Electronic medical record data patient information desensitization and inspection data standardization method

Abstract

The invention provides an electronic medical record data patient information desensitization and inspection data standardization method, and relates to the technical field of medical big data processing and privacy protection. The method comprises the following steps: collecting structured and unstructured electronic medical record data from a hospital information system; carrying out automatic identification and desensitization processing on the sensitive information of the patient; performing time alignment, missing value filling and normalization processing on laboratory inspection data; key features are extracted, a prediction model is constructed, and the adaptive capacity of the model in different medical policy environments is improved in combination with policy stage division and feature importance analysis. And the interpretability and prediction performance of the model are enhanced through multi-model fusion modeling and an SHAP interpretation mechanism. According to the method, efficient and compliant data desensitization is realized, the consistency of test data and modeling quality are improved, dynamic response policy changes are supported, and the method has good practicability and popularization value.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data processing, and in particular to a method for desensitizing patient information in electronic medical record data and standardizing test data. Background Art

[0002] With the continuous advancement of medical informatization, electronic medical record (EMR) systems have become the core data platform for medical institutions at all levels. EMRs contain a large amount of structured and unstructured clinical information, such as basic patient information, medical records, laboratory test results, and imaging reports. They are a vital data source for clinical decision support, disease prediction modeling, and medical quality assessment.

[0003] Against the backdrop of the increasing application of medical big data analysis and artificial intelligence, how to utilize this data efficiently and securely has become a research hotspot. Currently, various methods have been used to desensitize and preprocess electronic medical record data to meet privacy protection regulations and improve data availability. However, existing technologies still have the following significant shortcomings: existing desensitization methods often use simple field deletion, replacement, or fuzzification, lacking the ability to effectively identify sensitive information in unstructured text. For example, only structured fields such as name and ID number are pseudo-anonymized, while ignoring the personally identifiable information (PII) that may be contained in the free text of doctors' written notes and consultation opinions, resulting in the risk of data leakage after desensitization. In addition, some desensitization processes destroy the semantic integrity of the original data, affecting subsequent modeling and use.

[0004] Laboratory test data collected across different hospitals and time periods often suffer from inconsistent units, inconsistent test item naming, and varying reference value ranges. Direct use of these data can lead to model bias. Existing processing methods often rely on manual cleaning or simple normalization, making it difficult to achieve data consistency across institutions and timescales, limiting the application of large-scale joint modeling.

[0005] In practical applications, laboratory test data often contain missing data. Traditional methods often use methods such as deleting missing samples or filling in the mean, which can easily lead to information loss or introduce bias, affecting model prediction performance. In particular, in time series modeling tasks, improper handling of time series misalignment or missing data can severely impact the model's temporal logic.

[0006] Adjustments to healthcare policies often lead to shifts in clinical behavior patterns, which in turn cause data distribution drift. Existing modeling methods generally ignore this external factor, resulting in a decrease in the model's adaptability and unstable predictions after policy changes.

[0007] Most existing methods rely on empirical screening during the feature engineering phase, lacking unified standards and automated processes. Furthermore, the model training process lacks transparency, lacking in-depth analysis and visualization of key feature groups, making it difficult to ensure the credibility and widespread value of clinical decision-making systems. To address this issue, we propose a method for desensitizing patient information from electronic medical record data and standardizing test data. Summary of the Invention

[0008] The purpose of the present invention is to address the problems raised by the existing background technology. In order to achieve the above-mentioned invention purpose, the present invention provides the following technical solutions: a method for desensitizing patient information in electronic medical record data and standardizing test data, comprising the following steps:

[0009] Step 1: Patient information desensitization step: Collect patient data from the electronic medical record system, identify sensitive information fields, and use multiple desensitization algorithms such as replacement, masking, and truncation to process sensitive information;

[0010] Step 2: Standardization of test data: Align test data to hospitalization records based on test date. Extract key variables from the previous hospitalization. Compare the importance scores of demographics, LE and LMR, and laboratory test results across the three policy phases and assess statistical significance using the Wilcoxon test. Establish a unified standard format and coding rules for test data.

[0011] Step 3: Data modeling and classification: Use multiple classification models to model and analyze the processed data, use the training data set to train the model, use the test data set to evaluate the model performance and select the optimal model.

[0012] As a preferred technical solution of the present invention, in the patient information desensitization step, the collected patient data includes demographic information, LE and LMR information, and laboratory test results.

[0013] As a preferred technical solution of the present invention, in the patient information desensitization step, the identified sensitive information fields include the patient's name, ID number, and contact information.

[0014] As a preferred technical solution of the present invention, in the patient information desensitization step, the desensitization algorithm used is specifically:

[0015] For patient names, a replacement method was used to replace the real names with randomly generated anonymous names;

[0016] For ID card numbers, a masking method is used to retain only some key digits and replace the rest with characters;

[0017] For contact information, the truncation method is adopted to retain only part of the number segment and hide the key part.

[0018] As a preferred technical solution of the present invention, in the test data standardization step, formulating a unified test data standard format and coding rules includes:

[0019] Standardize the names of inspection items and use unified international or domestic inspection item codes;

[0020] Standardize the numerical range of test results and use unified units and precision;

[0021] Standardize the format of inspection reports and standardize the content and layout of reports.

[0022] As a preferred technical solution of the present invention, the classification model used in the data modeling and classification steps includes:

[0023] Logistic regression (scikit-learn v1.2.2) with solver = 'lbfgs', penalty = 'l2', C = 1.0, max_iter = 1000, random_state = 42.

[0024] Decision tree classifier (scikit-learn v1.2.2) with criterion = 'gini', max_depth = None, min_samples_split = 2, random_state = 42;

[0025] Random Forest Classifier (scikit-learn v1.2.2) with parameters set to n_estimators = 100, criterion = 'gini', max_depth = None, min_samples_split = 2, random_state = 42, n_jobs = -1.

[0026] As a preferred technical solution of the present invention, the classification model used in the data modeling and classification steps also includes:

[0027] A classifier based on HistGradientBoosting (HistGradientBoostingClassifier, scikit-learn v1.2.2) with parameters set to max_iter=100, learning_rate=0.1, l2_regularization=0.0, max_bins=255, class_weight='balanced', random_state=42;

[0028] XGBoost classifier (xgboost v1.7.6) with parameters set to n_estimators = 100, max_depth = 6;

[0029] learning_rate = 0.1;

[0030] objective='binary:logistic;

[0031] use_label_encoder=False, eval_metric='logloss', random_state=42;

[0032] LightGBM classifier (lightgbm v3.3.2), with parameter set to n_estimators=100,

[0033] max_depth=-1;

[0034] learning_rate=0.1, objective='binary', class_weight='balanced', boosting_type='gbdt', verbose=-1, random_state=42, n_jobs=-1.

[0035] As a preferred technical solution of the present invention, the classification model used in the data modeling and classification steps also includes an LSTM / GRU classifier (PyTorch), with parameters set as input_dim=feature dimension, hidden_size=64, num_layers=2, output_size=1, learning_rate=0.001, batch_size=64, num_epochs=10, and task='binary'.

[0036] As a preferred technical solution of the present invention, in the test data normalization step, statistical significance is evaluated using the Wilcoxon test, and the annotation rules are: ns: p>0.05; p≤0.05; p≤0.01; p≤0.001; p≤0.0001.

[0037] As a preferred technical solution of the present invention, in the data modeling and classification step, the processed electronic medical record data and test data are divided into a training data set and a test data set, with the training data set accounting for 70% and the test data set accounting for 30%.

[0038] Compared with the existing technology, the present invention has the following advantages: the present invention combines structured field hash encryption with unstructured text natural language recognition to accurately identify and desensitize patient sensitive information, while protecting patient privacy and retaining the key information required for clinical semantics and modeling.

[0039] This method matches laboratory test results to hospitalization records, avoiding modeling bias caused by time series misalignment. It also employs missing value filling, normalization, and encoding conversion standardization to effectively eliminate heterogeneity in data collected across hospitals and time periods, improving cross-institutional data integration capabilities.

[0040] This paper improves model stability and interpretability by integrating a multi-model fusion strategy (including logistic regression, random forest, XGBoost, LightGBM, and LSTM) with feature importance assessment and the SHAP interpretability tool. Experiments show that data processed using this method achieves an average AUC of over 0.83 in the readmission prediction task, significantly outperforming direct modeling using the raw data.

[0041] This method introduces policy stage division and Wilcoxon rank sum test mechanism to quantitatively evaluate the changes in the importance of each feature group under different policy stages. It can identify the high responsiveness of laboratory examination features to policy adjustments (p<0.01), thereby assisting the model to adjust the input feature weights in a timely manner and improving the model's adaptability to changes in the policy environment.

[0042] The present invention can be seamlessly integrated into hospital information systems (HIS), electronic medical record systems (EMR) or regional medical big data platforms, supports batch processing and real-time streaming updates, adapts to large-scale, high-concurrency data processing needs, and has good engineering implementation capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A data frame diagram provided by the present invention;

[0044] Figure 2 A data frame diagram provided by the present invention;

[0045] Figure 3 A data frame diagram provided by the present invention;

[0046] Figure 4 This is a data frame diagram provided by the present invention. DETAILED DESCRIPTION

[0047] To make the purpose, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them.

[0048] Therefore, the following detailed description of the embodiments of the present invention is not intended to limit the scope of the invention claimed for protection, but merely represents some embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features and technical solutions in the embodiments can be combined with each other. It should be noted that similar numbers and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0049] Example 1: A method for desensitizing patient information in electronic medical record data and standardizing test data, comprising the following steps: Step 1, patient information desensitization step: collecting patient data in the electronic medical record system, identifying sensitive information fields therein, and processing the sensitive information using a variety of desensitization algorithms such as replacement method, masking method, and truncation method;

[0050] Step 2: Standardization of test data: Align test data to hospitalization records based on test date. Extract key variables from the previous hospitalization. Compare the importance scores of demographics, LE and LMR, and laboratory test results across the three policy phases and assess statistical significance using the Wilcoxon test. Establish a unified standard format and coding rules for test data.

[0051] Step 3: Data modeling and classification: Use multiple classification models to model and analyze the processed data, use the training data set to train the model, use the test data set to evaluate the model performance and select the optimal model.

[0052] In the patient information desensitization step, the collected patient data include demographic information, LE and LMR information, and laboratory test results.

[0053] In the patient information desensitization step, the sensitive information fields identified include the patient's name, ID number, and contact information.

[0054] In the patient information desensitization step, the desensitization algorithm used is as follows:

[0055] For patient names, a replacement method was used to replace the real names with randomly generated anonymous names;

[0056] For ID card numbers, a masking method is used to retain only some key digits and replace the rest with characters;

[0057] For contact information, the truncation method is adopted to retain only part of the number segment and hide the key part.

[0058] In the inspection data standardization step, the unified inspection data standard format and coding rules are formulated, including:

[0059] Standardize the names of inspection items and use unified international or domestic inspection item codes;

[0060] Standardize the numerical range of test results and use unified units and precision;

[0061] Standardize the format of inspection reports and standardize the content and layout of reports.

[0062] The classification models used in the data modeling and classification steps include:

[0063] Logistic regression (scikit-learn v1.2.2) with solver = 'lbfgs', penalty = 'l2', C = 1.0, max_iter = 1000, random_state = 42.

[0064] Decision tree classifier (scikit-learn v1.2.2) with criterion = 'gini', max_depth = None, min_samples_split = 2, random_state = 42;

[0065] Random Forest Classifier (scikit-learn v1.2.2) with parameters set to n_estimators = 100, criterion = 'gini', max_depth = None, min_samples_split = 2, random_state = 42, n_jobs = -1.

[0066] The classification models used in the data modeling and classification steps also include:

[0067] A classifier based on HistGradientBoosting (HistGradientBoostingClassifier, scikit-learn v1.2.2) with parameters set to max_iter=100, learning_rate=0.1, l2_regularization=0.0, max_bins=255, class_weight='balanced', random_state=42;

[0068] XGBoost classifier (xgboost v1.7.6) with parameters set to n_estimators = 100, max_depth = 6;

[0069] learning_rate = 0.1;

[0070] objective='binary:logistic';

[0071] use_label_encoder=False, eval_metric='logloss', random_state=42;

[0072] LightGBM classifier (lightgbm v3.3.2) with parameters set to n_estimators = 100, max_depth = -1;

[0073] learning_rate=0.1, objective='binary', class_weight='balanced', boosting_type='gbdt', verbose=-1, random_state=42, n_jobs=-1.

[0074] The classification model used in the data modeling and classification steps also includes an LSTM / GRU classifier (PyTorch), with parameters set to input_dim = feature dimension, hidden_size = 64, num_layers = 2, output_size = 1, learning_rate = 0.001, batch_size = 64, num_epochs = 10, and task = 'binary'.

[0075] In the test data normalization step, statistical significance was assessed using the Wilcoxon test, with the following annotation rules: ns: p>0.05; p≤0.05; p≤0.01; p≤0.001; p≤0.0001.

[0076] In the data modeling and classification steps, the processed electronic medical record data and test data are divided into training data sets and test data sets, with the training data set accounting for 70% and the test data set accounting for 30%.

[0077] Working Principle of the Method for Desensitizing Patient Information in Electronic Medical Records and Standardizing Laboratory Data: This method addresses two key issues in electronic medical record data processing: protecting patient information privacy and effectively integrating and utilizing laboratory data. Its core approach is to first desensitize patient information to prevent privacy leaks, then standardize laboratory data to eliminate discrepancies between data sources, and finally utilize multiple classification models to unlock the value of this data. The entire process follows a conventional workflow from data collection to data cleaning, processing, and analysis, achieving secure and efficient utilization of electronic medical record data through a series of orderly steps.

[0078] Electronic medical record systems accumulate a large amount of data containing various patient information, such as demographic information (name, age, and gender), LE and LMR information, and laboratory test results. This data is the basis for subsequent processing and is extracted from the electronic medical record system through data interfaces to prepare for further information identification and desensitization.

[0079] Based on an understanding of the EMR data structure and privacy protection requirements, sensitive information fields are identified. These fields typically contain information that can directly or indirectly identify the patient, such as the patient's name, ID number, and contact information. This identification process can be achieved through pre-set field name matching rules or data type determination.

[0080] In order to ensure that patient privacy is not leaked, a variety of desensitization algorithms are used to process sensitive information:

[0081] Replacement method: Patient names are replaced with randomly generated anonymous names. This method replaces real names with meaningless common names, making it impossible to directly link the data to specific patients while preserving semantics.

[0082] Truncation method: For contact information, only a portion of the number segment is retained (such as the first three digits representing the operator and region), hiding the key parts. This method can prevent others from directly contacting the patient through the contact information, while retaining certain data characteristics for preliminary classification and screening.

[0083] Linking test data with hospitalization records is crucial for medical analysis. Aligning test data to hospitalization records based on test date establishes a temporal and business logic connection between the data. By comparing test dates with admission and discharge times in hospitalization records, the best matching hospitalization record is found and the relevant test data is linked to it. Key variables are extracted from previous hospitalizations, which may include the patient's baseline health indicators and previous medical diagnoses, providing more comprehensive information for subsequent data analysis.

[0084] The importance scores of demographics, left leukemia (LE) and left leukemia (LMR) data, and laboratory test results were compared across the three policy phases. Different policy environments may influence medical behavior and data. By comparing the importance of data across different phases, we can identify patterns in data changes with policy. Laboratory test results were most sensitive to policy changes, suggesting that policy adjustments may directly affect the selection and frequency of test items. The left leukemia (LE) and left leukemia (LMR) feature groups were ranked as the most important overall, demonstrating the centrality of these features in medical analysis. Statistical significance was assessed using the Wilcoxon test. This nonparametric test is independent of the data distribution and determines whether significant differences exist by comparing the rank sums of two data sets. Test results were annotated according to their p-values. Different annotation rules (ns: p>0.05; p≤0.05; p≤0.01; p≤0.001; p≤0.0001) can intuitively reflect the significance of data differences.

[0085] Because test data formats and standards vary across medical institutions and testing equipment, standardization is necessary: ​​Standardization of test item names: Develop a unified test item coding system to uniformly code the same test items with different names used by different medical institutions. For example, the different expressions "CBC-WBC" and "CBC-WBC" could be uniformly coded as "CBC-WBC." This eliminates data confusion caused by name differences and facilitates data integration and comparative analysis.

[0086] Standardize the numerical range of test results: Different testing devices may use different units and precisions to record test results, so the numerical range needs to be standardized. For example, convert blood glucose values ​​from "mg / dL" to "mmol / L" and specify a uniform precision (such as retaining two decimal places). By converting units and standardizing precision, test results from different sources can be compared.

[0087] Standardize the format of test reports: Standardize the content and layout of test reports, unifying report titles, fonts, and table formats. This improves data readability and processability, facilitating subsequent data extraction and analysis. For example, a test report must include basic information such as the test item name, test results, and reference range, arranged in a fixed format.

[0088] In order to mine the potential information in electronic medical record data, a variety of classification models are used for modeling analysis:

[0089] Logistic regression: Based on the principles of linear regression, this method uses a logistic function to map the output of linear regression to the interval [0, 1] for binary classification problems. In this method, solver = 'lbfgs' is set to optimize model parameters, penalty = 'l2' is used for regularization to prevent overfitting, C = 1.0 controls the regularization strength, max_iter = 1000 specifies the maximum number of iterations, and random_state = 42 ensures repeatability of the results.

[0090] Decision Tree Classifier: This classifies data using a tree structure, gradually assigning data to different categories by continuously partitioning features. criterion = 'gini' specifies the Gini impurity score as the classification criterion, max_depth = None specifies an unlimited tree depth, min_samples_split = 2 specifies that each internal node contains at least two samples, and random_state = 42 ensures consistent results across runs.

[0091] Random Forest Classifier: This model is an ensemble of multiple decision trees. It constructs multiple decision trees by randomly selecting features and samples, and then combines the results of these decision trees for classification. n_estimators = 100 indicates the use of 100 decision trees. Other parameter settings are similar to those for the decision tree classifier. n_jobs = -1 indicates the use of all available CPU cores for parallel computing, improving training efficiency.

[0092] HistGradientBoosting-based classifier: Based on the gradient boosting algorithm, a strong classifier is constructed by iteratively training a series of weak classifiers. max_iter = 100 specifies the maximum number of iterations, learning_rate = 0.1 controls the step size of each iteration, l2_regularization = 0.0 disables L2 regularization, max_bins = 255 is used for histogram optimization, class_weight = 'balanced' is used to address data imbalance, and random_state = 42 ensures reproducible results.

[0093] XGBoost Classifier: An efficient gradient boosting algorithm that more accurately approximates the optimal solution by performing a second-order Taylor expansion on the loss function. n_estimators = 100 specifies the use of 100 weak classifiers, max_depth = 6 specifies the maximum tree depth, learning_rate = 0.1 controls the learning rate, objective = 'binary:logistic' specifies the use of a binary logistic regression objective function, use_label_encoder = False disables the use of a label encoder, eval_metric = 'logloss' uses logarithmic loss as the evaluation metric, and random_state = 42 ensures consistent results.

[0094] LightGBM classifier: This is also a gradient boosting algorithm that uses a histogram algorithm and a depth-first tree growing strategy, which is characterized by high efficiency and low memory usage. n_estimators = 100, max_depth = -1 means no limit on the tree depth, learning_rate = 0.1, objective = 'binary' indicates a binary classification task, class_weight = 'balanced' handles data imbalance, boosting_type = 'gbdt' uses a gradient boosting decision tree, verbose = -1 means no detailed output, random_state = 42 ensures repeatability, and n_jobs = -1 enables parallel computing.

[0095] LSTM / GRU classifiers: Suitable for processing sequential data, such as time series test data. LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit) address the vanishing gradient problem of traditional recurrent neural networks through gating mechanisms, better capturing long-term dependencies in sequences. input_dim represents the dimension of the input features, hidden_size = 64 represents the number of neurons in the hidden layer, num_layers = 2 represents the number of layers in the network, output_size = 1 represents the output dimension, learning_rate = 0.001 controls the learning rate, batch_size = 64 represents the number of samples per training run, num_epochs = 10 represents the number of training epochs, and task = 'binary' represents a binary classification task.

[0096] The classification model is trained using the training dataset by continuously adjusting the model parameters to make the prediction results as close to the true labels as possible. During the training process, the model learns the mapping relationship between features and labels based on the input data. Then the trained model is evaluated using the test dataset, and common evaluation metrics include accuracy, recall, and F1 value. According to the evaluation results, the model with the best performance is selected for subsequent data analysis and prediction. For example, if a certain model has the highest accuracy on the test dataset, then this model is selected to classify and predict new electronic medical record data, providing support for medical research and clinical decision-making.

[0097] Test Example

[0098] I. Purpose of the Test

[0099] To verify the effectiveness and practicality of the method for de-identifying patient information and standardizing laboratory data, including evaluating the security of patient information de-identification, the accuracy of laboratory data standardization, and the performance of data modeling and classification.

[0100] II. Test Materials and Environment

[0101] (I) Test Data

[0102] A total of 5000 electronic medical record data were collected from three different sizes and types of medical institutions (large general hospitals, specialized hospitals, and primary community hospitals), covering patient demographic information, LE and LMR related information, and laboratory test results. These data contain rich diversity to simulate complex situations in actual applications.

[0103] (II) Test Environment

[0104] Hardware Environment: Servers equipped with Intel Xeon processors, 128GB of memory, and NVIDIA GPUs to meet the needs of large-scale data processing and model training.

[0105] Software Environment: Linux operating system, using Python 3.8 as the main programming language, combined with relevant data processing and machine learning libraries such as Pandas, Scikit-learn, XGBoost, LightGBM, and PyTorch.

[0106] III. Test Steps

[0107] (I) Patient Information De-identification Test

[0108] 1. Data Collection and Preliminary Processing

[0109] According to the established data collection process, data were extracted from the electronic medical record systems of the three medical institutions and preliminarily cleaned and organized to remove duplicate records and records missing key information.

[0110] 2. Sensitive Information Identification and Desensitization

[0111] Use pre-set rules and algorithms to identify sensitive information in the data (such as patient name, ID number, contact information), and apply replacement, masking, and truncation methods to desensitize it.

[0112] 3. Evaluation of desensitization effect

[0113] Privacy protection assessment: 500 pieces of desensitized data were randomly selected and examined by professional privacy assessors to assess whether sensitive information was effectively hidden and whether there was a risk of information leakage.

[0114] Data availability assessment: Use the desensitized data to perform simple statistical analysis (such as age distribution and gender ratio), compare the statistical results with the original data, and evaluate the impact of desensitization on data availability.

[0115] (2) Inspection data standardization test

[0116] 1. Data Alignment and Statistical Analysis

[0117] Test data were aligned to hospitalization records based on test date, and key variables were extracted from the previous hospitalization. The importance scores of demographics, LE and LMR, and laboratory test results were compared across the three policy phases, and statistical significance was assessed using the Wilcoxon test.

[0118] 2. Data standardization

[0119] Standardization of test item names: Establish a unified test item coding dictionary to map test items with different names used by different medical institutions to standard codes.

[0120] Standardization of the numerical range of test results: The numerical range of test results is standardized according to international standards, including unit conversion and precision adjustment.

[0121] Standardization of inspection report format: Develop a unified inspection report template and convert inspection reports in different formats into a standard template.

[0122] 3. Standardization effect evaluation

[0123] Data consistency assessment: Randomly select 300 standardized test data to check whether the test item name, result value range and report format meet the standard requirements.

[0124] Evaluation of data analysis results: Correlation analysis and cluster analysis were performed using the data before and after standardization to compare the accuracy and reliability of the analysis results.

[0125] (3) Data modeling and classification experiments

[0126] 1. Data partitioning

[0127] The processed electronic medical record data and test data are divided into training data sets (3500 copies) and test data sets (1500 copies) in a ratio of 7:3.

[0128] 2. Model training and evaluation

[0129] The training dataset is used to train logistic regression, decision tree classifier, random forest classifier, HistGradientBoosting based classifier, XGBoost classifier, LightGBM classifier and LSTM / GRU classifier.

[0130] Use the test data set to evaluate the trained model and calculate the model's accuracy, recall, and F1 value indicators.

[0131] 3. Model Selection and Performance Comparison

[0132] Based on the evaluation results, the model with the best performance is selected, and the advantages and disadvantages of different models in processing electronic medical record data are compared.

[0133] 4. Test Results

[0134] (1) Results of patient information desensitization test

[0135] 1. Privacy Protection Assessment

[0136] No sensitive information leakage was found in the 500 randomly selected anonymized data sets. Professional evaluators believe that the anonymization process effectively protects patients' privacy.

[0137] 2. Data Availability Assessment

[0138] After desensitization, the statistical analysis results of the data are basically consistent with the original data. For example, the errors in age distribution and gender ratio are both within 1%, indicating that desensitization has little impact on the availability of the data.

[0139] (2) Test results of inspection data standardization

[0140] 1. Data consistency assessment

[0141] Among the 300 randomly selected standardized test data, more than 98% of the test item names, result numerical ranges and report formats met the standard requirements, indicating that data standardization processing has high accuracy.

[0142] 2. Data analysis effect evaluation

[0143] Using standardized data for correlation analysis and cluster analysis significantly improves the accuracy and reliability of the results. For example, in correlation analysis, standardized data can more accurately reflect the relationship between different test indicators; in cluster analysis, clustering results are clearer and more reasonable.

[0144] (3) Data modeling and classification test results

[0145]

[0146] From the results, we can see that the LightGBM classifier performs best in terms of accuracy, recall and F1 value, and can classify and predict electronic medical record data more accurately.

[0147] The results of this experiment demonstrate the effectiveness and practicality of the "Method for Desensitizing Patient Information in Electronic Medical Records and Standardizing Laboratory Data." Desensitizing patient information effectively protects patient privacy while minimizing the impact on data availability. Standardizing laboratory data improves data consistency and analytical effectiveness. Various classification models demonstrated varying performance when processing electronic medical record data, with the LightGBM classifier performing optimally and being the preferred model for subsequent data analysis and prediction. This method provides a reliable solution for the secure processing and effective utilization of electronic medical record data.

[0148] The above embodiments are only used to illustrate the present invention and are not intended to limit the technical solutions described in the present invention. Although this specification has described the present invention in detail with reference to the above embodiments, the present invention is not limited to the above specific implementation methods. Therefore, any modification or replacement of the present invention; and all technical solutions and improvements thereof that do not depart from the spirit and scope of the invention are included in the scope of the claims of the present invention.

Claims

1. A method for desensitizing patient information in electronic medical record data and standardizing test data, characterized in that: The following steps are involved: Step 1: Patient information desensitization step: Collect patient data from the electronic medical record system, identify sensitive information fields, and use multiple desensitization algorithms such as replacement, masking, and truncation to process sensitive information; Step 2: Standardization of test data: Align test data to hospitalization records based on test date. Extract key variables from the previous hospitalization. Compare the importance scores of demographics, LE and LMR, and laboratory test results across the three policy phases and assess statistical significance using the Wilcoxon test. Establish a unified standard format and coding rules for test data. Step 3: Data modeling and classification: Use multiple classification models to model and analyze the processed data, use the training data set to train the model, use the test data set to evaluate the model performance and select the optimal model.

2. The method for desensitizing patient information in electronic medical record data and standardizing test data according to claim 1, characterized in that: In the patient information desensitization step, the collected patient data includes demographic information, LE and LMR information, and laboratory test results.

3. The method for desensitizing patient information in electronic medical record data and standardizing test data according to claim 1, characterized in that: In the patient information desensitization step, the identified sensitive information fields include the patient's name, ID number, and contact information.

4. The method for desensitizing patient information in electronic medical record data and standardizing test data according to claim 1, characterized in that: In the patient information desensitization step, the desensitization algorithm used is specifically: For patient names, a replacement method was used to replace the real names with randomly generated anonymous names; For ID card numbers, a masking method is used to retain only some key digits and replace the rest with characters; For contact information, the truncation method is adopted to retain only part of the number segment and hide the key part.

5. The method for desensitizing patient information in electronic medical record data and standardizing test data according to claim 1, characterized in that: In the test data standardization step, formulating a unified test data standard format and coding rules includes: Standardize the names of inspection items and use unified international or domestic inspection item codes; Standardize the numerical range of test results and use unified units and precision; Standardize the format of inspection reports and standardize the content and layout of reports.

6. The method for desensitizing patient information in electronic medical record data and standardizing test data according to claim 1, characterized in that: The classification models used in the data modeling and classification steps include: Logistic regression (scikit-learn v1.2.2) with solver = 'lbfgs', penalty = 'l2', C = 1.0, max_iter = 1000, random_state = 42. Decision tree classifier (scikit-learn v1.2.2) with criterion = 'gini', max_depth = None, min_samples_split = 2, random_state = 42; Random Forest Classifier (scikit-learn v1.2.2) with parameters set to n_estimators = 100, criterion = 'gini', max_depth = None, min_samples_split = 2, random_state = 42, n_jobs = -1.

7. The method for desensitizing patient information in electronic medical record data and standardizing test data according to claim 1, characterized in that: The classification model used in the data modeling and classification step also includes: A classifier based on HistGradientBoosting (HistGradientBoostingClassifier, scikit-learn v1.2.2) with parameters set to max_iter=100, learning_rate=0.1, l2_regularization=0.0, max_bins=255, class_weight='balanced', random_state=42; XGBoost classifier (xgboost v1.7.6) with parameters set to n_estimators = 100, max_depth = 6; learning_rate = 0.1; objective='binary:logistic'; use_label_encoder=False, eval_metric='logloss', random_state=42; LightGBM classifier (lightgbm v3.3.2) with parameters set to n_estimators = 100, max_depth = -1; learning_rate=0.1, objective='binary', class_weight='balanced', boosting_type='gbdt', verbose=-1, random_state=42, n_jobs=-1.

8. The method for desensitizing patient information in electronic medical record data and standardizing test data according to claim 1, characterized in that: The classification model used in the data modeling and classification steps also includes an LSTM / GRU classifier (PyTorch), with parameters set as input_dim=feature dimension, hidden_size=64, num_layers=2, output_size=1, learning_rate=0.001, batch_size=64, num_epochs=10, and task='binary'.

9. The method for desensitizing patient information in electronic medical record data and standardizing test data according to claim 1, characterized in that: In the test data normalization step, statistical significance was assessed using the Wilcoxon test, and the annotation rules were as follows: ns: p>0.05; p≤0.05; p≤0.01; p≤0.001; p≤0.0001.

10. The method for desensitizing patient information in electronic medical record data and standardizing test data according to claim 1, characterized in that: In the data modeling and classification step, the processed electronic medical record data and test data are divided into a training data set and a test data set, with the training data set accounting for 70% and the test data set accounting for 30%.

Images