Drug dosage prediction method based on patient similarity
The method addresses the reliance on expert knowledge in patient similarity calculations by using multi-feature analysis and machine learning to predict optimal drug dosages, enhancing accuracy and reducing adverse reactions and resource waste.
Patent Information
- Application Number
- CN202411403326.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-09
- Publication Date
- 2024-12-27
AI Technical Summary
Existing patient similarity calculation methods rely heavily on expert knowledge for weight setting and have a single measurement approach, leading to inconsistent and poorly interpretable drug dosages, which can cause adverse reactions and waste resources.
A method for predicting medication dosage based on patient similarity using multi-feature similarity measurement and feature correlation analysis, involving data preprocessing, encoding, pairing, and machine learning models to select the most similar historical patient data for training, with specific encoding techniques for different data types and correlation calculations.
Improves medical efficiency by matching patients with optimal medication dosages, reducing adverse reactions, and optimizing resource use through standardized data processing and accurate prediction.
Smart Images
Figure CN119207831A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of drug dosage prediction, and in particular relates to a drug dosage prediction method based on patient similarity. Background Art
[0002] In recent years, with the continuous emergence of new drugs, the number of clinical drugs has continued to increase. Faced with a wide range of drugs, medical staff cannot understand and master comprehensive drug information, and clinical pharmacists cannot be very familiar with the application of each type of drug. As a result, when clinicians choose drug dosages, it is inevitable that some under- or over-dosage will occur. At the same time, due to the uneven level of doctors, it is easy to match different drug dosages for the same patient. This irrational use of drugs will not only delay or even fail treatment, cause adverse reactions and drug-induced diseases, and harm patients, but also damage the reputation of medical institutions and doctors, and even waste social resources and increase the economic burden on patients. At present, drug prediction mostly uses patient similarity to achieve individualized diagnosis and treatment of patients in order to improve diagnostic accuracy. However, the existing patient similarity calculation method has the problem of a single measurement method and excessive reliance on expert knowledge in weight setting. Therefore, solving the problems caused by irrational drug administration and dosage, improving patient similarity matching, reducing the model's reliance on expert knowledge, and improving the interpretability of results are of great significance to promoting the construction of smart medical care. Summary of the invention
[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method for predicting medication dosage based on patient similarity, which solves the problem that the existing patient similarity calculation method has a single measurement method and the weight setting is too dependent on expert knowledge.
[0004] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a method for predicting medication dosage based on patient similarity, comprising the following steps:
[0005] S1: Obtain the real clinical historical data of historical patients and perform preprocessing to obtain the processed historical patient data set;
[0006] S2: Encode different types of features in the historical patient data set in a corresponding manner to obtain the representation data of the historical patients;
[0007] S3: obtaining the data of the current patient, and preprocessing and encoding the data in a corresponding manner according to different types of features in the data of the current patient, to obtain the representation data of the current patient;
[0008] S4: Based on the characterization data of historical patients and the characterization data of current patients, the similarity ranking of historical patients and current patients in the historical patient dataset is calculated using a multi-feature similarity measurement strategy and feature correlation analysis;
[0009] S5: Select the historical patient data with the highest similarity and a fixed percentage in the similarity ranking between the historical patients and the current patients as the training set for the current patient, and process it using the machine learning model to obtain the predicted result of the current patient's medication dosage.
[0010] The beneficial effects of the present invention are as follows: in the present invention, by performing standardized processing and detailed analysis on the patient's data, the most similar patient group is matched for each patient in terms of characteristics and comprehensive patient levels. By selecting some of the most similar patient groups as training sets and using machine learning models to predict the patient's optimal medication dosage, medical efficiency can be improved, the problems of poor consistency and poor interpretability of clinical medication dosage can be alleviated, adverse reactions can be prevented or mitigated, and patients can obtain the best treatment effect with the least expenditure, effectively utilizing medical resources and reducing waste.
[0011] Further: the encoding of different types of features in the historical patient data set in a corresponding manner includes:
[0012] For the multivariate data in the historical patient dataset, one-hot encoding is used for encoding;
[0013] For the continuous data in the historical patient dataset, scaling and standardization are used for encoding to map the continuous data to data ranging from 0 to 1.
[0014] The beneficial effect of the above further scheme is: converting the data in the historical patient dataset into a unified format to facilitate subsequent processing, while improving the efficiency and generalization of drug dosage prediction.
[0015] Further: the specific steps of S4 are as follows:
[0016] S401: Pair the current patient with each of the historical patients in pairs to obtain a pairing combination of each of the historical patients and the current patient;
[0017] S402: Calculate the similarity value of each characterization data in each pairing combination of the historical patient and the current patient according to the characterization data of the historical patient and the characterization data of the current patient;
[0018] S403: according to the historical patient characterization data and the current patient characterization data, using feature correlation analysis, calculate the correlation between each characterization data and the medication dosage, and use a softmax function to convert the correlation between each characterization data and the medication dosage into a weight of each characterization data and the medication dosage;
[0019] S404: Multiply the similarity value of each characterization data in the pairing combination of each historical patient and the current patient by the weight of the characterization data and the medication dosage, and add the multiplication results to obtain the comprehensive similarity value of each historical patient and the current patient, and sort them to obtain the similarity ranking of historical patients and current patients in the historical patient data set.
[0020] The beneficial effect of the above further scheme is: by calculating the similarity between historical patients and current patients, a comprehensive and accurate prediction result can be provided for the current patient, thereby improving medical efficiency and the safety of patient medication.
[0021] Further: the similarity value of each characterization data in the pairing combination of each historical patient and the current patient is calculated in S402, and each characterization data includes:
[0022] Binary representation data: When the feature values of the binary representation data of the historical patient and the current patient are equal, the similarity value is set to 1, otherwise it is set to 0;
[0023] Multivariate representation data: When the eigenvalues of the multivariate representation data of the historical patient and the current patient are equal, the similarity value is set to 1. Otherwise, the smaller eigenvalue is used as the numerator and the larger eigenvalue is used as the denominator. The fractional value is the similarity value of the multivariate representation data.
[0024] Continuous representation data: The cosine similarity is used to calculate the similarity value of the continuous representation data.
[0025] The beneficial effect of the above further scheme is that for different types of characterization data, different methods are used to calculate similarity values, which can enhance adaptability, predict the dosage of medication based on more comprehensive data, and improve the accuracy of the prediction.
[0026] Further: the calculation expression of the binary characterization data is as follows:
[0027]
[0028] in, is the similarity value of the binary feature data of patient a and patient b, a is patient a, b is patient b, S a is the eigenvalue of the binary representation data of patient a, S b is the eigenvalue of the binary characterization data of patient b.
[0029] The beneficial effect of the above further scheme is: using the above formula to calculate the binary characterization data can increase the reliability of the calculation results of the binary characterization data and ensure the accuracy and reliability of the prediction results.
[0030] Further: The calculation expression of the multivariate characterization data is as follows:
[0031]
[0032] in, is the similarity value of patient a and patient b in the multivariate representation data, min(,) is the smaller of the two data, max(,) is the larger of the two data, C a is the multivariate representation data of patient a, C b is the multivariate characterization data of patient b.
[0033] The beneficial effect of the above further scheme is: using the above formula to calculate the multivariate characterization data can increase the reliability of the calculation results of the multivariate characterization data and ensure the accuracy and reliability of the prediction results.
[0034] Further: the similarity value of the continuous characterization data is obtained by using the cosine similarity calculation, and its expression is as follows:
[0035]
[0036] in, is the similarity value of the continuous characterization data between patient a and patient b, i is the sequence number of the characterization data, n is the number of continuous characterization data, is the i-th data of the continuous representation data of patient a, is the i-th data of the continuous representation data of patient b, ∑ is the summation symbol, and i is the sequence number of the data in the continuous representation data.
[0037] The beneficial effect of the above further scheme is: by using cosine similarity to calculate the similarity value of continuous characterization data alone, the efficiency of the calculation and the accuracy of the calculation results can be ensured, thereby improving the accuracy and reliability of the prediction.
[0038] Further: In S403, the correlation between each characterization data and the dosage is calculated by using feature correlation analysis, and the expression is as follows:
[0039]
[0040] Among them, r is the Pearson correlation coefficient, X j is the jth eigenvalue representing the data, Y j is the dosage of the jth characterization data, and are the average values of X and Y respectively, j is the sequence number of the data, and m is the number of data.
[0041] The beneficial effect of the above further solution is that by calculating the Pearson correlation coefficient, the linear strength and direction of the characteristic data and the medication dosage can be intuitively indicated. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 The figure is a flow chart of a method for predicting medication dosage based on patient similarity. DETAILED DESCRIPTION
[0043] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.
[0044] like Figure 1 As shown, a method for predicting medication dosage based on patient similarity includes the following steps:
[0045] S1: Obtain the real clinical historical data of historical patients and perform preprocessing to obtain the processed historical patient data set;
[0046] S2: Encode different types of features in the historical patient data set in a corresponding manner to obtain the representation data of the historical patients;
[0047] S3: obtaining the data of the current patient, and preprocessing and encoding the data in a corresponding manner according to different types of features in the data of the current patient, to obtain the representation data of the current patient;
[0048] S4: Based on the characterization data of historical patients and the characterization data of current patients, the similarity ranking of historical patients and current patients in the historical patient dataset is calculated using a multi-feature similarity measurement strategy and feature correlation analysis;
[0049] S5: Select the historical patient data with the highest similarity and a fixed percentage in the similarity ranking between the historical patients and the current patients as the training set for the current patient, and process it using the machine learning model to obtain the predicted result of the current patient's medication dosage.
[0050] In one embodiment of the present invention, the real clinical historical data of historical patients is obtained in S1 and preprocessed. The specific steps of the preprocessing are as follows:
[0051] A1: Obtain the real clinical historical data of historical patients and delete the data with missing important features in the real clinical historical data; it can solve the problems of many missing values and irregular data records in the real clinical historical data;
[0052] A2: Clean up the special characters in the real clinical historical data and delete unnecessary characters;
[0053] A3: Checking and correcting the information in the real clinical historical data can solve the errors caused by manual data entry, such as typing a decimal point as a period, such as "54.2kg" is mistakenly entered as "54.2kg";
[0054] A4: Use the IQR quartile method to perform outlier analysis on the continuous values of real clinical data and delete outlier sample data, that is, delete abnormal values;
[0055] A5: The numerical units and numerical types of each data in the real clinical data are standardized and limited to obtain the processed historical patient data set; in this embodiment, the numerical units of the data are standardized, and the weight is recorded in kg; the numerical types of the data are standardized, and the age is mapped from a specific numerical value to an age category.
[0056] In one embodiment of the present invention, different types of features in the historical patient data set are encoded in S2 in a corresponding manner, and the specific encoding methods include:
[0057] For the multivariate data in the historical patient data set, one-hot encoding is used to encode and convert it into digital data. In this embodiment, for the age data in the historical patient data set, one-hot encoding is used to obtain children [0, 0, 1], adults [0, 1, 0], and the elderly [1, 0, 0];
[0058] For the continuous data in the historical patient dataset, scaling and standardization are used for encoding to map the continuous data to data ranging from 0 to 1.
[0059] In one embodiment of the present invention, the specific steps of S4 are as follows:
[0060] S401: Pair the current patient with each of the historical patients in pairs to obtain a pairing combination of each of the historical patients and the current patient;
[0061] S402: Calculate the similarity value of each characterization data in each pairing combination of the historical patient and the current patient according to the characterization data of the historical patient and the characterization data of the current patient;
[0062] S403: according to the historical patient characterization data and the current patient characterization data, using feature correlation analysis, calculate the correlation between each characterization data and the medication dosage, and use a softmax function to convert the correlation between each characterization data and the medication dosage into a weight of each characterization data and the medication dosage;
[0063] S404: Multiply the similarity value of each characterization data in the pairing combination of each historical patient and the current patient by the weight of the characterization data and the medication dosage, and add the multiplication results to obtain the comprehensive similarity value of each historical patient and the current patient, and sort them to obtain the similarity ranking of historical patients and current patients in the historical patient data set.
[0064] In S402, the similarity value of each characterization data in the paired combination of each historical patient and the current patient is calculated, and each characterization data includes:
[0065] Binary representation data: When the feature values of the binary representation data of the historical patient and the current patient are equal, the similarity value is set to 1, otherwise it is set to 0;
[0066] Multivariate representation data: When the eigenvalues of the multivariate representation data of the historical patient and the current patient are equal, the similarity value is set to 1. Otherwise, the smaller eigenvalue is used as the numerator and the larger eigenvalue is used as the denominator. The fractional value is the similarity value of the multivariate representation data.
[0067] Continuous representation data: The cosine similarity is used to calculate the similarity value of the continuous representation data.
[0068] In this embodiment, the binary representation data includes clinical data such as the patient's gender, and the calculation expression of the binary representation data is as follows:
[0069]
[0070] in, is the similarity value of the binary feature data of patient a and patient b, a is patient a, b is patient b, S a is the eigenvalue of the binary representation data of patient a, S b is the feature value of the binary representation data of patient b;
[0071] In this embodiment, the multivariate representation data includes clinical data such as the patient's examination site, disease, and symptoms. The calculation expression of the multivariate representation data is as follows:
[0072]
[0073] in, is the similarity value of patient a and patient b in the multivariate representation data, min(,) is the smaller of the two data, max(,) is the larger of the two data, C a is the multivariate representation data of patient a, C b is the multivariate representation data of patient b;
[0074] In this embodiment, the continuous characterization data includes clinical data such as the patient's height and weight, which are encoded by scaling and standardization to obtain data from 0 to 1. The similarity value of the continuous characterization data is calculated by cosine similarity, and its expression is as follows:
[0075]
[0076] in, is the similarity value of the continuous characterization data between patient a and patient b, i is the sequence number of the characterization data, n is the number of continuous characterization data, is the i-th data of the continuous representation data of patient a, is the i-th data of the continuous representation data of patient b, ∑ is the summation symbol, and i is the sequence number of the data in the continuous representation data.
[0077] Among them, in S403, the correlation between each characterization data and the dosage is calculated by using feature correlation analysis, and its expression is as follows:
[0078]
[0079] Among them, r is the Pearson correlation coefficient, X j is the jth eigenvalue representing the data, Y j is the dosage of the jth characterization data, and are the average values of X and Y respectively, j is the sequence number of the data, and m is the number of data.
[0080] In one embodiment of the present invention, S5 selects data of historical patients with high similarity and a specific percentage as a training set for the current patient, and processes them using a machine learning model. Due to the complexity and processing rate of the data, a highly efficient integrated machine learning model is selected, such as a random forest model and an XGBoost extreme gradient boosting model, etc.; taking the XGBoost extreme gradient boosting model as an example, the selected similar historical patient data is used as a training set, and the labeled data is the actual injection dose of similar historical patients. After completing the model training, a trained model is obtained, and the current patient information is used as input, and the model outputs the predicted medication dose of the current patient.
[0081] In one embodiment of the present invention, after S5 obtains the predicted value of the medication dosage, the SHAP interpretable machine learning model is used to perform feature attribution interpretation of the predicted value; the SHAP interpretable machine learning model is a method for explaining the prediction results of a machine learning model, which provides an interpretable framework to explain the decision-making process of the model by calculating the contribution of each feature to the prediction result. The core idea is to decompose the prediction result into the contribution of each feature value based on the concept of Shapley value in game theory.
[0082] Based on the SHAP interpretable machine learning model, the feature subsets of the current patient are arranged and combined, the contribution value of each feature value in different subsets is calculated, and then the weighted sum is performed based on these contribution values to obtain the overall contribution of each feature. By calculating the contribution of the feature to the prediction result, the SHAP interpretable machine learning model can provide the relative importance ranking of each feature and visualize the contribution value of the feature, so as to better understand the prediction results of the model;
[0083] The beneficial effects of the present invention are as follows: in the present invention, by performing standardized processing and detailed analysis on the patient's data, the most similar patient group is matched for each patient in terms of characteristics and comprehensive patient levels. By selecting some of the most similar patient groups as training sets and using machine learning models to predict the patient's optimal medication dosage, medical efficiency can be improved, the problems of poor consistency and poor interpretability of clinical medication dosage can be alleviated, adverse reactions can be prevented or mitigated, and patients can obtain the best treatment effect with the least expenditure, effectively utilizing medical resources and reducing waste.
Claims
1. A method for predicting medication dosage based on patient similarity, characterized in that: The following steps are involved: S1: Obtain the real clinical historical data of historical patients and perform preprocessing to obtain the processed historical patient data set; S2: Encode different types of features in the historical patient data set in a corresponding manner to obtain the representation data of the historical patients; S3: obtaining the data of the current patient, and preprocessing and encoding the data in a corresponding manner according to different types of features in the data of the current patient, to obtain the representation data of the current patient; S4: Based on the characterization data of historical patients and the characterization data of current patients, the similarity ranking of historical patients and current patients in the historical patient dataset is calculated using a multi-feature similarity measurement strategy and feature correlation analysis; S5: Select the historical patient data with the highest similarity and a fixed percentage in the similarity ranking between the historical patient and the current patient as the training set for the current patient, and process it using the machine learning model to obtain the predicted result of the current patient's medication dosage.
2. The method for predicting medication dosage based on patient similarity according to claim 1, characterized in that: The different types of features in the historical patient data set are encoded in a corresponding manner, including: For the multivariate data in the historical patient dataset, one-hot encoding is used for encoding; For the continuous data in the historical patient dataset, scaling and standardization are used for encoding to map the continuous data to data ranging from 0 to 1.
3. The method for predicting medication dosage based on patient similarity according to claim 1, characterized in that: The specific steps of S4 are as follows: S401: Pair the current patient with each of the historical patients in pairs to obtain a pairing combination of each of the historical patients and the current patient; S402: Calculate the similarity value of each characterization data in each pairing combination of the historical patient and the current patient according to the characterization data of the historical patient and the characterization data of the current patient; S403: according to the historical patient characterization data and the current patient characterization data, using feature correlation analysis, calculate the correlation between each characterization data and the medication dosage, and use a softmax function to convert the correlation between each characterization data and the medication dosage into a weight of each characterization data and the medication dosage; S404: Multiply the similarity value of each characterization data in the pairing combination of each historical patient and the current patient by the weight of the characterization data and the medication dosage, and add the multiplication results to obtain the comprehensive similarity value of each historical patient and the current patient, and sort them to obtain the similarity ranking of historical patients and current patients in the historical patient data set.
4. The method for predicting medication dosage based on patient similarity according to claim 3, characterized in that: In S402, the similarity value of each characterization data in the paired combination of each historical patient and the current patient is calculated, and each characterization data includes: Binary representation data: When the feature values of the binary representation data of the historical patient and the current patient are equal, the similarity value is set to 1, otherwise it is set to 0; Multivariate representation data: When the eigenvalues of the multivariate representation data of the historical patient and the current patient are equal, the similarity value is set to 1. Otherwise, the smaller eigenvalue is used as the numerator and the larger eigenvalue is used as the denominator. The fractional value is the similarity value of the multivariate representation data. Continuous representation data: The cosine similarity is used to calculate the similarity value of the continuous representation data.
5. The method for predicting medication dosage based on patient similarity according to claim 4, characterized in that: The calculation expression of the binary characterization data is as follows: in, is the similarity value of the binary feature data of patient a and patient b, a is patient a, b is patient b, S a is the eigenvalue of the binary representation data of patient a, S b is the eigenvalue of the binary characterization data of patient b.
6. The method for predicting medication dosage based on patient similarity according to claim 4, characterized in that: The calculation expression of the multivariate characterization data is as follows: in, is the similarity value of patient a and patient b in the multivariate representation data, min(,) is the smaller of the two data, max(,) is the larger of the two data, C a is the multivariate representation data of patient a, C b is the multivariate characterization data of patient b.
7. The method for predicting medication dosage based on patient similarity according to claim 4, characterized in that: The similarity value of the continuous characterization data is obtained by using the cosine similarity calculation, and its expression is as follows: in, is the similarity value of the continuous characterization data between patient a and patient b, i is the sequence number of the characterization data, n is the number of continuous characterization data, is the i-th data of the continuous representation data of patient a, is the i-th data of the continuous representation data of patient b, ∑ is the summation symbol, and i is the sequence number of the data in the continuous representation data.
8. The method for predicting medication dosage based on patient similarity according to claim 3, characterized in that: In S403, the correlation between each characterization data and the dosage is calculated by using feature correlation analysis, and the expression is as follows: Among them, r is the Pearson correlation coefficient, X j is the jth eigenvalue representing the data, Y j is the dosage of the jth characterization data, and are the average values of X and Y respectively, j is the sequence number of the data, and m is the number of data.
Citation Information
Cited By
Method for generating pediatric surgery appendix postoperative anti-infection scheme
CN120833922A