after changes in medication regimens or quarterly; and the baseline layer is reassessed when diagnostic information changes.

[0070] The model parameters are optimized based on accumulated treatment feedback data. Similarity weight optimization uses Bayesian optimization to adjust the weight parameters in the overall similarity calculation, aiming to maximize the actual effectiveness of the recommended treatment plan. This optimization is performed monthly. Risk warning threshold optimization adjusts the warning threshold based on warning accuracy feedback, aiming to maximize precision while maintaining a recall rate greater than 80%. This optimization is performed quarterly. Cluster structure optimization periodically re-evaluates the cluster structure based on accumulated patient data. Optimization includes the number of clusters and cluster centers, based on intra-cluster consistency and inter-cluster discriminability. This optimization is performed semi-annually.

[0071] The technical solution of the present invention will be further described below through specific embodiments.

[0072] In the first embodiment, an initial treatment plan was recommended for a newly diagnosed epilepsy patient. Patient Zhang, male, 25 years old, was newly diagnosed with generalized tonic-clonic seizures of unknown etiology and without significant comorbidities. First, data collection and profiling were performed, including collecting basic patient information, diagnostic information, and initial EEG results to construct a multidimensional patient profile. Then, cluster matching was performed. Baseline matching categorized the patient into the generalized seizures-unknown etiology cluster. Response layer matching calculated the association between the patient and each response layer cluster under the baseline layer. The results showed the highest association (0.78) with the cluster showing a good response to sodium valproate monotherapy. Finally, a treatment plan was recommended. Initial treatment plans for 42 historical patients in the matched clusters were analyzed. Sodium valproate monotherapy was used in 28 cases (75% efficacy), and levetiracetam monotherapy was used in 10 cases (70% efficacy). Sodium valproate monotherapy was recommended as the first-line treatment, with a recommended starting dose of 500 mg daily, gradually increasing to 1000 mg daily. Regular monitoring of liver function and blood drug concentration was also advised.

[0073] In the second embodiment, medication adjustment recommendations were made for patients with treatment-resistant epilepsy. Patient Li, female, 35 years old, with focal epilepsy, had been treated with carbamazepine and lamotrigine combination therapy for one year, but still experienced 2 to 3 seizures per month. First, a profile update was performed, determining the disease stage as treatment-resistant, with a poor response to both sodium channel blockers. Then, cluster matching was performed: response-level matching categorized the patient into the poor sodium channel blocker response cluster, while state-level matching categorized her into the partially controlled-breakthrough cluster. Finally, treatment recommendations were made, and the subsequent adjustment plans for 18 historical patients in the cluster were analyzed. Among them, levetiracetam was added in 8 cases with an improvement rate of 62%, zonisamide was added in 5 cases with an improvement rate of 60%, and sodium valproate was replaced as the basic regimen in 5 cases with an improvement rate of 40%. The assessment results showed that the current combination of the two sodium channel blockers was not well controlled. The recommended adjustment direction was to consider adding a non-sodium channel mechanism drug, with levetiracetam as the first choice and zonisamide as an alternative. It was also suggested to continue to keep an episodic diary and have a follow-up evaluation one month after medication adjustment.

[0074] In the third embodiment, dynamic early warning of seizure risk is implemented for patients in the stable phase. Patient Wang, whose condition has been stable for one year, has experienced poor sleep for the past week and missed a medication dose yesterday. Risk factor assessment is performed: Increased recent seizure frequency score is 0, decreased medication adherence score is 60, decreased sleep quality score is 70, increased emotional stress score is 40, proximity to historical seizure cycles score is 20, and known trigger exposure score is 50. Risk score is calculated: Risk score = (0×0.25 + 60×0.20 + 70×0.15 + 40×0.10 + 20×0.15 + 50×0.15) / 1.0 = (0 + 12 + 10.5 + 4 + 3 + 7.5) = 37 Based on the risk score, the risk level is determined to be medium risk, triggering a notification on the patient's end. The notification message states that there are recent risk factors for an attack, and patients should pay attention to ensuring adequate sleep and taking medication on time, while also sending prevention suggestions.

[0075] This invention also provides an epilepsy patient information management system, which includes a data acquisition module, a data processing module, a clustering analysis module, an association matching module, a recommendation generation module, and a feedback optimization module. The data acquisition module is used to acquire multi-source data of epilepsy patients, including an in-hospital data interface unit, an examination data access unit, a patient self-report collection unit, and an equipment data access unit. The data processing module is used to standardize the multi-source data and construct a patient data model, extract multi-dimensional features from the patient data model, and construct a patient profile, including a data cleaning unit, a standardization processing unit, a feature engineering unit, and a data storage unit. The clustering analysis module is used to calculate a comprehensive similarity based on the similarity of historical patients' treatment responses and disease progression, and to perform multi-level clustering of historical patients based on the comprehensive similarity to obtain multiple clusters. The association matching module is used to determine the degree of association between the patient to be managed and each cluster based on the similarity between the patient profile of the patient to be managed and the patient profiles of historical patients in each cluster, as well as the distribution breadth of the features of the patient to be managed in each cluster. The recommendation generation module generates management suggestions based on the degree of correlation and the current management scenario. It also includes a risk assessment unit to calculate risk scores and trigger corresponding early warning responses based on risk levels. The feedback optimization module collects treatment feedback data from patients under management and updates their patient profiles and cluster affiliations based on this data. It also optimizes model parameters based on accumulated feedback data.

[0076] The modules mentioned above interact with each other through data interfaces. The data acquisition module transmits the collected multi-source data to the data processing module. The data processing module transmits the constructed patient profile to the cluster analysis module and the association matching module. The cluster analysis module transmits the clustering results to the association matching module. The association matching module transmits the association matching results to the recommendation generation module. The management suggestions generated by the recommendation generation module are provided to the doctor's workstation, patient application, and family application through the application service interface. The feedback optimization module collects feedback data from the application service interface and feeds back the optimization results to the data processing module, the cluster analysis module, and the recommendation generation module, forming a complete closed-loop management process.

[0077] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent transformations or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for managing information on epilepsy patients, characterized in that, Includes the following steps: S1: Obtain multi-source heterogeneous data of epilepsy patients, standardize the multi-source heterogeneous data, and construct a unified patient data model; S2: Extract static features, dynamic features, and temporal features from the patient data model, and construct a multidimensional patient profile based on the static features, dynamic features, and temporal features; S3: Calculate the comprehensive similarity between historical patients based on the similarity of their treatment response and the similarity of their disease progression, and perform multi-level clustering of historical patients based on the comprehensive similarity to obtain multiple clusters; S4: Calculate the degree of association between the patient to be managed and each cluster based on the similarity between the multidimensional patient profile of the patient to be managed and the multidimensional patient profile of the historical patients in each cluster; S5: Based on the degree of correlation and the current management scenario, generate management suggestions for the patient to be managed; S6: Collect the treatment feedback data of the patients to be managed, and update the multidimensional patient profile and the cluster to which the patients to be managed belong based on the treatment feedback data.

2. The method of claim 1, wherein, The multi-level clustering includes: Based on static features, basic layer clustering of historical patients is performed to obtain multiple basic layer clusters; Within each basic layer cluster, response layer clustering is performed based on treatment response similarity to obtain multiple response layer clusters; Within each response layer cluster, state layer clustering is performed based on the similarity of disease progression to obtain multiple state layer clusters.

3. The method according to claim 1, characterized in that, The calculation of the treatment response similarity includes: Segment the treatment records of historical patients to extract keywords related to treatment plans; Calculate the effect description feature value of each keyword in a preset neighborhood, where the effect description feature value is the average of the effect tendency scores of each word segment in the preset neighborhood; The similarity of treatment responses between any two historical patients is calculated based on the similarity of the keyword sets in their treatment records and the similarity of the effect description feature values ​​of the same keywords.

4. The method according to claim 1, characterized in that, The calculation of the similarity in disease progression includes: Obtain the control state sequences of historical patients, and use the dynamic time warping algorithm to calculate the similarity of control state sequences of any two historical patients. Determine whether the disease stages of any two historical patients are the same to obtain the disease stage matching degree; The similarity of disease progression is obtained by weighted summation of the control state sequence similarity and the disease stage matching degree.

5. The method according to claim 1, characterized in that, Calculating the association between the patients to be managed and each cluster includes: Calculate the distribution breadth of each feature dimension of the patients to be managed in the target cluster to obtain the overall distribution breadth; Calculate the cosine similarity between the multidimensional patient profiles of the patient to be managed and each historical patient in the target cluster to obtain the overall similarity. The degree of association between the patients to be managed and the target cluster is obtained by weighted summation of the overall distribution breadth and the overall similarity.

6. The method according to claim 1, characterized in that, Step S5 specifically includes: Identify the current management scenario, which includes at least one of the following: initial diagnosis scenario, follow-up scenario, medication adjustment scenario, attack scenario, and daily scenario; Select the corresponding recommendation strategy based on the current management scenario; Based on the degree of correlation, historical patient treatment plans and effect data are statistically analyzed from the matched clusters, and management suggestions are generated according to the recommendation strategy.

7. The method according to claim 1, characterized in that, The method also includes a risk warning step: Obtain current values ​​of multiple risk factors for patients under management, including at least two of the following: changes in seizure frequency, medication adherence, sleep quality, emotional stress, historical seizure cycles, and trigger exposure; Risk scores are calculated based on the current values ​​and preset weights of each risk factor. The corresponding early warning response is triggered based on the risk level to which the risk score belongs.

8. The method according to claim 1, characterized in that, The timing features include: Seizure pattern characteristics, including at least one of seizure periodicity, seizure clustering, and seizure trend; Treatment response trajectory characteristics, including at least one of the following: number of medication adjustments, average duration of medication use, and treatment response pattern; The disease progression features include at least one of the following: disease stages, sequence of changes in control status, and exacerbation events.

9. The method according to claim 1, characterized in that, The multidimensional patient profile and cluster of the patients to be managed are updated based on the treatment feedback data, including: Dynamic features are updated in real time based on treatment feedback data, and time-series features are updated periodically. The cluster affiliation of the state layer is reassessed after each follow-up visit, the cluster affiliation of the response layer is reassessed after changes in the medication regimen or at a preset cycle, and the cluster affiliation of the basic layer is reassessed when diagnostic information changes. The parameters of the clustering model are periodically optimized based on accumulated feedback data.

10. An epilepsy patient information management system, characterized in that, include: The data acquisition module is used to acquire multi-source heterogeneous data from epilepsy patients; The data processing module is used to standardize the multi-source heterogeneous data and construct a unified patient data model; The feature extraction module is used to extract static features, dynamic features and time-series features from the patient data model to construct a multi-dimensional patient profile. The clustering analysis module is used to calculate a comprehensive similarity based on the similarity of treatment response and disease course evolution of historical patients, and to perform multi-level clustering of historical patients based on the comprehensive similarity to obtain multiple clusters; The association matching module is used to calculate the degree of association between the patient to be managed and each cluster based on the similarity between the multidimensional patient profile of the patient to be managed and the multidimensional patient profile of the historical patients in each cluster. The recommendation generation module is used to generate management suggestions for patients to be managed based on the degree of correlation and the current management scenario; The feedback update module is used to collect treatment feedback data and update the multidimensional patient profile and the cluster to which the patient to be managed belongs based on the treatment feedback data.

Images