CN113555079B - Medication prediction system based on neural network - Google Patents

Abstract

Translated fromChinese

本发明提供了一种基于神经网络的用药预测系统，其特征在于，主要包括：痊愈总时长预测模块基于基础信息利用预先训练好的分类模型预测得到患者痊愈的总时长；剩余时长预测模块基于基础信息以及跟踪数据利用预先训练好的时序模型预测得到患者达到痊愈状态时还需要的时长；时长设定与判断模块判断已用药时长Δt与剩余时长T之和是否与患者痊愈的总时长相近，判断为否时，药效生存分析模块通过COX回归模型得到以用药停止为终结事件的药效曲线，进而对该药效曲线进行优化，时序模型微调模块基于优化后药效曲线对时序模型进行微调得到个性化时序模型；用药建议模块利用个性化时序模型预测得到用药预测结果作为医生用药调整的参考。

The invention provides a medication prediction system based on a neural network, which is characterized in that it mainly includes: a total healing time prediction module based on basic information and using a pre-trained classification model to predict the total healing time of a patient; the remaining time prediction module is based on basic information. Information and tracking data use the pre-trained time series model to predict the time required for the patient to reach the recovery state; the duration setting and judgment module judges whether the sum of the medication duration Δt and the remaining duration T is similar to the total duration of the patient's recovery, and judges When it is no, the drug efficacy survival analysis module obtains the drug efficacy curve with the discontinuation of drug use as the end event through the COX regression model, and then optimizes the drug efficacy curve. Personalized time series model; the medication recommendation module uses the personalized time series model to predict the medication prediction results as a reference for doctors to adjust medication.

Description

Medication prediction system based on neural network

Technical Field

The invention belongs to the technical field of medical care informatics, and particularly relates to a medication prediction system based on a neural network.

Background

The artificial neural network is one of the most active branches of computational intelligence and machine learning research, and has been widely applied to different fields of machine fault diagnosis, voice recognition, security management and the like. With the rapid increase of medical data scale, the artificial neural network becomes a new technology in the medical field due to the characteristics of self-learning, self-optimization and the like, and features or related data which are interested by medical staff can be extracted from a large amount of medical data by using the new technology, so that diagnosis and treatment references are provided for the medical staff.

In recent years, with the development of the prosperous medical science and technology and the explosive increase of medical knowledge, the contents to be learned by medical staff are increasing, and it is difficult to ensure the appropriateness, economy, effectiveness and safety of medication during the medication process. Unreasonable unsafe medication situations are ubiquitous, and medical care personnel need to pay attention to safe medication supervision for both chronic patients and acute patients. At present, medical staff can actively adjust the medicine only when the detection indexes of patients are obviously different, and the medicine is not timely adjusted according to feedback information after the patients take the medicine in the early period. How to extract useful information from wide medical knowledge and feedback information of a patient after personalized medication through a neural network, and giving medication adjustment information in time according to the useful information is the key point of safety medication supervision.

The existing neural network model can predict the medicine amount of a patient at the next moment, but the neural network model only predicts the medicine amount, the medicine type cannot be changed, the neural network model is not specific to the individual patient, the prediction result deviates due to individual difference, and the applicability of the neural network model is poor.

Disclosure of Invention

In order to solve the problems, the invention provides a medication prediction system for timely adjusting medication of a patient through a personalized neural network, and the invention adopts the following technical scheme:

the invention provides a medicine use prediction system based on a neural network, which is characterized by comprising the following components: the basic information acquisition module is used for acquiring basic information of the patient, and the basic information at least comprises a diagnosis conclusion of the patient and prescription conditions; the total healing duration prediction module is used for predicting the total healing duration of the patient by utilizing a pre-trained classification model based on basic information and taking the total healing duration as the first total healing duration T₁(ii) a The tracking data acquisition module is used for acquiring feedback information of the patient in the medicine taking process as tracking data, and the tracking data comprises the taken medicine time delta t of the patient; residual duration prediction module based on basic informationThe time length needed by the patient when the patient reaches the recovery state is obtained through forecasting by utilizing a pre-trained time sequence model and is used as the residual time length T; a time length setting and judging module for setting the sum of the used time length delta T and the residual time length T as a second total healing time length T₂And judging the total first healing duration T₁And the total duration of the second healing period T₂Whether they are similar; the pharmacodynamic survival analysis module is used for obtaining a pharmacodynamic curve taking the medication stoppage as a termination event through the COX regression model based on the basic information and the tracking data when the duration setting and judging module judges that the duration is not the duration setting and judging module; the drug effect curve optimizing module is used for optimizing the drug effect curve so as to obtain an optimized drug effect curve; the time sequence model fine tuning module is used for taking parameters corresponding to the optimized pharmacodynamic curve as a fine tuning direction, and fine tuning is carried out on the time sequence model according to the fine tuning direction, so that an individualized time sequence model is obtained; and the medication suggestion module predicts the medication of the patient at the next moment by utilizing the personalized time sequence model based on the tracking data, so that a medication prediction result is obtained and is used as a reference for the medication adjustment of the doctor.

The invention also provides a medicine prediction system based on the neural network, which is characterized in that: the training process of the pre-trained classification model and the pre-trained time sequence model comprises the following steps: step S1-1, acquiring a training basic data set and a training tracking data set; step S1-2, obtaining the actual healing duration of different patients according to the basic data set for training; step S1-3, building a classification model, training the classification model by using the actual healing time as a real label and using a training basic data set until the classification model is converged, thereby obtaining a trained classification model; and step S1-4, constructing a time sequence model, and training the time sequence model by using a training basic data set and a training tracking data set until the time sequence model is converged, thereby obtaining the trained time sequence model.

The invention also provides a medicine prediction system based on the neural network, which is characterized in that: the tracking data set for training is obtained through the following steps: step S2-1, obtaining the history of medication process of different patientsAn order data set; step S2-2, dividing the historical time sequence data set into a complete data set and an incomplete data set, taking the complete data set as a current complete data set, and separating the incomplete data set to obtain a right deletion data set; step S2-3, calculating the average duration N of the recording time of the historical time sequence data corresponding to all patients in the historical time sequence data set; step S2-4, determining the recording time n corresponding to each right deleted data in the right deleted data set_iIf the average duration is longer than N, if so, the right deleted data is supplemented into the current complete data set to form a new current complete data set, and if not, the right deleted data is used as alternative data; step S2-5, calculating consistency index C between each alternative data and the current complete data set_R(ii) a Step S2-6, judging consistency index C corresponding to each alternative data_RWhether greater than the standard conformance indexC₀(ii) a And S2-7, when the judgment in the step S2-6 is yes, supplementing the alternative data into the current complete data set to form a new current complete data set as a tracking data set for training.

The invention also provides a medicine prediction system based on the neural network, which is characterized in that: wherein, the standard consistency indexC₀The method comprises the following steps: step S3-1, based on the basic data set for training, using COX regression model to carry out regression analysis on the recovery conditions of the diseases corresponding to each patient, thereby obtaining the recovery curves corresponding to the patients; step S3-2, obtaining the prediction data of COX regression model according to the cure curve, calculating the consistency between the prediction data and the corresponding basic data set for training, and obtaining the standard consistency indexC₀。

The invention also provides a medicine prediction system based on the neural network, which is characterized in that: wherein, the time sequence model is an LSTM model.

The invention also provides a medicine prediction system based on the neural network, which is characterized in that: the basic information also comprises basic information, detection indexes, disease symptoms and related indexes of the patient, and the tracking data comprises medication time, medicine dosage, symptom feedback, index change, medication delay and related matters.

According to the medicine use prediction system based on the neural network, the total healing time prediction module predicts the total healing time of the patient by using the pre-trained classification model, and the residual time prediction module predicts the time required by the patient to reach the healing state by using the pre-trained time sequence model, so that the healing time of the patient without medicine use is predicted by the classification neural network, the time required by the patient for healing is predicted by the time sequence data of a period of time taken by the patient through the time sequence neural network, if the two healing times are not consistent, the effect of taking the current medicine or medicine amount by the patient is not in accordance with the expected requirement and is adjusted in time, but the doctor considers the medicine use adjustment through the examination index of the patient, and the medicine use time of the patient can be shortened by adjusting the medicine in time, so that the patient can be healed quickly, realizes the safe medication supervision of patients. In addition, the time sequence model fine tuning module fine tunes the time sequence model according to the parameters corresponding to the optimized drug effect curve to obtain the personalized time sequence model, and then the personalized time sequence model is used for predicting to obtain the drug use prediction result of the patient at the next time, so that different patients correspond to the personalized time sequence model, the corresponding drug use prediction result is more targeted and has higher accuracy, and the appropriateness, effectiveness and safety of drug use are improved.

Drawings

Fig. 1 is a block diagram of a neural network-based medication prediction system according to an embodiment of the present invention.

Fig. 2 is a schematic diagram of basic information content according to an embodiment of the present invention.

FIG. 3 is a diagram illustrating contents of trace data according to an embodiment of the present invention.

FIG. 4 is a flowchart illustrating a training process of a classification model and a timing model according to an embodiment of the present invention.

FIG. 5 is a flowchart of a training tracking data set acquisition process according to an embodiment of the present invention.

Figure 6 is a schematic diagram of a healing curve according to an embodiment of the present invention.

FIG. 7 is a graph illustrating a dosing curve according to an embodiment of the present invention.

FIG. 8 is a graph illustrating an optimized dosing curve according to an embodiment of the present invention.

Fig. 9 is a flowchart of the operation of the neural network-based medication prediction system according to the embodiment of the present invention.

Detailed Description

In order to make the technical means, the creation features, the achievement purposes and the effects of the invention easy to understand, a neural network-based medication prediction system of the invention is specifically described below with reference to the embodiments and the accompanying drawings.

Examples

As shown in fig. 1, the neural network-basedmedication prediction system 1 includes a basicinformation acquisition module 11, a total healingtime prediction module 12, a trackingdata acquisition module 13, a remainingtime prediction module 14, a time setting andjudgment module 15, a drug effectsurvival analysis module 16, a drug effectcurve optimization module 17, a time sequence model fine-tuning module 18, and amedication suggestion module 19.

The basicinformation acquisition module 11 acquires basic information of a patient.

As shown in fig. 2, the basic information (as shown in fig. 2) includes basic information of the patient, detection index, disease symptom, related index (i.e., other index), diagnosis conclusion, and prescription condition.

The basic information is sex, age, height, weight and the like of the patient, the detection indexes are blood pressure, blood routine, urine routine and the like of the patient, the related indexes are medical images and other existing diseases, the diagnosis conclusion is the disease type of the patient, and the prescription condition is the prescription prescribed by a doctor and comprises the name of a medicine, the components of the medicine, the dosage and the like.

The total healingduration prediction module 12 predicts the total healing duration of the patient by using a pre-trained classification model based on the basic information and takes the total healing duration as the first total healing duration T₁。

The trackingdata obtaining module 13 obtains feedback information of the patient during the medication process as tracking data, wherein the tracking data includes the medication time Δ t of the patient.

As shown in fig. 3, the tracking data includes time, drug dosage, symptom feedback, index change, medication delay, and related items (i.e., other items).

Time 1, 2, 3, 4, and the healing time is the medication time of the patient, such as the medication of the 1 st time, the medication of the 2 nd time, and the like; the dosage of the medicine is normal dosage, halved dosage and the like; symptom feedback uses 4, 3, 2, 1 and 0 to represent symptom degree, 4 is a little relieved of symptom, 0 is recovery, and the degrees corresponding to 3, 2 and 1 are analogized in sequence; index changes are similar to symptom feedback, withnumbers 6 to 0 reflecting the degree of index change; the delay of medication is to record the deviation of the patient's medication time from the prescribed time, for example, 0.5h represents 0.5h earlier for taking the medicine, 1h represents 1h later for taking the medicine; the relevant items are other data of the patient and are not described herein.

The remainingduration prediction module 14 predicts the duration required by the patient to reach the healing state by using a pre-trained time sequence model based on the basic information and the tracking data, and uses the duration as the remaining duration T. Specifically, a pre-trained time sequence model is used for predicting a state (namely a healing state) that the medicine dosage of the patient at a certain moment is absent, and the time required by the patient to reach the healing state is obtained according to the time corresponding to the state and is used as the residual time T.

As shown in fig. 4, the training process of the above-mentioned pre-trained classification model and pre-trained time sequence model includes the following steps:

step S1-1, a training basic data set and a training tracking data set are obtained.

As shown in fig. 5, the training trace data set is obtained by the following steps:

and step S2-1, acquiring historical time sequence data sets in the medication process of different patients.

And step S2-2, dividing the historical time sequence data set into a complete data set and an incomplete data set, taking the complete data set as a current complete data set, and separating the incomplete data set to obtain a right deletion data set.

The complete data in the complete data set refers to the complete data from the observation start point to the occurrence of the terminal event (i.e. the event that the patient is observed to die or be terminal).

Incomplete data means that the patient's observed cutoff is not due to a termination event, but due to other causes, including loss of visit (loss of patient contact), withdrawal (withdrawal of the patient from the study due to non-research or non-treatment factors), and termination (the time specified in the design has met the limit of terminating observation, but the patient has not yet healed).

Right deletion data means that the starting time of administration of the patient is known but the recovery time is unknown in follow-up observation of the patient.

The trace data acquired by the tracedata acquisition module 13 is right-deleted data.

And step S2-3, calculating the average duration N of the recording time of the historical time sequence data corresponding to all patients in the historical time sequence data set.

Step S2-4, determining the recording time n corresponding to each right deleted data in the right deleted data set_iAnd if the average duration is longer than the average duration N, the right deleted data is supplemented into the current complete data set to form a new current complete data set if the average duration is yes, and the right deleted data is used as alternative data if the average duration is not yes.

Step S2-5, calculating consistency index C between each alternative data and the current complete data set_R。

Step S2-6, judging consistency index C corresponding to each alternative data_RWhether greater than the standard conformance indexC₀If yes, the process proceeds to step S2-7, if no, the candidate data are discarded, and then the process proceeds to step S2-5, where a consistency index C corresponding to the next candidate data is calculated_R。

Wherein, the standard consistency indexC₀Specifically, the calculation is performed in the following steps S3-1 to S3-2.

And step S3-1, performing regression analysis on the disease healing conditions corresponding to each patient by using a COX regression model based on the training basic data set, thereby obtaining a healing curve corresponding to the patient.

In the disease survival analysis, each healing factor of a patient is used as a dependent variable, and the COX regression model can be used for obtaining the instantaneous death rate of the disease corresponding to the patient at a certain moment (namely, the healing rate of the patient at a certain moment). Wherein, the healing factors can be the course of medication, the timing of medication, the frequency of medication, the dosage and the like.

As shown in fig. 6, the horizontal axis represents the survival time of the disease, the vertical axis represents the survival rate of the disease, the two recovery curves represent the recovery curves of two patients under the same disease, and when the survival rate of the disease approaches to 0, the patient is cured. The healing time of different patients is different, and the healing curve can be changed by adjusting various healing factors, so that the healing time is shortened.

Step S3-2, obtaining the prediction data of COX regression model according to the cure curve, calculating the consistency between the prediction data and the corresponding basic data set for training, and obtaining the standard consistency indexC₀。

In this embodiment, the consistency calculation steps are specifically as follows:

1. all samples were paired with each other for a total of N x (N-1)/2 pairs, where N is the number of samples. If a total of 20 patient data sets are assumed, N = 20.

2. The excluded pairs in the sample are: (1) the observation endpoint was not reached in both of the pairs, i.e. recovery was complete; (2) one patient a was administered for a shorter time than the other patient B, however patient a had not reached the event endpoint (this pairing did not tell who healed first). The number of pairs left at this time is denoted as M.

If it is assumed that the exclusion of 5 groups does not reach the observation end point; excluding 5 groups with non-uniform timing, 10 patients remained, i.e. log matched M = 10.

3. And calculating the number of pairs of the rest M pairs, wherein the number of pairs with the predicted result consistent with the actual result is recorded as K, namely (if the predicted medication time of one patient with longer medication time is longer than that of the other patient, or the medication time of one patient with higher predicted cure probability is longer than that of the other patient, the predicted result is consistent with the actual result, and the result is called as consistent.

4. As a result: calculate C-index = K/M, i.e. the standard conformance index C₀=K₀/M₀. Assuming that the prediction is consistent with reality, denoted as K =10, C-index =10/20= 0.5.

And step S2-7, supplementing the alternative data into the current complete data set to form a new current complete data set as a tracking data set for training.

And step S1-2, obtaining the actual healing time of different patients according to the basic data set for training.

And step S1-3, building a classification model, training the classification model by using the actual healing time as a real label and using the training basic data set until the classification model is converged, thereby obtaining the trained classification model.

And step S1-4, constructing a time sequence model, and training the time sequence model by using a training basic data set and a training tracking data set until the time sequence model is converged, thereby obtaining the trained time sequence model.

The time sequence model is an LSTM model, and the LSTM model is a special RNN, can learn long dependency relationships and is a RNN with better memorability.

The time length setting and judgingmodule 15 sets the sum of the used time length delta T and the residual time length T as the second total healing time length T₂And judging the total first healing duration T₁And the total duration of the second healing period T₂Whether or not they are close.

When the time length setting anddetermination module 15 determines yes, the patient medication does not need to be adjusted.

When the time length setting and determiningmodule 15 determines that the event is negative, the pharmacodynamic-survival analyzing module 16 obtains a pharmacodynamic curve in which the drug administration is stopped as a termination event through a COX regression model based on the basic information and the trace data.

In this embodiment, assuming that the predicted recovery period T1 of the patient is 15 days, and the medication period Δ T =6 and the remaining period T =12, the second total recovery period T is₂= Δ T + T =18,. At this time, the duration setting and determiningmodule 15 determines whether the medication effect is not good.

The pharmacodynamiccurve optimizing module 17 optimizes the pharmacodynamic curve to obtain an optimized pharmacodynamic curve.

In this embodiment, the pharmacodynamiccurve optimizing module 17 optimizes the pharmacodynamic curve by adjusting the biased regression coefficient of independent variables (e.g., different drugs, different dosages) in the COX regression model, thereby obtaining an optimized pharmacodynamic curve with the shortest recovery time. Wherein, the biased return coefficient can be parameters such as medication time, medication frequency, dosage, medicine type and the like.

Suppose that the pharmacodynamic-survival analysis module 16 has a dose curve based on the underlying information and the follow-up data and obtained by a COX regression model as shown in fig. 7. In fig. 7, the vertical axis represents the target value of drug administration, the horizontal axis represents time (unit: day), the curve corresponding to "a patient" represents the curve obtained by actual drug administration, and the curve corresponding to "COX regression" represents the drug administration curve obtained by the pharmacodynamic-survival analysis module 16 using the COX regression model.

Next, the pharmacodynamiccurve optimization module 17 adjusts the bias return coefficient so that the target value of the medication curve is reduced, and the final optimized curve is shown in fig. 8.

The time sequence modelfine tuning module 18 uses the parameters corresponding to the optimized pharmacodynamic curve (i.e. the parameters such as the drug and the dosage corresponding to the optimized pharmacodynamic curve) as the fine tuning direction, and fine tunes the time sequence model according to the fine tuning direction, thereby obtaining the personalized time sequence model.

Themedication suggestion module 19 predicts the medication of the patient at the next time by using the personalized time sequence model based on the tracking data, so as to obtain a medication prediction result as a reference for the medication adjustment of the doctor.

As shown in fig. 9, the operation process of the neural network-basedmedication prediction system 1 includes the following steps:

step S4-1, the basicinformation obtaining module 11 obtains the basic information of the patient, and then step S4-2 is entered;

step S4-2, the total healingduration prediction module 12 predicts the total healing duration of the patient by using a pre-trained classification model based on the basic information, and the total healing duration is used as a first total healing duration T₁Then, it goes to step S4-3;

step S4-3, the trackingdata obtaining module 13 obtains the feedback information of the patient in the medicine taking process as the tracking data, the tracking data includes the time length delta t of the taken medicine of the patient, and then the step S4-4 is entered;

step S4-4, the residualduration prediction module 14 predicts the duration needed by the patient to reach the recovery state by using a pre-trained time sequence model based on the basic information and the tracking data to be used as the residual duration T, and then the step S4-5 is carried out;

step S4-5, the duration setting and judgingmodule 15 sets the sum of the used duration Deltat and the remaining duration T as the second total healing duration T₂And judging the total first healing duration T₁And the total duration of the second healing period T₂If the judgment result is yes, the medicine is not adjusted, and the process enters an end state;

step S4-6, the pharmacodynamicsurvival analysis module 16 obtains a pharmacodynamic curve taking the drug administration stop as a termination event through a COX regression model based on the basic information and the tracking data, and then the step S4-7 is carried out;

step S4-7, the pharmacodynamiccurve optimization module 17 optimizes the pharmacodynamic curve to obtain an optimized pharmacodynamic curve, and then the step S4-8 is carried out;

step S4-8, the time sequence modelfine tuning module 18 takes the parameters corresponding to the optimized pharmacodynamic curve as a fine tuning direction, fine tuning is carried out on the time sequence model according to the fine tuning direction, so as to obtain an individualized time sequence model, and then the step S4-9 is carried out;

and step S4-9, themedication suggestion module 19 predicts the medication of the patient at the next moment by using the personalized time sequence model based on the tracking data, so as to obtain a medication prediction result as a reference for the medication adjustment of the doctor, and then enters an ending state.

According to themedication prediction system 1 based on the neural network provided by the embodiment, the total healingtime prediction module 12 predicts the total healing time of the patient by using the pre-trained classification model, and the residualtime prediction module 14 predicts the time required by the patient to reach the healing state by using the pre-trained time sequence model, so that the healing time of the patient without medication is predicted by the classification neural network, the time required by the patient for long time to heal is predicted by the time sequence data of the patient taking a period of time through the time sequence neural network, if the two healing times are not consistent, the effect of the current medicine or the dosage taken by the patient is not in accordance with the expected need to be adjusted in time, but the doctor considers medication adjustment through the examination index of the patient, and the medication adjustment can shorten the medication time of the patient in time, the patient can be quickly cured, and the safe medication supervision of the patient is realized. In addition, the time sequence modelfine tuning module 18 fine tunes the time sequence model according to the parameters corresponding to the optimized drug effect curve to obtain an individualized time sequence model, and then the medication prediction result of the patient at the next time is obtained by using the individualized time sequence model prediction, so that different patients correspond to the individualized time sequence model, the corresponding medication prediction result is more targeted and has higher accuracy, and the medication appropriateness, effectiveness and safety are improved.

In the above embodiment, the complete data set and the incomplete data set are firstly classified into the historical time series data set, and then the average duration N and the standard consistency index are usedC₀And further brushing and selecting right deleted data in the incomplete data set to form a new current complete data set as a tracking data set for training, so that under the condition that the historical time sequence data set is largely incomplete, as much useful data as possible are obtained from the historical time sequence data set as the tracking data set for training, and the two neural network models are ensured to have enough training data.

The above-described embodiments are merely illustrative of specific embodiments of the present invention, and the present invention is not limited to the description of the above-described embodiments.

Claims (4)

Translated fromChinese

1.一种基于神经网络的用药预测系统，其特征在于，包括：1. a kind of medication prediction system based on neural network, is characterized in that, comprises:基础信息获取模块，用于获取患者的基础信息，该基础信息至少包括所述患者的诊断结论以及处方情况；a basic information acquisition module for acquiring basic information of a patient, the basic information at least including the patient's diagnosis conclusion and prescription;痊愈总时长预测模块，基于所述基础信息，利用预先训练好的分类模型预测得到所述患者痊愈的总时长，作为第一痊愈总时长T1；The total duration of recovery prediction module, based on the basic information, utilizes the pre-trained classification model to predict the total duration of recovery of the patient, as the first total duration of recovery T1;跟踪数据获取模块，获取所述患者在服药过程中的反馈信息作为跟踪数据，该跟踪数据中包括所述患者的已用药时长Δt；A tracking data acquisition module, which acquires the feedback information of the patient in the process of taking medicine as the tracking data, and the tracking data includes the patient's medication duration Δt;剩余时长预测模块，基于所述基础信息以及所述跟踪数据，利用预先训练好的时序模型预测得到所述患者达到痊愈状态时还需要的时长，作为剩余时长T；The remaining duration prediction module, based on the basic information and the tracking data, utilizes the pre-trained time sequence model to predict the duration that the patient needs to reach the recovery state, as the remaining duration T;时长设定与判断模块，将所述已用药时长Δt与所述剩余时长T之和设定为第二痊愈总时长T2，并判断所述第一痊愈总时长T1与所述第二痊愈总时长T2是否相近；Duration setting and judging module, the sum of the described medication duration Δt and the remaining duration T is set as the second total duration of recovery T2, and it is judged that the first total duration of recovery T1 and the second total duration of recovery are Whether T2 is similar;药效生存分析模块，在所述时长设定与判断模块判断为否时，基于所述基础信息以及所述跟踪数据，通过COX回归模型得到以用药停止为终结事件的药效曲线；The drug efficacy survival analysis module, when the duration setting and judgment module judges it to be no, based on the basic information and the tracking data, obtain the drug efficacy curve with the drug discontinuation as the terminating event through the COX regression model;药效曲线优化模块，对所述药效曲线进行优化从而得到优化后药效曲线；A drug efficacy curve optimization module, which optimizes the drug efficacy curve to obtain an optimized drug efficacy curve;时序模型微调模块，将所述优化后药效曲线对应的参数作为微调方向，根据该微调方向对所述时序模型进行微调，从而得到个性化时序模型；以及A timing model fine-tuning module, which uses the parameter corresponding to the optimized drug effect curve as a fine-tuning direction, and fine-tunes the timing model according to the fine-tuning direction, thereby obtaining a personalized timing model; and用药建议模块，基于所述跟踪数据，利用所述个性化时序模型对所述患者下一时刻的用药进行预测，从而得到用药预测结果作为医生用药调整的参考，The medication suggestion module, based on the tracking data, uses the personalized time series model to predict the medication of the patient at the next moment, so as to obtain the medication prediction result as a reference for the doctor's medication adjustment,所述预先训练好的分类模型以及所述预先训练好的时序模型的训练过程包括如下步骤：The training process of the pre-trained classification model and the pre-trained time series model includes the following steps:步骤S1-1，获取训练用基础数据集以及训练用跟踪数据集；Step S1-1, obtaining a basic data set for training and a tracking data set for training;步骤S1-2，根据所述训练用基础数据集得到不同病患的实际痊愈时长；Step S1-2, obtain the actual recovery duration of different patients according to the basic data set for training;步骤S1-3，搭建分类模型，将所述实际痊愈时长作为真实标签，利用所述训练用基础数据集训练所述分类模型，直到所述分类模型收敛，从而得到所述训练好的分类模型；Step S1-3, building a classification model, using the actual recovery duration as a true label, and using the basic data set for training to train the classification model until the classification model converges, thereby obtaining the trained classification model;步骤S1-4，搭建时序模型，利用所述训练用基础数据集以及所述训练用跟踪数据集对所述时序模型进行训练，直到所述时序模型收敛，从而得到所述训练好的时序模型，Step S1-4, build a time series model, use the basic data set for training and the tracking data set for training to train the time series model, until the time series model converges, so as to obtain the trained time series model,所述训练用跟踪数据集通过如下步骤得到：The training tracking data set is obtained through the following steps:步骤S2-1，获取不同病患用药过程中的历史时序数据集；Step S2-1, obtaining historical time series data sets during the medication process of different patients;步骤S2-2，将所述历史时序数据集分为完整数据集以及不完整数据集，将所述完整数据集作为当前完整数据集，并从所述不完整数据集中分离得到右删失数据集；Step S2-2, dividing the historical time series data set into a complete data set and an incomplete data set, using the complete data set as the current complete data set, and separating from the incomplete data set to obtain a right-censored data set ;步骤S2-3，计算所述历史时序数据集中所有所述病患对应的历史时序数据的记录时间的平均时长N；Step S2-3, calculating the average duration N of the recording time of the historical time series data corresponding to all the patients in the historical time series data set;步骤S2-4，判断所述右删失数据集中各个右删失数据对应的记录时间ni是否大于所述平均时长N，判断为是时将所述右删失数据补入所述当前完整数据集从而形成新的当前完整数据集，判断为否时将所述右删失数据作为备选数据；Step S2-4, judge whether the recording time ni corresponding to each right-censored data in the right-censored data set is greater than the average duration N, and add the right-censored data into the current complete data set when it is judged that it is Thereby a new current complete data set is formed, and the right-censored data is used as candidate data when it is judged to be no;步骤S2-5，计算每一个所述备选数据与所述当前完整数据集之间的一致性指数CR；Step S2-5, calculating the consistency index CR between each of the candidate data and the current complete data set;步骤S2-6，判断每一个所述备选数据对应的所述一致性指数CR是否大于标准一致性指数C0；Step S2-6, judging whether the consistency index CR corresponding to each candidate data is greater than the standard consistency index C0;步骤S2-7，在所述步骤S2-6判断为是时，将所述备选数据补入所述当前完整数据集从而形成新的当前完整数据集，作为训练用跟踪数据集。Step S2-7, when the determination in step S2-6 is yes, add the candidate data to the current complete data set to form a new current complete data set, which is used as a training tracking data set.2.根据权利要求1所述的基于神经网络的用药预测系统，其特征在于：2. the medication prediction system based on neural network according to claim 1, is characterized in that:其中，所述标准一致性指数C₀通过如下步骤计算得到：Wherein, the standard consistency indexC₀ is calculated by the following steps:步骤S3-1，基于所述训练用基础数据集，利用COX回归模型对各个所述病患对应的疾病痊愈情况进行回归分析，从而得到与所述病患对应的痊愈曲线；Step S3-1, based on the basic data set for training, use the COX regression model to carry out regression analysis on the disease recovery situation corresponding to each of the patients, thereby obtaining a recovery curve corresponding to the patient;步骤S3-2，根据所述痊愈曲线得到所述COX回归模型的预测数据，计算所述预测数据与对应的所述训练用基础数据集之间的一致性，得到标准一致性指数C₀。Step S3-2, obtaining the prediction data of the COX regression model according to the recovery curve, calculating the consistency between the prediction data and the corresponding basic data set for training, and obtaining a standard consistency indexC₀ .3.根据权利要求1所述的基于神经网络的用药预测系统，其特征在于：3. the medication prediction system based on neural network according to claim 1, is characterized in that:其中，所述时序模型为LSTM模型。Wherein, the time series model is an LSTM model.4.根据权利要求1所述的基于神经网络的用药预测系统，其特征在于：4. the medication prediction system based on neural network according to claim 1, is characterized in that:其中，所述基础信息还包括所述患者的基本信息、检测指标、疾病症状以及相关指标，Wherein, the basic information also includes the basic information, detection indicators, disease symptoms and related indicators of the patient,所述跟踪数据包括时间、药品用量、症状反馈、指标变化、用药延迟以及相关事项。The tracking data includes time, drug dosage, symptom feedback, index changes, drug delay and related matters.

Images