An antibiotic combination recommendation model and its application

Abstract

The present invention discloses an antibiotic combination recommendation model and its application. The model is constructed through the following training process: S1, obtaining different antibiotic combination historical data sets, and extracting the historical compatibility frequency of any antibiotic combination in the historical data set, as well as the historical effectiveness after the combination is used; S2, based on the historical compatibility frequency of any of the aforementioned antibiotic combinations and its historical effectiveness, determining the recommended type of the antibiotic combination combination; S3, using the molecular fingerprint information of each antibiotic in the aforementioned antibiotic combination as input data, and its recommended type as output data, and performing classification learning through a fully connected deep neural network to obtain the antibiotic combination recommendation model. The model realizes intelligent recommendation of antibiotic combination by combining drug feature information with clinical drug feature information, thereby improving the safety and effectiveness of patients' combination medication.

Description

Technical Field

[0001] The present invention relates to the technical field of antibiotic combination therapy, and in particular to an antibiotic combination therapy recommendation model and its application. Furthermore, the present invention provides an antibiotic combination therapy recommendation model, and an evaluation method for the recommended type of antibiotic combination therapy in clinical practice using the antibiotic combination therapy recommendation model. Background Art

[0002] The problem of antibiotic combination therapy is often faced in clinical anti-infection treatment. The purpose of clinical antibiotic combination therapy is mainly to treat complex mixed infections, expand the antibacterial spectrum, reduce the occurrence of drug resistance and alleviate toxic and side effects. However, in actual application, complex disease conditions and a wide variety of antibiotic drugs often make it difficult for clinical workers to quickly screen and confirm effective antibiotic combinations. In fact, in the past few years, 30% of the new antibiotics approved by the U.S. Food and Drug Administration (FDA) are combination antibiotics.

[0003] At present, most of the clinical studies examining the efficacy of antibiotic combination therapy are systematic clinical studies, and there is still a lack of relevant feasibility evaluation technology to guide clinical workers in conducting antibiotic combination therapy. Therefore, it is difficult to accurately evaluate the rationality and feasibility of clinical antibiotic combination therapy. Summary of the invention

[0004] The purpose of this application is to solve the problem in the prior art that the rationality and feasibility of clinical antibiotic combination therapy are difficult to accurately evaluate, and to provide an antibiotic combination therapy recommendation model and its application.

[0005] In order to achieve the above invention purpose, the technical solution adopted in this application is:

[0006] In a first aspect, the present application provides an antibiotic combination recommendation model, which is constructed through the following training process:

[0007] S1. Obtain a historical data set of different antibiotic combination medications, and extract the historical compatibility frequency of any antibiotic combination in the historical data set, as well as the historical effectiveness rate after the combination is used;

[0008] S2. Based on the historical compatibility frequency and historical effectiveness of any of the aforementioned antibiotic combinations, determine the recommended type of combined medication for the antibiotic combination;

[0009] S3. Using the molecular fingerprint information of each antibiotic in the aforementioned antibiotic combination as input data and its recommended type as output data, classification learning is performed through a fully connected deep neural network to obtain the antibiotic combination recommendation model.

[0010] In an optional embodiment, the historical data set includes a clinical patient medical record data set that has received antibiotic treatment and has undergone at least two infection-related examinations.

[0011] In an optional embodiment, the clinical patients do not include patients who have received immunointervention therapy and suffer from immune disorders, or patients who have only received one antibiotic treatment, or patients who have not received combined antibiotic treatment.

[0012] In an optional embodiment, the infection-related examination items include C-reactive protein examination and procalcitonin examination; the determination process of the historical effectiveness rate in step S1 includes:

[0013] For any antibiotic combination, the historical effectiveness of the antibiotic combination after compatible use is determined based on the test results of the C-reactive protein test and the test results of the procalcitonin test corresponding to the antibiotic combination.

[0014] In an optional embodiment, in step S1, after obtaining the antibiotic combination medication history data set, the step further includes:

[0015] For any antibiotic in the antibiotic combination medication history data set, perform generic name unification processing to obtain the standard name of each antibiotic;

[0016] Based on the standard name of each antibiotic, the molecular fingerprint information of each antibiotic in the antibiotic combination medication history dataset is determined.

[0017] In an optional embodiment, the process of determining the molecular fingerprint information includes:

[0018] Determining the chemical structural formula of each antibiotic based on the standard name of each antibiotic;

[0019] Based on the chemical structural formula of each antibiotic, the MACCS key of each antibiotic is determined using the MACCS molecular access system as the molecular fingerprint information of each antibiotic.

[0020] In an optional embodiment, the process of determining the recommended type to which any antibiotic combination belongs in step S2 includes:

[0021] Based on the historical efficacy, the antibiotic combination is divided into a strong antibiotic combination and a weak antibiotic combination, wherein the historical efficacy of all antibiotic groups in the strong antibiotic combination is not less than 50%, and the historical efficacy of all antibiotic groups in the weak antibiotic combination is less than 50%;

[0022] Based on the historical compatibility frequency, the antibiotic combinations in the strong antibiotic combination and the weak antibiotic combination are sorted by compatibility frequency in descending order of the historical compatibility frequency;

[0023] The recommendation type to which the antibiotic combinations whose compatibility frequency is ranked in the top 16% among the strong antibiotic combinations is determined as a strongly recommended type, the recommendation type to which the antibiotic combinations whose compatibility frequency is ranked in the bottom 16% among the weak antibiotic combinations is determined as a not recommended type, and the recommendation type to which the remaining antibiotic combinations among the strong antibiotic combinations and the weak antibiotic combinations are determined as a weakly recommended type.

[0024] In a second aspect, the present application provides a method for evaluating the recommended types of clinical antibiotic combination therapy, and uses the above-mentioned antibiotic combination therapy recommendation model to evaluate the recommended types of clinical antibiotic combination therapy.

[0025] In an optional embodiment, the evaluation method comprises the following steps:

[0026] Obtaining the standard name of each antibiotic in the antibiotic combination to be recommended, and determining the molecular fingerprint information of each antibiotic based on the standard name of each antibiotic;

[0027] Based on the molecular fingerprint information of each antibiotic, using the above-mentioned antibiotic combination recommendation model, determine the probability score of the clinical recommendation type of the antibiotic combination to be recommended;

[0028] Determining a recommendation type corresponding to the antibiotic combination to be recommended based on the probability score of the clinical recommendation type of the antibiotic combination to be recommended;

[0029] The recommendation types include a non-recommended type, a weakly recommended type and a strongly recommended type.

[0030] In an optional embodiment, determining the recommendation type corresponding to the antibiotic combination to be recommended based on the probability score of the clinical recommendation type of the antibiotic combination to be recommended includes:

[0031] The highest probability score among the probability scores of the clinical recommendation types of the antibiotic combination to be recommended is determined, and the recommendation type corresponding to the highest probability score is determined as the recommendation type corresponding to the antibiotic combination to be recommended.

[0032] The present invention discloses an intelligent recommendation method and system for antibiotic combination therapy based on clinical medical record big data analysis. By combining drug feature information with clinical medication feature information, intelligent recommendation for antibiotic combination therapy is achieved, thereby improving the safety and effectiveness of patient combination therapy.

[0033] The intelligent recommendation method and system for antibiotic combination based on clinical medical record big data analysis disclosed in the present application integrates large-scale antibiotic combination prescription data, combines clinical prescription compatibility frequency and efficiency information to achieve clinical grading decision-making, uses DNN framework for learning and modeling, realizes intelligent recommendation of antibiotic combination, and puts forward feasible decision-making opinions for antibiotic combination methods that lack clinical experience guidance. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A flowchart schematically illustrates a training process of an antibiotic combination medication recommendation model provided in an embodiment of the present application;

[0035] Figure 2 A flowchart schematically shows a method for evaluating the recommended type of clinical antibiotic combination therapy provided by an embodiment of the present application;

[0036] Figure 3 ROC and PRC curves of the test set data in the antibiotic combination recommendation system;

[0037] Figure 4 This is a test result graph of the accuracy of the recommendation of the antibiotic combination recommendation system of the present application based on retrospective clinical data. DETAILED DESCRIPTION

[0038] In order to further explain the technical means and results taken by the present invention to achieve the predetermined invention purpose, the specific implementation methods, technical solutions and features of the present invention application are described in detail below with preferred embodiments. The specific features, structures, or characteristics of the multiple embodiments in the following description can be combined in any suitable form.

[0039] The present invention provides an intelligent recommendation method and system for antibiotic combination based on clinical medical record big data analysis. First, clinical medical record information is collected for antibiotic combination, antibiotic naming is unified according to antibiotic common name information, non-combination medical records are eliminated, and antibiotic combination effectiveness statistics and prescription frequency statistics are performed in combination with antibiotic prescription information and patient infection index monitoring information to achieve standardization of heterogeneous data and obtain an antibiotic combination clinical feature database. Afterwards, the antibiotic combination grading threshold is optimized in combination with effectiveness statistics and prescription frequency indicators, and classification modeling is performed using a DNN algorithm to obtain a stable antibiotic combination grading decision scheme. Finally, the structural characteristics of antibiotics are discretely encoded using the MACCS method, a three-classification model is built using a DNN framework, and model training is performed in combination with antibiotic combination clinical grading decision labels to form an antibiotic combination intelligent recommendation system with an antibiotic input terminal and a three-classification recommendation probability score output neuron. The accuracy of the antibiotic combination intelligent recommendation system is evaluated based on new clinical retrospective cohort data.

[0040] The present application embodiment provides a recommendation model for antibiotic combination therapy. Figure 1 The flowchart of the training process of the recommendation model is schematically shown. Figure 1 As shown in Figure 1, the model is constructed through the following training process:

[0041] S101. Obtain a historical data set of combined use of different antibiotics, and extract the historical compatibility frequency of any antibiotic combination in the historical data set, as well as the historical effectiveness rate after the combined use;

[0042] S102. Determine the recommended type of combined medication for any of the aforementioned antibiotic combinations based on their historical compatibility frequency and historical effectiveness;

[0043] S103, using the molecular fingerprint information of each antibiotic in the aforementioned antibiotic combination as input data, and its recommended type as output data, performing classification learning through a fully connected deep neural network to obtain the antibiotic combination recommendation model.

[0044] In step S101, the historical data set may include a medical record data set of clinical patients who have received antibiotic treatment and have undergone at least two infection-related examinations.

[0045] In an optional embodiment, the clinical patients do not include patients who have received immunointervention therapy and suffer from immune disorders, or patients who have only received one antibiotic treatment, or patients who have not received combined antibiotic treatment.

[0046] In the form of pairwise compatibility of antibiotics, the historical compatibility frequency and historical effectiveness of antibiotic combination prescriptions are counted to achieve heterogeneous feature acquisition of medical record information of antibiotic combination in different clinical centers. The clinical medication characteristics of antibiotic combinations are described based on the historical prescription compatibility frequency and the historical prescription effectiveness, and used as the data source for recommendation grading and antibiotic combination recommendation models. The characteristic data of antibiotic combination use include specific antibiotic combinations, prescription frequency and effectiveness index. The prescription frequency reflects the clinical medication habits of the antibiotic combination, and the effectiveness takes into account the efficacy of the combination in clinical practice.

[0047] In an optional embodiment, the infection-related examination items include C-reactive protein examination and procalcitonin examination; the historical effectiveness determination process in step S1 includes: for any antibiotic combination, based on the examination results of the C-reactive protein examination and the procalcitonin examination corresponding to the antibiotic combination, determining the historical effectiveness of the antibiotic combination after use.

[0048] In an optional embodiment, in step S1, after obtaining the antibiotic combination medication history data set, it also includes: for any antibiotic in the antibiotic combination medication history data set, performing common name unification processing to obtain the standard name of each antibiotic; based on the standard name of each antibiotic, determining the molecular fingerprint information of each antibiotic in the antibiotic combination medication history data set.

[0049] In an optional embodiment, the process of determining the molecular fingerprint information includes: determining the chemical structure formula of each antibiotic based on the standard name of each antibiotic; based on the chemical structure formula of each antibiotic, using the MACCS molecular access system to determine the MACCS key of each antibiotic as the molecular fingerprint information of each antibiotic.

[0050] MACCS is used to construct molecular fingerprint information of antibiotics based on their substructures. MACCS contains a subset of 166 bonds. Substructures are a subset associated with chemical structures. They can be represented by a bit array of 0 and 1, where each bit position represents the presence or absence of a structural fragment. Ultimately, each antibiotic can be represented as a binary vector based on its substructure:

[0051]

[0052] in, v n is a binary value and n is the number of substructures.

[0053] In step S102, when determining the recommended type of combined use of the antibiotic combination based on the historical compatibility frequency and historical effectiveness of any of the aforementioned antibiotic combinations, the historical prescription compatibility frequency is first fixed to the median, and the samples are divided into prescription effectiveness levels in units of 0.1. By using a fully connected deep neural network (DNN) for classification learning, with the stability of the prediction results as the evaluation index, the classification effect is best when the prescription effectiveness threshold is set to 50%. After the prescription effectiveness threshold is set to 50%, the antibiotic combinations are sorted according to their prescription compatibility frequencies in each type of sample, with a step size of 1, to improve the hierarchical decision-making scheme. The antibiotic prescription compatibility frequency and effectiveness index are used as the basis for antibiotic recommendation grading, and the optimal threshold range for the three categories of strong recommendation, weak recommendation, and non-recommendation is obtained.

[0054] Exemplarily, the recommended type of antibiotic combination finally determined may include: Based on the historical effectiveness, the antibiotic combination is divided into a strong antibiotic combination and a weak antibiotic combination, wherein the historical effectiveness of all antibiotic groups in the strong antibiotic combination is not less than 50%, and the historical effectiveness of all antibiotic groups in the weak antibiotic combination is less than 50%; Based on the historical compatibility frequency, the antibiotic combinations in the strong antibiotic combination and the weak antibiotic combination are sorted by compatibility frequency in descending order of the historical compatibility frequency; the recommendation type to which the antibiotic combinations in the strong antibiotic combination whose compatibility frequency is ranked in the top 16% is determined as a strongly recommended type, the recommendation type to which the antibiotic combinations in the weak antibiotic combination whose compatibility frequency is ranked in the bottom 16% is determined as a not recommended type, and the recommendation type to which the remaining antibiotic combinations in the strong antibiotic combination and the weak antibiotic combination belong is determined as a weakly recommended type.

[0055] In step S103, the benchmark dataset can be used for training. The benchmark dataset decision label division includes 55 antibiotics, 657 antibiotic combinations, and three types of labels, namely, 105 Class 1 (not recommended), 447 Class 2 (weakly recommended), and 105 Class 3 (strongly recommended). 80% of the benchmark dataset is used to train the model, and the remaining 20% ​​is used as an independent test dataset.

[0056] Specifically, when using the DNN framework to develop an antibiotic combination recommendation model, the DNN model is designed as a three-classification model with four fully connected hidden layers and a softmax layer with three output neurons, which can predict the recommendation type for each given antibiotic combination. V sub, and concatenate them into a vector, which is fed into the feedforward neural network. The activity values ​​of the output neurons range from 0 to 1 and can be considered as the probability score of belonging to a certain recommendation type. If the value of the corresponding output neuron is greater than the value of any other neuron, then the given antibiotic combination is predicted to be the recommended level of that neuron.

[0057] The standard gradient descent method and Adam optimizer are used for training, and the cross entropy function is used as the loss function for the classification task. Three evaluation indicators are used to evaluate the model construction, including precision, recall rate and F1 score, which are defined as:

[0058]

[0059] Among them, TP is the number of true positives, FN is the number of false negatives, and FP is the number of false positives. Precision describes the proportion of predicted true positive samples to all predicted positive samples, while recall refers to the proportion of predicted true positive samples to all true positive samples. The F1 score can be interpreted as the harmonic mean of precision and recall.

[0060] The embodiment of the present application also provides a method for evaluating the recommended types of clinical antibiotic combination therapy, which uses the above-mentioned antibiotic combination therapy recommendation model to evaluate the recommended types of clinical antibiotic combination therapy. Figure 2 The schematic diagram of the flow chart of the evaluation method provided in this embodiment is shown as follows: Figure 2 As shown, the evaluation method may include the following steps:

[0061] S201, obtaining the standard name of each antibiotic in the antibiotic combination to be recommended, and determining the molecular fingerprint information of each antibiotic based on the standard name of each antibiotic;

[0062] S202, based on the molecular fingerprint information of each antibiotic, using the aforementioned antibiotic combination recommendation model, determining a probability score of the clinical recommendation type of the antibiotic combination to be recommended;

[0063] S203. Determine the recommendation type corresponding to the antibiotic combination to be recommended based on the probability score of the clinical recommendation type of the antibiotic combination to be recommended; wherein the recommendation type includes a not recommended type, a weakly recommended type, and a strongly recommended type.

[0064] In step S201, first, the information of the antibiotic group to be recommended input by the user through an input device such as a mouse, keyboard, voice acquisition device, image acquisition device, handwriting input device, etc. can be obtained, and each antibiotic name included in the antibiotic group information to be recommended can be identified. Exemplarily, an interactive interface can be provided to the user for receiving and uploading the antibiotic group information to be recommended input by the user. The antibiotic name obtained can be a common name or a standard name. In the case where the antibiotic name obtained is a common name, the common name can be converted into a standard name.

[0065] In step S203, determining the recommendation type corresponding to the antibiotic combination to be recommended based on the probability score of the clinical recommendation type of the antibiotic combination to be recommended may include: determining the highest probability score among the probability scores of the clinical recommendation types of the antibiotic combination to be recommended, and determining the recommendation type corresponding to the highest probability score as the recommendation type corresponding to the antibiotic combination to be recommended.

[0066] The embodiment of the present application also provides a method for constructing a database of antibiotic combination drug characteristics based on multi-center heterogeneous medical record data, the method comprising: performing heterogeneous feature extraction operations from clinical medical record data and medication order data, comprising the following steps: unifying medication order information with the generic name of the antibiotic drug, and using this as a keyword to search in medication prescriptions and orders, and identifying medical records in which antibiotic combination drugs are used; using inflammatory indicators (C-reactive protein, procalcitonin, white blood cell count and body temperature) as keywords, searching in medical record information, and extracting medical records in which antibiotic combination drugs are used twice or more during the period of antibiotic combination drug use; Medical records of examination items on the list; extracting patients' inflammatory index information from medical record data as a feature for identifying the effectiveness of combination antibiotics, and using it as one of the features for recommending combination antibiotics; extracting medication order information from medical record data, using the generic name of antibiotics as a search keyword, searching for medication order information, and counting the frequency of prescriptions for combination medications, and using this as one of the features for recommending combination antibiotics; using the clinical effectiveness rate and prescription frequency of combination antibiotics as key indicators, achieving standardized representation of clinical heterogeneity data for combination antibiotics, and using this as the basis for graded recommendations of antibiotics.

[0067] Based on the above-mentioned antibiotic combination feature database information based on medical record data, a clinical grading decision-making scheme for antibiotic combination was established: the effectiveness of antibiotic prescriptions was expressed as a percentage (100% × number of effective cases / number of ineffective cases), and the samples were divided into prescription effectiveness levels in units of a certain step size, and the prediction result stability index was used as the basis to find the optimal threshold range for classification effect; after the prescription effectiveness threshold was optimized, the antibiotic combinations were sorted according to their prescription usage frequency in various samples, and the optimal threshold range for classification effect was found based on a step size of 1 and the prediction result stability index; the antibiotic prescription frequency and effectiveness index were used as the basis for antibiotic decision grading to obtain the optimal threshold range for the three categories of strong recommendation, weak recommendation and no recommendation.

[0068] Based on the above-mentioned clinical feature database of antibiotic combination and decision-making classification labels, an intelligent decision-making system for antibiotic combination based on clinical medical record big data analysis was constructed: MACCS was used to encode the substructure of antibiotic molecules, and the presence or absence of structural fragments at each position was represented by a bit array of 0 and 1. The antibiotic molecular fingerprint was encoded based on the binary vector of the substructure; the DNN framework was used to develop an antibiotic combination decision-making model, which was designed as a three-classification model with four fully connected hidden layers and a softmax layer with three output neurons, which can predict the intelligent recommendation for each given antibiotic combination.

[0069] The intelligent decision-making system for antibiotic combination based on clinical medical record big data analysis includes: the input antibiotic drug combination is compared with the drug information in the external database to realize automatic modeling of antibiotic drug molecules; the clinical data benchmark database information is compared through a feedforward neural network, and the probability score of the corresponding graded output neuron is output; the predicted probability scores of the three output neurons are combined to compare and select the recommended grade with the highest probability value, highlight it, and indicate the reasonable recommendation level of the antibiotic combination; the antibiotic combination recommendation system based on medical record data has the ability to read antibiotic drug combination information, perform drug fingerprint modeling, perform model calculation through a feedforward neural network, and finally output antibiotic combination decision information.

[0070] In order to verify the rationality of the antibiotic combination recommendation system of this application, the performance of the antibiotic combination recommendation method based on clinical big data was evaluated. A 10-fold cross validation (10-CV) was performed on the training data set to find the optimal hyperparameters, and then the performance of ACRM was evaluated on an independent test data set.

[0071] Figure 3The ROC and PRC curves of the test dataset on the antibiotic combination recommendation system are as follows: The receiver operating characteristic curve (ROC) and precision-recall curve (PRC) are used to measure the predictive performance of the model. The area under the ROC curve (AUROC) is equal to the probability that the classifier ranks a randomly selected positive instance higher than a randomly selected negative instance; PRC shows the trade-off between precision and recall at different thresholds. The average AUROC is used as the main indicator to evaluate the overall performance. The closer the AUROC value is to 1, the better the overall performance.

[0072] On the 10-CV training set, the macro-average AUROC (macro-AUROC), micro-average AUROC (micro-AUROC), macro-average AUPRC (macro-AUPRC), and micro-average AUPRC (micro-AUPRC) of ACRM are 0.830, 0.902, 0.701, and 0.832, respectively. In the test set, the macro-average AUROC, micro-average AUROC, macro-average AUPRC, and micro-average AUPRC are 0.767, 0.862, 0.609, and 0.776, respectively. Three evaluation indicators are used, including precision, recall, and F1 score, and the results are shown in Table 1.

[0073]

[0074] Precision describes the proportion of predicted true positive samples to all predicted positive samples, while recall refers to the proportion of predicted true positive samples to all true positive samples. The F1 score can be interpreted as the harmonic mean of precision and recall. The AUROCs for not recommending (category 1), weakly recommending (category 2), and strongly recommending (category 3) are 0.820, 0.589, and 0.895, respectively.

[0075] Figure 4 The test result diagram for testing the recommendation accuracy of the antibiotic combination recommendation system of this application based on retrospective clinical data is as follows:

[0076] A new retrospective study cohort was established according to the above clinical data collection method. The new clinical antibiotic combination cohort included 383 patients who received antibiotic combination therapy, and the frequency of prescription and effectiveness of combination therapy were counted. According to the clinical antibiotic combination recommendation method, the obtained antibiotic combination therapy combinations were graded according to the prescription frequency and effectiveness. The combinations with an effectiveness rate higher than 50% and ranked in the top 16.0% of the prescription frequency were listed as strongly recommended, and the antibiotic combinations with an effectiveness rate lower than 50% and ranked in the bottom 16% of the prescription frequency were not recommended. The remaining antibiotic combinations were listed as weakly recommended. The obtained antibiotic combinations were entered into the combination therapy recommendation system for prediction, and the results are shown in Table 2.

[0077]

[0078]

[0079]

[0080]

[0081] According to the comparison between the prediction results and the clinical decision results, among the clinically strongly recommended decision combinations, 73.33% of the prediction results were consistent with the actual clinical distribution, 61.54% of the clinically weakly recommended antibiotic combination prediction results were consistent with the corresponding clinical distribution, and 72.22% of the clinically not recommended antibiotic combination prediction results were consistent with the corresponding clinical decision situations.

[0082] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A method for evaluating the recommended types of clinical antibiotic combination therapy, characterized in that: The preset antibiotic combination recommendation model was used to evaluate the recommended types of antibiotic combination in clinical practice; The antibiotic combination recommendation model is used to evaluate the rationality and feasibility of clinical antibiotic combination therapy and is constructed through the following training process: S1. Obtain a historical data set of different antibiotic combination medications, and extract the historical compatibility frequency of any pair of antibiotic combinations in the historical data set, as well as the historical effectiveness rate after the combination is used; S2. Based on the historical compatibility frequency and historical effectiveness of any of the aforementioned antibiotic combinations, determine the grading threshold to determine the recommended type of combined medication for the antibiotic combination; S3, using the molecular fingerprint information of each antibiotic in the aforementioned antibiotic combination as input data, and its recommended type as output data, performing classification learning through a fully connected deep neural network to obtain the three-classification recommendation model for the antibiotic combination; The process of determining the recommended type to which any pair of antibiotic combinations belongs in step S2 includes: Based on the historical effectiveness, the antibiotic combination is divided into a strong antibiotic combination and a weak antibiotic combination, wherein the historical effectiveness of all antibiotic groups in the strong antibiotic combination is not less than 50%, and the historical effectiveness of all antibiotic groups in the weak antibiotic combination is less than 50%; Based on the historical compatibility frequency, the antibiotic combinations in the strong antibiotic combination and the weak antibiotic combination are sorted by compatibility frequency in descending order of the historical compatibility frequency; The recommendation type to which the antibiotic combinations whose compatibility frequency is ranked in the top 16% among the strong antibiotic combinations is determined as a strongly recommended type, the recommendation type to which the antibiotic combinations whose compatibility frequency is ranked in the bottom 16% among the weak antibiotic combinations is determined as a not recommended type, and the recommendation type to which the remaining antibiotic combinations among the strong antibiotic combinations and the weak antibiotic combinations are determined as a weakly recommended type.

2. The evaluation method according to claim 1, characterized in that: The historical data set includes a clinical patient medical record data set that has received antibiotic treatment and has undergone at least two infection-related examinations.

3. The evaluation method according to claim 2, characterized in that: The clinical patients do not include patients who have received immunointervention therapy and suffer from immune disorders, or patients who have only received one type of antibiotic treatment, or patients who have not received combined antibiotic treatment.

4. The evaluation method according to claim 2, characterized in that: The infection-related examination items include C-reactive protein examination and procalcitonin examination; the determination process of the historical effectiveness rate in step S1 includes: For any pair of the antibiotic combinations, the historical effectiveness of the antibiotic combination after use is determined based on the test results of the C-reactive protein test and the test results of the procalcitonin test corresponding to the antibiotic combination.

5. The evaluation method according to claim 1, characterized in that: In step S1, after obtaining the antibiotic combination medication history data set, the following steps are further included: For any antibiotic in the antibiotic combination medication history data set, perform generic name unification processing to obtain the standard name of each antibiotic; Based on the standard name of each antibiotic, the molecular fingerprint information of each antibiotic in the antibiotic combination medication history dataset is determined.

6. The evaluation method according to claim 5, characterized in that: The process of determining the molecular fingerprint information includes: Determining the chemical structural formula of each antibiotic based on the standard name of each antibiotic; Based on the chemical structural formula of each antibiotic, the MACCS key of each antibiotic is determined using the MACCS molecular access system as the molecular fingerprint information of each antibiotic.

7. The evaluation method according to claim 1, characterized in that: The evaluation method comprises the following steps: Obtaining the standard name of each antibiotic in the antibiotic combination to be recommended, and determining the molecular fingerprint information of each antibiotic based on the standard name of each antibiotic; Based on the molecular fingerprint information of each antibiotic, using the antibiotic combination recommendation model, determining a probability score of the clinical recommendation type of the antibiotic combination to be recommended; Determining a recommendation type corresponding to the antibiotic combination to be recommended based on the probability score of the clinical recommendation type of the antibiotic combination to be recommended; The recommendation types include a non-recommended type, a weakly recommended type and a strongly recommended type.

8. The evaluation method according to claim 7, characterized in that: The determining the recommendation type corresponding to the antibiotic combination to be recommended based on the probability score of the clinical recommendation type of the antibiotic combination to be recommended includes: The highest probability score among the probability scores of the clinical recommendation types of the antibiotic combination to be recommended is determined, and the recommendation type corresponding to the highest probability score is determined as the recommendation type corresponding to the antibiotic combination to be recommended.

Images