Multi-source data fusion-based method and system for predicting drug-induced adverse reactions

Abstract

The present application provides a kind of multi-source data fusion's drug-induced adverse reaction prediction method and system, including obtaining at least two kinds of different sources of drug-related data and health record data, each kind of data is cleaned and format standardization processing, conflict detection is carried out based on pre-set conflict evaluation rule and the record of existence conflict is identified, the record of existence conflict is based on pre-set reliability evaluation rule and is carried out reliability score and is selected reliable field according to the scoring result Data fusion generates fusion data set, constructs and trains prediction model based on fusion data set, the risk prediction probability of the multi-dimensional characteristics of the object to be predicted is input into the trained prediction model to obtain. The present application improves the data reliability and feature coverage dimension by multi-source data fusion, and realizes the accurate prediction of drug-induced adverse reactions through the prediction model.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and medical data mining, and in particular to a method and system for predicting drug-induced adverse reactions by multi-source data fusion. Background Technology

[0002] Drug treatments may cause adverse reactions such as hypertension, and early prediction is crucial for the prevention and treatment of these adverse reactions. Existing machine learning-based prediction methods mostly rely on single features (such as drug dosage) from adverse event reporting systems (FAERS). However, the factors leading to adverse reactions are complex and related to multiple factors such as clinical data and lifestyle. It is difficult to achieve accurate prediction using only a single data source and a single feature.

[0003] Currently, the main technical bottlenecks in constructing predictive models for drug-induced adverse reactions are as follows: (1) Data level: Single data source has limited coverage dimensions. When multiple data sources are fused, problems such as logical conflicts and abnormal formats are likely to occur, making it difficult to guarantee data reliability and providing high-quality data support for model training. (2) Model level: Traditional models lack adaptive processing mechanisms for multi-source features. Fixed weights cannot match the dynamic importance of features during training, resulting in insufficient prediction accuracy and generalization ability, making it difficult to adapt to different clinical scenarios.

[0004] Therefore, how to achieve reliable fusion of multi-source data and high-precision prediction of drug-induced adverse reactions is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] This invention provides a method and system for predicting drug-induced adverse reactions by fusing multi-source data, so as to achieve reliable fusion of multi-source data and high-precision prediction of drug-induced adverse reactions.

[0006] On one hand, the present invention provides a method for predicting drug-induced adverse reactions based on multi-source data fusion, comprising: Obtain drug-related data and health record data from at least two different sources; The drug-related data and the health record data are cleaned and standardized to obtain their respective standardized datasets. Based on the preset conflict assessment rules, conflict detection is performed on each standardized dataset to identify conflicting records; For records with conflicts, a reliability score is calculated based on preset reliability assessment rules. Reliable fields are selected based on the score results to complete data fusion and generate a fused dataset. A prediction model is built and trained based on the fused dataset; By inputting the multi-dimensional features of the object to be predicted into the trained prediction model, the probability of predicting the risk of drug-induced adverse reactions can be obtained.

[0007] On the other hand, the present invention also provides a prediction system for drug-induced adverse reactions based on multi-source data fusion, comprising: The acquisition module is used to acquire drug-related data and health record data from at least two different sources; The processing module is used to clean and standardize the acquired data of various types to obtain corresponding standardized datasets; The detection module is used to perform conflict detection on each standardized dataset based on preset conflict assessment rules and identify conflicting records. The generation module is used to score the reliability of conflicting records based on preset reliability assessment rules, select reliable fields based on the scoring results to complete data fusion, and generate a fused dataset. Build modules are used to build and train prediction models based on fused datasets; The prediction module is used to input the multi-dimensional features of the object to be predicted into the trained prediction model to obtain the probability of predicting the risk of drug-induced adverse reactions.

[0008] The present invention provides a method and system for predicting drug-induced adverse reactions based on multi-source data fusion. This method and system involves constructing a multi-source data acquisition and standardization preprocessing module to clean and standardize the acquired drug-related data and health record data from at least two different sources. A conflict detection and reliability scoring mechanism is built based on a multi-dimensional validation data reliability assessment method. Conflict detection is performed on each standardized dataset, and conflicting records are identified. The records with conflicts are then scored for reliability, and reliable fields are selected based on the scoring results to complete data fusion and generate a fused dataset. A prediction model is built and trained based on the fused dataset. The multi-dimensional features of the object to be predicted are input into the trained prediction model to obtain the risk prediction probability of drug-induced adverse reactions. This achieves reliable multi-source data fusion and high-precision prediction of drug-induced adverse reactions. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating the method for predicting drug-induced adverse reactions based on multi-source data fusion provided in this embodiment of the invention. Figure 2This is a network framework diagram for predicting drug-induced adverse reactions based on multi-source data fusion; Figure 3 This is a schematic diagram of the structure of the drug-induced adverse reaction prediction system based on multi-source data fusion provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0012] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0013] Figure 1 This is a flowchart illustrating the method for predicting drug-induced adverse reactions based on multi-source data fusion provided in this embodiment of the invention.

[0014] like Figure 1 As shown, the method for predicting drug-induced adverse reactions based on multi-source data fusion provided in this embodiment of the invention mainly includes the following steps: 101. Obtain drug-related data and health record data from at least two different sources; 102. The drug-related data and the health record data are cleaned and standardized to obtain their respective standardized datasets; 103. Based on the preset conflict assessment rules, perform conflict detection on each standardized dataset and identify conflicting records; 104. For records with conflicts, reliability scores are performed based on preset reliability assessment rules. Reliable fields are selected according to the score results to complete data fusion and generate a fused dataset. 105. Construct and train a prediction model based on the fused dataset; 106. Input the multi-dimensional features of the object to be predicted into the trained prediction model to obtain the probability of predicting the risk of drug-induced adverse reactions.

[0015] In this embodiment, the conflict assessment rule refers to the evaluation criteria used to measure the degree of inconsistency between corresponding records of the same object in different data sources. Specifically, it includes three evaluation dimensions: logical consistency, format anomalies, and numerical anomalies, and the conflict score is calculated through weighted summation. The reliability assessment rule refers to the evaluation criteria used to assess the credibility of records in each data source. Specifically, it includes three evaluation dimensions: data collection standardization, clinical benchmark alignment, and key indicator completeness, and the reliability score is also calculated through weighted summation. The fused dataset refers to the final dataset formed by merging data from multiple sources after conflict resolution, where each record selects field values ​​from the most reliable data source. The prediction model refers to a binary classification model built on a fully connected neural network, used to output the probability of adverse reactions.

[0016] The multi-source data fusion method for predicting drug-induced adverse reactions provided in this embodiment takes the prediction of drug-induced hypertension as a typical application scenario. The specific implementation process is as follows: Data can be obtained from two different sources: adverse event data from the Fatal Event Reporting System (FAERS), referred to as FAERS data, and clinical record data from the Sim-EHR (Sim-EHR) analog electronic health record system, referred to as Sim-EHR data. Both types of data contain multi-dimensional information such as patient demographics, medication history, physiological measurements, adverse reaction labels, and lifestyle indicators.

[0017] The acquired FAERS and Sim-EHR data were cleaned and standardized independently. Through operations such as outlier detection and removal, conversion of special format data, and marking unconvertible data as missing values, standardized Ext-FAERS and standardized Sim-EHR datasets were formed.

[0018] Conflict detection is performed on two standardized datasets based on preset conflict assessment rules. For each record in each dataset, the corresponding field is scored from three dimensions: logical consistency, format abnormality, and numerical abnormality. The scores are weighted and summed to obtain the conflict score. Records with conflict scores exceeding a preset threshold are identified as conflicting records.

[0019] A reliability score is obtained by weighted summation of scores from three dimensions: data collection standardization, alignment with clinical benchmarks, and completeness of key indicators. For conflicting records of the same patient, field values ​​from records with higher reliability scores are selected field by field and combined with consistent records without conflicts to generate a high-reliability fusion dataset.

[0020] Furthermore, a prediction model can be built and trained based on the fused dataset. This embodiment uses a fully connected neural network as the prediction model. Its structure includes an input layer dimension matching the dimensions of the selected features, multiple fully connected layers with ReLU activation function in the hidden layers, a Dropout layer after each hidden layer, and an output layer using the Sigmoid activation function to output the probability of adverse reactions. The Adam optimizer and binary cross-entropy loss function are used to iteratively train the model until convergence.

[0021] Finally, the multi-dimensional features of the patient to be predicted (after the same standardization process as the training data) are input into the trained prediction model. The model outputs the probability value of adverse reaction between 0 and 1. Based on the threshold determined by the validation set optimization, it is determined whether there is a high risk of drug-induced adverse reaction.

[0022] This embodiment realizes a complete process of acquiring, cleaning and standardizing, conflict detection and resolution, data fusion, model building and training, and risk prediction of multi-source drug-related data and health record data. It solves the problems of limited coverage dimensions of single data sources, logical conflicts that are easy to occur when fusing multi-source data, and difficulty in ensuring data reliability. It provides reliable technical support for the accurate prediction of drug-induced adverse reactions, and realizes end-to-end risk probability output through a fully connected neural network, which facilitates clinical decision-making.

[0023] In some embodiments, in step 101 above, acquiring at least two types of drug-related data and health record data from different sources includes: Obtain adverse drug event data from the adverse event reporting system; To obtain clinical record data from the electronic health record system; The adverse drug event data and clinical record data both include patient demographic information, medication history, physiological index measurements, adverse reaction labels, and lifestyle indicators.

[0024] In this embodiment, the adverse event reporting system refers to a database used to collect spontaneous reports of adverse reactions after drug marketing; the electronic health record system refers to a digital system used by medical institutions to store patients' clinical diagnosis and treatment information, such as the Sim-EHR analog electronic health record system. Adverse drug event data records information on adverse reaction events that occur after drug use, including drug exposure and event descriptions; clinical record data records longitudinal health information generated by patients during clinical diagnosis and treatment, including medical history, physical examinations, and laboratory tests.

[0025] In detail, adverse drug event data can be obtained from the adverse event reporting system, which includes information such as drug dosage, adverse reaction type, patient gender, and age; and clinical record data can be obtained from the Sim-EHR simulated electronic health record system, which includes clinical characteristics such as patient medical history, blood pressure measurements (systolic and diastolic), smoking history, and alcohol consumption history. Both types of data must include the following five core information fields: patient demographic information (e.g., age, gender, body mass index), medication history (drug name, dosage, duration of use), physiological indicator measurements (e.g., blood pressure, heart rate, blood glucose), adverse reaction labels (whether hypertension occurred and its severity), and lifestyle indicators (smoking, alcohol consumption, exercise habits). By simultaneously acquiring these two types of data, subsequent data fusion can cover both drug exposure information and clinical health information, compensating for the shortcomings of a single data source.

[0026] In some embodiments, in step 102 above, the various types of data are cleaned and standardized to obtain corresponding standardized datasets, including: using the Z-Score method to detect abnormal records in the data and removing abnormal records with an absolute Z value greater than a preset abnormal threshold; converting special format data to convert string format values ​​into numerical data; marking data that cannot be converted into missing values ​​to form a standardized dataset.

[0027] In this embodiment, the Z-Score method is a commonly used outlier detection method in statistics. It measures the degree of abnormality by calculating the deviation of each data point from the mean (in units of standard deviation). The preset outlier threshold is the critical Z-score value for determining whether a record is abnormal; in this embodiment, it is set to 3. Special format data refers to data in non-standard numerical representations, such as interval data (e.g., "120-140"), fractional data (e.g., "3 / 5"), and strings with units (e.g., "120mmHg").

[0028] In detail, perform the following operations for each type of data source (FAERS data and Sim-EHR data).

[0029] The first step is anomaly detection and removal. For each numerical field in the dataset (such as age, blood pressure, medication dosage, etc.), calculate the mean and standard deviation of that field. For each numerical field in each record, calculate its Z-score, which is (field value - mean) / standard deviation. If the absolute Z-score of any key field in the record is greater than a preset anomaly threshold (set to 3 in this embodiment), then the entire record is determined to be a statistically significant outlier and removed from the dataset. For example, if a patient's systolic blood pressure is 250 mmHg, and the mean of this field is 130 mmHg and the standard deviation is 20, then the Z-score is (250-130) / 20=6, which is greater than 3, and the record is removed.

[0030] The second step is the conversion of data with special formats. For non-standard numerical representations in the data, appropriate parsing rules are used for conversion. For example, for range formats such as "120-140mmHg", it is parsed as an average of 130; for fractions such as "3 / 5", a division operation is performed to obtain 0.6; for strings with units such as "120mmHg", the unit is removed to extract the value 120. For non-numerical fields such as dates, they are uniformly converted to the standard date format.

[0031] The third step is missing value labeling. For fields that still cannot be successfully converted into numerical data or cannot be parsed into a standard format after the above transformation, they are marked as NaN (missing value) to ensure that all fields have a unified data type and format specification. For records containing fields marked as missing values, if any key prediction field (including medication dosage, systolic blood pressure, diastolic blood pressure, and adverse reaction label) is missing, the record is removed to ensure the integrity and reliability of the training data; if the missing field is a non-key auxiliary field, the record is retained.

[0032] Step 4: Standardize the mapping of the drug name field and establish a unified mapping dictionary between the generic name of the drug and its aliases, trade names, and abbreviations in various data sources.

[0033] After the above four steps, standardized Ext-FAERS dataset and standardized Sim-EHR dataset are formed, providing data with uniform format and reliable quality for subsequent collision detection.

[0034] In some embodiments, step 103 above involves performing conflict detection on each standardized dataset based on preset conflict assessment rules to identify conflicting records, including: Each record is scored based on three dimensions: logical consistency, format anomalies, and numerical anomalies. The scores of the three dimensions are weighted and summed according to the preset dimension weights to obtain the conflict score for each record. Records with conflict scores exceeding a preset conflict threshold are identified as conflicting records.

[0035] In this embodiment, the logical consistency dimension is used to evaluate the degree of logical consistency between records of the same object in different data sources, such as whether the medication history and the trend of physiological indicator changes are consistent; the format anomaly dimension is used to detect whether the data format conforms to preset specifications, such as date format, encoding range, etc.; the numerical anomaly dimension is used to detect whether the numerical value is within a reasonable physiological or pharmacological range. The preset dimension weights are logical consistency 0.4, format anomaly 0.2, and numerical anomaly 0.4, respectively. The preset conflict threshold is an empirical value used to distinguish between normal records and conflicting records; in this embodiment, it is set to 0.6 (out of 1) based on the data distribution.

[0036] In detail, for each record (corresponding record of the same patient) in the two standardized datasets Ext-FAERS and Sim-EHR, perform the following operations.

[0037] First, a score is awarded based on logical consistency. This score is determined by examining whether there are inconsistencies between the medication history records and physiological indicator changes of the same patient in two data sources. For example, if a patient's blood pressure measurements in the Sim-EHR dataset show consistently low blood pressure, the logical consistency score will be low; conversely, it will be high. The score uses a continuous value between 0 and 1, with 1 for complete consistency and 0 for complete inconsistency.

[0038] Secondly, scoring is based on format anomalies. Check whether the format of each field conforms to the expected specifications. For example, are all date fields in "YYYY-MM-DD" format? Are gender fields only in "Male / Female" or "M / F" format? Is the diagnostic code within the ICD code range? The higher the format conformity, the higher the score.

[0039] Next, scoring is based on numerical abnormalities. The numerical field is checked to see if it falls within a reasonable physiological range. For example, the normal range for systolic blood pressure is generally 90-180 mmHg. Values ​​outside this range are considered abnormal, and the closer the value is to the normal center value, the higher the score.

[0040] After scoring across the three dimensions, a weighted sum is calculated based on preset weights (logical consistency 0.4, formatting anomaly 0.2, numerical anomaly 0.4) to obtain a conflict score for each record. The conflict score is then compared to a preset conflict threshold (0.6 in this embodiment). If the score exceeds 0.6, the record is considered to have a conflict and requires further reliability assessment and conflict resolution procedures; if the score does not exceed 0.6, the record is considered to have no conflict and can be directly retained.

[0041] This embodiment, through the aforementioned three-dimensional weighted scoring mechanism, can comprehensively and quantitatively identify conflicting records in multi-source data. It considers both the format and numerical rationality of the data itself, as well as the logical consistency between different sources, thus avoiding the decrease in reliability caused by simple elimination or arbitrary fusion.

[0042] In some embodiments, in step 104 above, the conflicting records are scored based on preset reliability assessment rules, and reliable fields are selected according to the scoring results to complete data fusion and generate a fused dataset, including: Conflicting records were scored based on three dimensions: data collection standardization, alignment with clinical benchmarks, and completeness of key indicators. The scores of the three dimensions are weighted and summed according to the preset dimension weights to obtain the reliability score of each conflict record. For multiple conflicting records of the same object, the field from the record with the higher reliability score is selected as the fused dataset.

[0043] In this embodiment, data collection standardization refers to whether the data collection process conforms to standard operating procedures, such as whether it is generated by a certified system and whether the collection time is clearly defined; clinical benchmark alignment refers to whether the data values ​​are consistent with recognized clinical guidelines or reference standards; and key indicator completeness refers to whether the records contain the fields most important to the prediction task (such as medication dosage, blood pressure value, and adverse reaction label) and are not missing. The preset dimension weights are 0.3 for data collection standardization, 0.5 for clinical benchmark alignment, and 0.2 for key indicator completeness.

[0044] In detail, for each record that is determined to have a conflict (including the corresponding record in the Ext-FAERS dataset and the corresponding record in the Sim-EHR dataset), its reliability score is calculated.

[0045] First, the data collection compliance dimension is scored. This assesses the authority and standardization of the data source for the record. For example, FAERS, as a national adverse event reporting system, has a standardized data collection process and scores highly; Sim-EHR, as simulated data, has compliance depending on the generation rules, and is scored based on the actual situation. Simultaneously, it checks whether the record includes standardized information such as collection timestamps and collector identification.

[0046] Secondly, the score is based on clinical benchmark alignment. Key medical values ​​recorded (such as blood pressure and medication dosage) are compared with reference standards in clinical practice guidelines (such as the JNC hypertension guidelines). The closer the value is to the guideline recommendation or the center of the normal range, the higher the score. For example, a systolic blood pressure of 120 mmHg is closer to the normal benchmark than 160 mmHg, and therefore scores higher.

[0047] Secondly, scoring is based on the completeness of key indicators. The records are checked to ensure they fully include the five core fields: demographic information, medication history, blood pressure measurements, adverse reaction labels, and lifestyle indicators. The fewer missing fields, the higher the score.

[0048] After scoring each of the three dimensions, the scores are weighted and summed according to their respective weights (0.3, 0.5, 0.2) to obtain the reliability score of the record. For conflicting records of the same patient (from the FAERS and Sim-EHR datasets respectively), the reliability scores of the two records are compared, and then the field value from the record with the higher score is selected field by field. For example, if the blood pressure value recorded in the FAERS dataset is more reliable, the blood pressure value from the FAERS dataset is used; if the medication history in the Sim-EHR dataset is more reliable, the medication history from the Sim-EHR dataset is used. All consistent records without conflicts are directly retained. After the above field-level optimization fusion, a highly reliable fused dataset is finally generated.

[0049] This embodiment achieves effective reliability assessment and intelligent fusion of conflicting records. By comprehensively scoring data from three dimensions—data collection standardization, clinical benchmark alignment, and key indicator completeness—and selecting fields from records with higher reliability scores for fusion, the problem of multi-source data conflict is effectively resolved, ensuring the reliability and completeness of the fused dataset.

[0050] In some embodiments, before building and training the prediction model based on the fused dataset, the method further includes: Continuous features in the fused dataset are discretized using binning. Calculate the mutual information value between each discretized feature and the adverse reaction label; Features whose mutual information values ​​are greater than or equal to a preset mutual information threshold are retained to form a candidate feature set; The candidate feature set is further filtered through a feature group scoring mechanism to determine the optimal input feature set.

[0051] In this embodiment, binning refers to dividing continuous numerical values ​​into several discrete intervals. This embodiment uses equal-frequency binning, ensuring that each interval contains approximately the same number of samples. Mutual information is an indicator in information theory that measures the degree of interdependence between two variables; a higher value indicates a stronger correlation. The preset mutual information threshold is set to 0.7 in this embodiment. The feature group scoring mechanism involves grouping features by category and then comprehensively scoring each group, reducing redundancy while ensuring sufficient information content.

[0052] In detail, this embodiment adds a feature selection step before model training to reduce input dimensionality and improve model performance. The specific process is as follows.

[0053] First, the continuous features (such as age, systolic blood pressure, diastolic blood pressure, and medication dosage) in the fused dataset are discretized using equal-frequency binning. The value range of each feature is divided into several quantile intervals, ensuring that each interval contains approximately the same number of samples. For example, age is divided into four intervals: "under 30 years old," "30-45 years old," "45-60 years old," and "over 60 years old." Discretized features facilitate the calculation of mutual information between the features and the labels.

[0054] Next, the mutual information value between each discretized feature and the adverse reaction label (whether hypertension occurred) is calculated. Mutual information is calculated based on the ratio of the joint probability distribution to the marginal probability distribution, with values ​​between 0 and 1. A higher value indicates a stronger predictive ability of the feature for the label. In this embodiment, a mutual information threshold of 0.7 is set, retaining all features with mutual information values ​​greater than or equal to 0.7 to form a candidate feature set, while eliminating features that are weakly correlated with or unrelated to the label.

[0055] Next, the candidate feature sets are scored by feature groups. Candidate features are categorized according to their information type, such as demographics, medication history, physiological indicators, and lifestyle. For each group, the sum of mutual information of all features within the group and the average correlation between features are calculated to obtain a contribution score for that group. By comparing the scores of each group, groups with higher scores are retained, and highly redundant features within each group are removed (e.g., retaining both systolic and diastolic blood pressure may be redundant; only one needs to be selected). Finally, the optimal input feature set is determined, with a dimensionality significantly lower than the total number of original features.

[0056] This embodiment effectively reduces the feature dimension, eliminates irrelevant or redundant features, reduces the risk of model overfitting, shortens training time, improves the model's generalization ability and prediction accuracy, and enables subsequent fully connected neural networks to focus on high-contribution features.

[0057] In some embodiments, a dynamic weight allocation mechanism is employed during the training process of constructing and training the prediction model, including: At the beginning of each training cycle, the initial weights of each input feature are calculated based on the temperature-scaling Softmax function; A momentum factor is introduced to perform weighted smoothing of the initial weights of the current period and the final weights of the previous period; The smoothed weights are used as the final weights for each feature in the current period for model training.

[0058] In this embodiment, the temperature-scaling Softmax function refers to the standard Softmax function with temperature parameter adjustment, which controls the smoothness of the weight distribution. The momentum factor is used to control the influence of historical weights on current weights; in this embodiment, it is set to 0.15. A training cycle, or one epoch, refers to completing one forward and backward propagation of the entire training set.

[0059] In detail, at the beginning of each training epoch t, ​​for each feature of the input layer, its importance score in the current model state is first calculated (e.g., based on gradient or loss contribution). Then, a temperature-scaling Softmax function is applied to convert these scores into initial weights. Specifically, the importance score of each feature is divided by the temperature parameter τ (τ=0.5 in this example), and then the scaled scores of all features are exponentially normalized to obtain the initial weight distribution. The temperature parameter τ=0.5 makes the weight distribution sharper than the standard Softmax (τ=1), that is, important features have higher weights and unimportant features have lower weights, which is beneficial for highlighting key features.

[0060] It should be noted that for the first training epoch (t=1), the model has not yet undergone any parameter updates. At this time, the importance score is calculated based on the initial connection weights from the input layer to the first hidden layer (i.e., the weights after random initialization and without training), and this is used as the initial weights for the first round. At the same time, the final weights of the previous epoch do not exist in the first epoch, so they are initialized to a uniform distribution, that is, all feature weights are equal, and subsequent epochs are updated normally according to the smoothing formula.

[0061] Momentum factor can be introduced (This embodiment) The initial weights of the current period are weighted and smoothed by the final weights of the previous period (t-1). The specific formula is: Final weight of the current period = (1- ) × current initial weight + ×Previous epoch final weights. This smoothing process prevents weight changes from being too drastic, maintaining the stability of the training process. For the first epoch, the previous epoch final weights can be initialized with a uniform distribution.

[0062] The smoothed weights can be used as the final weights of each feature in the current cycle. In the forward propagation of model training, each input feature is multiplied by its corresponding weight and then fed into the network, so that the model can dynamically adjust the importance of features in each cycle.

[0063] This embodiment uses a dynamic weight allocation mechanism to enable the model to adaptively adjust the importance of each feature at different training stages, overcoming the shortcomings of traditional fixed weights that cannot match dynamic changes in features. This significantly improves the model's convergence speed and prediction accuracy, and enhances the model's adaptability to multi-source heterogeneous features.

[0064] In some embodiments, after obtaining the predicted probability of drug-induced adverse reactions, the method further includes: Based on the dynamic weights of each feature during model training, the contribution of each feature to the current prediction result is calculated. The features are sorted from highest to lowest contribution, and a predetermined number of features with the highest contribution are selected. Generate a predictive explanatory report that includes high-contribution features and their degree of contribution.

[0065] In this embodiment, contribution refers to the degree of influence of a certain input feature on the current prediction result. This embodiment directly uses the final weight of the feature in the current period in the dynamic weight allocation mechanism as the contribution. The preset number can be set according to clinical needs, such as taking the first 5 or the first 10 features. The prediction interpretation report is a readable document output in a structured format to help clinicians understand the basis of the prediction.

[0066] In detail, after the model outputs a risk probability for a patient to be predicted, it obtains the dynamic weight values ​​corresponding to each input feature during the prediction process (i.e., the final weights at the current training cycle or inference time), and directly uses these weight values ​​as the contribution of each feature to the prediction result. Since the dynamic weight allocation mechanism is continuously updated during training, the weights of the most recent cycle can be directly used during inference.

[0067] Features can be sorted from highest to lowest contribution value, and a preset number (e.g., 5) of features can be selected. These features are the factors that have the greatest impact on the current prediction results. A prediction explanation report can be generated, listing the specific names, values, and corresponding contribution values ​​of the selected features. For example, the report might show: "Systolic blood pressure: 145 mmHg, contribution 0.32; Medication dosage: 50 mg, contribution 0.25; Age: 62 years, contribution 0.18; Smoking history: Yes, contribution 0.12; Alcohol consumption history: Yes, contribution 0.08." This report can be output in text or tabular format for clinicians' reference.

[0068] This embodiment enhances the interpretability of the model by generating a prediction explanation report, enabling clinicians not only to obtain the probability of risk prediction, but also to understand which factors led to the result, thereby adjusting medication regimens or interventions in a more targeted manner, and improving the practicality and credibility of the model in clinical practice.

[0069] In some embodiments, the prediction model is a fully connected neural network, comprising: The input layer has dimensions that match the feature dimensions of the filtered optimal input feature set. Hidden layers, consisting of multiple fully connected layers, use the ReLU activation function; Dropout layers are placed after each hidden layer and randomly drop some neurons according to a preset dropout rate to prevent the model from overfitting. The output layer uses the Sigmoid activation function to output the probability value of adverse reactions.

[0070] This embodiment specifically defines the network architecture of the prediction model, such as Figure 2 As shown. The structure of this fully connected neural network from input to output is as follows: Input layer: The number of neurons in this layer is equal to the dimension of the optimal input feature set selected in Example 6. Each neuron corresponds to one input feature and receives the standardized feature value.

[0071] Hidden Layers: Three fully connected layers are used, containing 64, 32, and 16 neurons respectively. Each neuron in each layer is connected to all neurons in the previous layer and uses the ReLU activation function, meaning the output is the positive part of the input, and the output is 0 if the input is negative. This activation function helps accelerate convergence and alleviate the vanishing gradient problem. The number of neurons decreases layer by layer, achieving progressive abstraction and compression of features.

[0072] Dropout layer: A Dropout layer is set after each hidden layer, with a default dropout rate of 0.3. During the forward propagation of each training iteration, the output of the current hidden layer neuron is randomly set to 0 with a 30% probability, thus preventing the network from over-relying on any single neuron and effectively preventing overfitting. During test inference, the Dropout layer is not enabled, and all neurons participate in the computation.

[0073] Output layer: The Sigmoid activation function is used to map the output of the last hidden layer to a probability value between 0 and 1. This value directly represents the probability of the drug-induced adverse reaction (such as hypertension) predicted by the model. Based on a classification threshold optimized on the validation set (e.g., 0.5), when the output probability is greater than or equal to the threshold, it is judged as high risk; otherwise, it is judged as low risk.

[0074] The entire network uses the Adam optimizer, with a batch size of 128, 50 training epochs, a learning rate of 0.001, and a loss function of binary cross-entropy. Iterative training is performed until the model performance converges.

[0075] In some embodiments, Figure 2 This is a network framework diagram for predicting drug-induced adverse reactions based on multi-source data fusion, such as... Figure 2 As shown, it adopts a two-stage pipeline structure, clearly demonstrating the complete processing chain from raw multi-source data to the final risk prediction output: 201. Raw Multi-Source Data Input Unit: The raw data input for this prediction method includes two heterogeneous data sources: FAERS data and Sim-EHR data. The FAERS data source covers adverse drug event reports from Q1 2010 to Q3 2024, while the EHR data source covers clinical health data such as patient history, blood pressure measurement, and lifestyle, together ensuring the breadth of feature coverage of the multi-source fusion.

[0076] 202. Data Standardization Preprocessing Unit: It is responsible for cleaning and standardizing the raw data. After processing, it outputs two standardized datasets, Ext-FAERS and Sim-EHR, which are referred to as the standardized Ext-FAERS dataset and the standardized Sim-EHR dataset.

[0077] 203. Conflict Locating Submodule: For each corresponding record in the Ext-FAERS and Sim-EHR datasets, it calculates a conflict score based on three dimensions: logical consistency, format anomaly, and numerical anomaly, with preset weights of 0.4, 0.2, and 0.4, respectively. This quantitatively assesses the degree of conflict in the records and accurately locates conflicting data records and their corresponding fields.

[0078] 204. Conflict Resolution and Fusion Output Unit: For the identified conflict records, reliability scores are calculated based on three dimensions: data collection standardization, clinical benchmark alignment, and key indicator completeness, with weights of 0.3, 0.5, and 0.2, respectively. Fields with higher scores are selected as the basis for fusion. Consistent records without conflicts are retained, and finally, a highly reliable multi-source dataset, Multi-FHED, is generated by fusion, which serves as the data base for subsequent model training.

[0079] 205. Feature screening input module (feature screening + Input layer) It receives multi-source features from the Multi-FHED dataset and performs high-contribution feature screening: continuous features are discretized by equal-frequency binning, the mutual information between the features and the hypertension label is calculated, highly correlated features with mutual information ≥0.7 are retained, and further screening is performed by the feature group scoring mechanism to obtain the optimal input feature set and eliminate redundant low-contribution features.

[0080] 206. Dynamic weight assignment hidden layer (layer 1 Hidden): Dynamic weight allocation is performed in the first hidden layer after feature input: the initial weights of the features are calculated based on temperature scaling Softmax (temperature parameter τ=0.5), and the weight update process is smoothed by combining momentum factor (αw=0.15). The importance of features is dynamically adapted with the training rounds as the cycle, which strengthens the role of high-contribution features and suppresses the interference of low-value features.

[0081] 207. Second fully connected hidden layer: It corresponds to 32 neurons, inheriting the feature output of the previous layer, and performs nonlinear feature transformation through the ReLU activation function to further extract higher-order feature combinations; at the same time, it uses the Dropout (rate=0.3) mechanism to randomly deactivate some neurons, suppress model overfitting, and improve the model's generalization ability.

[0082] 208. Third fully connected hidden layer: It corresponds to 16 neurons, receives the feature output from the second hidden layer, and continues to complete the deep nonlinear feature mapping through the ReLU activation function, further compressing the feature dimension and abstracting the core prediction features. Combined with the Dropout mechanism, it continuously ensures the generalization performance of the model.

[0083] 209. Output Layer: It receives the feature output from the last hidden layer, performs numerical mapping using the Sigmoid activation function, and outputs a probability value of drug-induced hypertension with a value range of [0,1]. Subsequently, based on the determination threshold optimized by the validation set, it outputs the classification result of whether the patient has a high risk of drug-induced hypertension.

[0084] 210. Prediction Results and Performance Output Unit: The output consists of two types of results: first, the probability of hypertension risk prediction for a single sample, providing a safety warning reference for clinical medication; and second, the comparison results of the overall model performance indicators (accuracy, AUC, recall, F1 score, etc.), which are used to verify the effectiveness of model performance and multi-source data fusion, and support the evaluation of model effectiveness.

[0085] Based on the same general inventive concept, this invention also protects a prediction system for drug-induced adverse reactions based on multi-source data fusion. The prediction system for drug-induced adverse reactions based on multi-source data fusion provided by this invention will be described below. The prediction system for drug-induced adverse reactions based on multi-source data fusion described below can be referred to in correspondence with the prediction method for drug-induced adverse reactions based on multi-source data fusion described above.

[0086] Figure 3 This is a schematic diagram of the structure of the drug-induced adverse reaction prediction system based on multi-source data fusion provided in an embodiment of the present invention, as shown below. Figure 3As shown, the multi-source data fusion-based drug-induced adverse reaction prediction system of this embodiment includes: The acquisition module 31 is used to acquire drug-related data and health record data from at least two different sources; Processing module 32 is used to clean and standardize the acquired data of various types to obtain corresponding standardized datasets; The detection module 33 is used to perform conflict detection on each standardized dataset based on preset conflict assessment rules and identify records with conflicts. The generation module 34 is used to score the reliability of conflicting records based on preset reliability assessment rules, select reliable fields based on the scoring results to complete data fusion, and generate a fused dataset. Module 35 is used to build and train a prediction model based on the fused dataset; The prediction module 36 is used to input the multi-dimensional features of the object to be predicted into the trained prediction model to obtain the risk prediction probability of drug-induced adverse reactions.

[0087] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device may include: a processor 410, a communication interface 420, a memory 430, and a communication bus 440. The processor 410, communication interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions stored in the memory 430 to execute a method for predicting drug-induced adverse reactions based on multi-source data fusion.

[0088] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0089] It should be noted that all relevant information that may be involved in the various embodiments of the present invention is processed in strict accordance with the requirements of laws and regulations, following the principles of legality, legitimacy, and necessity, based on the reasonable purpose of the business scenario, and is information that users actively provide or generate during the use of the product / service, as well as information obtained with user authorization.

[0090] The information processed by this invention may vary depending on the specific product / service scenario and should be based on the specific scenario in which the user uses the product / service. This may involve user account information, device information, or other related information. This invention will treat the relevant information and its processing with the utmost diligence.

[0091] This invention places great emphasis on the security of relevant information and has adopted reasonable and feasible security protection measures that comply with industry standards to protect user information and prevent unauthorized access, public disclosure, use, modification, damage or loss of relevant information.

[0092] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0093] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting drug-induced adverse reactions based on multi-source data fusion, characterized in that, include: Obtain drug-related data and health record data from at least two different sources; The drug-related data and the health record data are cleaned and standardized to obtain their respective standardized datasets. Based on the preset conflict assessment rules, conflict detection is performed on each standardized dataset to identify conflicting records; For records with conflicts, a reliability score is calculated based on preset reliability assessment rules. Reliable fields are selected based on the score results to complete data fusion and generate a fused dataset. A prediction model is built and trained based on the fused dataset; By inputting the multi-dimensional features of the object to be predicted into the trained prediction model, the probability of predicting the risk of drug-induced adverse reactions can be obtained.

2. The method for predicting drug-induced adverse reactions based on multi-source data fusion according to claim 1, characterized in that, Obtain drug-related data and health record data from at least two different sources, including: Obtain adverse drug event data from the adverse event reporting system; To obtain clinical record data from the electronic health record system; The adverse drug event data and clinical record data both include patient demographic information, medication history, physiological index measurements, adverse reaction labels, and lifestyle indicators.

3. The method for predicting drug-induced adverse reactions based on multi-source data fusion according to claim 1, characterized in that, The acquired data of various types were cleaned and standardized to obtain corresponding standardized datasets, including: The Z-Score method is used to detect abnormal records in the data, and abnormal records with an absolute Z value greater than a preset abnormal threshold are removed. Perform format conversion on data with special formats, converting string-formatted numbers into numeric data; Data that cannot be formatted is marked as missing values, and records with missing values ​​in key fields are removed. Standardize the mapping of drug name fields in various types of data, establish a unified mapping dictionary between generic drug names and aliases, trade names, and abbreviations in various data sources, and form a standardized dataset.

4. The method for predicting drug-induced adverse reactions based on multi-source data fusion according to claim 1, characterized in that, Based on pre-defined conflict assessment rules, conflict detection is performed on each standardized dataset to identify conflicting records, including: Each record is scored based on three dimensions: logical consistency, format anomalies, and numerical anomalies. The scores of the three dimensions are weighted and summed according to the preset dimension weights to obtain the conflict score for each record. Records with conflict scores exceeding a preset conflict threshold are identified as conflicting records.

5. The method for predicting drug-induced adverse reactions based on multi-source data fusion according to claim 1, characterized in that, For conflicting records, a reliability score is assigned based on preset reliability assessment rules. Reliable fields are selected according to the score results to complete data fusion, generating a fused dataset, including: Conflicting records were scored based on three dimensions: data collection standardization, alignment with clinical benchmarks, and completeness of key indicators. The scores of the three dimensions are weighted and summed according to the preset dimension weights to obtain the reliability score of each conflict record. For multiple conflicting records of the same object, the field from the record with the higher reliability score is selected as the fused dataset.

6. The method for predicting drug-induced adverse reactions based on multi-source data fusion according to claim 1, characterized in that, Before building and training a prediction model based on the fused dataset, the following steps are also included: Continuous features in the fused dataset are discretized using binning. Calculate the mutual information value between each discretized feature and the adverse reaction label; Features whose mutual information values ​​are greater than or equal to a preset mutual information threshold are retained to form a candidate feature set; The candidate feature set is further filtered through a feature group scoring mechanism to determine the optimal input feature set.

7. The method for predicting drug-induced adverse reactions based on multi-source data fusion according to claim 1, characterized in that, The training process of building and training the prediction model employs a dynamic weight allocation mechanism, including: At the beginning of each training cycle, the initial weights of each input feature are calculated based on the temperature-scaling Softmax function; A momentum factor is introduced to perform weighted smoothing of the initial weights of the current period and the final weights of the previous period; The smoothed weights are used as the final weights for each feature in the current period for model training. In the first training cycle, the initial weights are calculated based on the initial connection weights from the input layer to the first hidden layer, and the final weights in the previous cycle are initialized to a uniform distribution.

8. The method for predicting drug-induced adverse reactions based on multi-source data fusion according to claim 7, characterized in that, After obtaining the predicted probability of drug-induced adverse reactions, the following is also included: Based on the dynamic weights of each feature during model training, the contribution of each feature to the current prediction result is calculated. The features are sorted from highest to lowest contribution, and a predetermined number of features with the highest contribution are selected. Generate a predictive explanatory report that includes high-contribution features and their degree of contribution.

9. The method for predicting drug-induced adverse reactions based on multi-source data fusion according to claim 1, characterized in that, The prediction model is a fully connected neural network, including: The input layer has dimensions that match the feature dimensions of the filtered optimal input feature set. Hidden layers, consisting of multiple fully connected layers, use the ReLU activation function; Dropout layers are placed after each hidden layer and randomly drop some neurons according to a preset dropout rate to prevent the model from overfitting. The output layer uses the Sigmoid activation function to output the probability value of adverse reactions.

10. A prediction system for drug-induced adverse reactions based on multi-source data fusion, characterized in that, include: The acquisition module is used to acquire drug-related data and health record data from at least two different sources; The processing module is used to clean and standardize the acquired data of various types to obtain corresponding standardized datasets; The detection module is used to perform conflict detection on each standardized dataset based on preset conflict assessment rules and identify conflicting records. The generation module is used to score the reliability of conflicting records based on preset reliability assessment rules, select reliable fields based on the scoring results to complete data fusion, and generate a fused dataset. Build modules are used to build and train prediction models based on fused datasets; The prediction module is used to input the multi-dimensional features of the object to be predicted into the trained prediction model to obtain the probability of predicting the risk of drug-induced adverse reactions.

Images