CN111696678A - Deep learning-based medication decision method and system - Google Patents

Deep learning-based medication decision method and system Download PDF

Info

Publication number
CN111696678A
CN111696678A CN202010543306.7A CN202010543306A CN111696678A CN 111696678 A CN111696678 A CN 111696678A CN 202010543306 A CN202010543306 A CN 202010543306A CN 111696678 A CN111696678 A CN 111696678A
Authority
CN
China
Prior art keywords
patient
medication
information
vector
scheme
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202010543306.7A
Other languages
Chinese (zh)
Other versions
CN111696678B (en
Inventor
吴嘉
余庚花
张璇
田晓明
常柳
庄庆贺
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Central South University
Original Assignee
Central South University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Central South University filed Critical Central South University
Priority to CN202010543306.7A priority Critical patent/CN111696678B/en
Publication of CN111696678A publication Critical patent/CN111696678A/en
Application granted granted Critical
Publication of CN111696678B publication Critical patent/CN111696678B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H10/00—ICT specially adapted for the handling or processing of patient-related medical or healthcare data
    • G16H10/60—ICT specially adapted for the handling or processing of patient-related medical or healthcare data for patient-specific data, e.g. for electronic patient records
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H20/00—ICT specially adapted for therapies or health-improving plans, e.g. for handling prescriptions, for steering therapy or for monitoring patient compliance
    • G16H20/10—ICT specially adapted for therapies or health-improving plans, e.g. for handling prescriptions, for steering therapy or for monitoring patient compliance relating to drugs or medications, e.g. for ensuring correct administration to patients
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Medical Informatics (AREA)
  • Public Health (AREA)
  • General Health & Medical Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • Primary Health Care (AREA)
  • Epidemiology (AREA)
  • Biomedical Technology (AREA)
  • Physics & Mathematics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Databases & Information Systems (AREA)
  • Pathology (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Medicinal Chemistry (AREA)
  • Chemical & Material Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Medical Treatment And Welfare Office Work (AREA)

Abstract

The invention provides a medication decision method and a medication decision system based on deep learning, which are characterized in that an analysis data set is obtained by collecting relevant information of a patient, relevant information of a medication scheme and actual curative effect evaluation information of the medication scheme on the actual curative effect of the patient, the analysis data set is analyzed and processed through a constructed curative effect prediction model so as to predict the predicted curative effect evaluation information of each medication scheme on different therapeutic effects of the patient, and the medication scheme suitable for the individualized requirement of the patient is recommended to the patient according to the predicted curative effect evaluation information. The medication decision method and the medication decision system have higher accuracy and stability when recommending the medication scheme, and are suitable for medical clinical application.

Description

Deep learning-based medication decision method and system
Technical Field
The invention belongs to the technical field of medical artificial intelligence auxiliary diagnosis, and particularly relates to a medication decision method and a medication decision system based on deep learning.
Background
In the medical artificial intelligence research, the intelligent auxiliary diagnosis system can help doctors to analyze and diagnose the state of an illness, the efficiency of the patients in hospitalizing is improved, the diagnosis and treatment time and period of the patients are shortened, meanwhile, the workload of the doctors is effectively reduced, and the problem of insufficient medical resources of the current society can be solved.
The diversity of the medicines and different pathological characteristics of patients complicate the medicine treatment, various influence factors can influence the dosage amount of the medicine, the individual judgment of doctors is not enough, the medication error is probably caused by incomplete information acquisition or insufficient personal knowledge storage, for example, for patients with non-small cell lung cancer, the medication scheme which is reasonably suitable for personalized targeted medicine of the patients is the most critical link in the diagnosis of doctors, and the doctors are required to have very rich experience. The intelligent medication decision method and the intelligent medication decision system can provide a predictive decision for the medication of the patient for reference of a doctor by acquiring the patient and historical medication information related to the patient and then performing corresponding calculation and processing through a medication decision model. Therefore, the intelligent medication decision system can better improve the medical quality, save the medical cost, timely, accurately and completely provide professional medication decision support for doctors, help the doctors to make correct and effective medication decisions and improve the success rate of medication.
Existing medication decision methods generally recommend a next medication regimen to a patient only through analysis of the patient's historical medication records, and rely on physician medication bias, i.e., on implicit feedback to the patient to provide recommended decisions on medication regimens. However, due to the sparseness of medication records, the performance of the curative effect prediction model is limited only by the implicit feedback method based on the empirical preference of the medication of a doctor, so that the accuracy of medication decision is influenced. Therefore, the existing medicine decision system based on machine learning cannot be widely applied in actual clinic.
Disclosure of Invention
In view of this, the present invention provides a medication decision method and system based on deep learning, so as to improve the accuracy of medication recommendation.
A deep learning-based medication decision method comprises the following steps:
collecting patient-related information, medication regimen-related information, and actual efficacy assessment information of the medication regimen on the actual efficacy outcome of the patient to obtain an analysis dataset,
constructing a curative effect prediction model to analyze the relationship among the patients and the relationship among the medication schemes according to the analysis data set, predicting curative effect evaluation information of the medication schemes on the treatment effect of different patients by combining the actual curative effect evaluation information,
recommending the medication scheme suitable for the individualized requirement of the patient to the patient according to the predicted curative effect evaluation information.
Preferably, the step of acquiring the predicted curative effect evaluation information by the prediction model comprises:
acquiring a relation matrix representing the relation between the patient and the medication scheme, wherein relation elements in the relation matrix represent the relation between the patient and the medication scheme, if the patient uses the medication scheme, the corresponding relation elements are represented as the actual curative effect evaluation value of the medication scheme on the patient, and if the patient does not use the medication scheme, the corresponding relation elements are set to be a first fixed value,
obtaining patient potential vectors and the drug potential vectors according to the relationship matrix, each of the patient potential vectors representing a relationship between one patient and a different medication regimen, each of the drug potential vectors representing a relationship between one medication regimen and a different patient,
respectively acquiring a patient auxiliary information matrix and a medication scheme auxiliary information matrix according to the patient-related information and the medication scheme-related information, wherein elements in the patient auxiliary information matrix represent information related to the patient, elements in the medication scheme auxiliary information matrix represent information related to the medication scheme,
extracting, by an SDAE autoencoder, a patient latent factor vector and a medication latent factor vector from the patient side information matrix and the medication side information matrix, respectively, the patient latent factor vector and the medication latent factor vector being hidden layer output information of the autoencoder,
and cooperatively filtering the drug potential vector, the drug regimen potential vector, the patient factor vector and the medication regimen factor vector through a neural network so as to learn the similarity between the patients and the similarity between the medication regimens and predict the treatment effect of each medication regimen on different patients by combining the relationship between the patients and the historical medication regimens of the patients so as to obtain the predicted treatment effect evaluation information.
Preferably, the analysis data set includes a patient set and a medication scheme set, and the step of obtaining a patient potential vector and the medication potential vector according to the relationship matrix includes:
setting all values in the relationship matrix other than the first fixed value to a second fixed value to obtain a similarity relationship matrix,
according to the similarity relation matrix, obtaining each patient label vector representing each patient and each medicine label vector representing each medicine scheme, wherein except the ith element, the ith patient label vector in the patient set is the second fixed value, the rest elements are set as the first fixed value, except the jth element, the jth medicine scheme label vector in the medicine scheme set is the second fixed value, and the rest elements are set as the first fixed value,
extracting a patient-medication relation feature vector of the relation between the ith patient and each medication in the set of medications from the similarity relation matrix,
extracting a patient-medication regimen-patient relationship feature vector for the relationship between the jth of the medication regimen and each patient in the set of patients from the similarity relationship matrix,
concatenating the patient-medication regimen relationship feature vector with the ith patient label vector to obtain the patient potential vector for the ith patient,
and connecting the medication scheme-patient relation feature vector with the jth patient label vector to obtain the drug potential vector corresponding to the jth medication scheme.
Preferably, after the elements in the patient auxiliary information matrix and the medication regimen auxiliary information matrix are randomly set to the first fixed values, the patient latent factor vector and the medication regimen latent factor vector are extracted by the SDAE self-encoder.
Preferably, the SDAE auto-encoder is trained to learn the patient latent factor vector and the medication regimen latent vector by minimizing reconstruction errors.
Preferably, the loss function of the efficacy prediction model consists of a minimized reconstruction error of the SDAE self-encoder during feature extraction and a prediction error of the efficacy prediction model for the efficacy prediction,
the prediction error is determined based on the predicted efficacy assessment information and the actual efficacy assessment information.
Preferably, recommending the medication regimen to the patient that is appropriate for the patient's personalized requirements based on the predicted efficacy assessment information comprises:
and sorting the predicted curative effect evaluation values in the predicted curative effect evaluation information from large to small, and selecting the medicine schemes corresponding to the first K predicted curative effect evaluation values in the sorting to recommend to the patient.
Preferably, discrete said patient potential vectors and said medication regime potential vectors are converted into a continuous vector representation by an embedding layer for input to said neural network.
A deep learning based medication decision system comprising: a data collection module, an efficacy prediction module, and a decision module configured to collect patient-related information, medication regimen-related information, and actual efficacy assessment information of the medication regimen on the actual efficacy effect of the patient, to obtain an analysis dataset,
the curative effect prediction module is configured to extract patient characteristic information and medication scheme characteristics according to the analysis data set, train the curative effect prediction model according to the extracted characteristic information to obtain predicted curative effect evaluation information of each medication scheme on different patient treatment effects,
the decision module is configured to recommend the medication scheme suitable for the individualized requirement of the patient to the patient according to the predicted curative effect evaluation information.
Preferably, the efficacy prediction module comprises:
an input layer configured to obtain a relationship matrix representing a relationship between the patient and the medication scheme according to the analysis data, wherein relationship elements in the relationship matrix represent the relationship between the patient and the medication scheme, if the patient uses the medication scheme, the corresponding relationship elements represent the actual efficacy evaluation value of the medication scheme on the patient, if the patient does not use the medication scheme, the corresponding relationship elements are set to a first fixed value, patient potential vectors and the medication potential vectors are obtained according to the relationship matrix, each patient potential vector represents a relationship between one patient and a different medication scheme, each medication potential vector represents a relationship between one medication scheme and a different patient, and patient auxiliary information is respectively obtained according to the patient related information and the medication scheme related information A matrix and a medication order assistance information matrix, elements of the patient assistance information matrix representing information including the patient related information, elements of the medication order assistance information matrix representing the medication order related information,
an embedding layer configured to convert the patient potential vector and the medication regimen vector embedding from discrete data to a continuous vector representation for input to the neural network layer,
a first SDAE self-encoder configured to extract a patient latent factor vector from the patient assistance information matrix, the patient latent factor vector being hidden layer output information of the first SDAE self-encoder,
a second SDAE self-encoder configured to extract a medication potential factor vector from the medication side information matrix, the medication potential factor vector being hidden layer output information of the second SDAE self-encoder,
a neural network layer configured to perform collaborative filtering processing on the drug potential vector, the drug regimen potential vector, the patient factor vector, and the medication regimen factor vector, thereby obtaining the predicted efficacy evaluation information.
The invention has one of the following beneficial effects: the relationship among the patients and the relationship among the medication schemes are analyzed through a curative effect prediction model, the predicted curative effect evaluation information of the treatment effect of each medication scheme on different patients is predicted by combining the actual curative effect evaluation information, the medication scheme is recommended according to the predicted curative effect evaluation information, and the recommendation method does not depend on the historical record relationship between the patients and the medication schemes, so that the recommendation accuracy is high.
The invention has the second beneficial effect: and performing collaborative filtering processing on the patient auxiliary relevant information and the representation of the historical medication records of the patient and the medication scheme through the characteristic extraction of the patient auxiliary relevant information and the medication scheme auxiliary relevant information so as to improve the stability of the curative effect prediction model.
Drawings
FIG. 1 is a schematic diagram of an intelligent medical auxiliary diagnosis system using the medication decision method provided by the invention;
FIG. 2 is a block diagram of a deep learning based medication decision system according to the present invention;
FIG. 3 is a diagram showing the case where the ratio ks of cases in which the classification of the medication scheme is correct is changed with the value of K;
FIG. 4 is a diagram illustrating the relationship between the importance parameter of the efficacy assessment value and the importance parameter of the auxiliary information and the assessment index Acc;
fig. 5 is a schematic diagram of the relationship between the importance parameter of the efficacy evaluation value and the importance parameter of the auxiliary information and the evaluation index Sen;
FIG. 6 is a diagram illustrating the relationship between the importance parameter of the efficacy evaluation value and the importance parameter of the auxiliary information and the evaluation index Spec;
FIG. 7 is a schematic representation of the change in the sample index in the case of triple line therapy in adenocarcinoma patients;
FIG. 8 is a schematic diagram showing the change of the index in the case of the three-line treatment in a patient with squamous carcinoma;
fig. 9 is a schematic diagram showing the accuracy of the medication decision system of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be described clearly and completely with reference to the accompanying drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without any creative effort, shall fall within the protection scope of the present invention. It should be noted that "…" in this description of the preferred embodiment is only for technical attributes or features of the present invention.
In order to truly apply the medicine decision system to clinic so as to assist doctors to make medicine scheme decisions suitable for patients, the invention provides a medicine decision method based on deep learning. By the medication decision method, decision of a medication scheme of the targeted drug of the patient with the non-small cell lung cancer can be realized, so that a doctor can be assisted to provide the most suitable targeted drug for the patient with the lung cancer, and time cost of the patient and the doctor is effectively saved. Fig. 1 is a schematic diagram of an intelligent medical auxiliary diagnosis system using the medication decision method provided by the present invention, the auxiliary diagnosis system mainly comprises a doctor, a patient, a medication decision module (operating in the computer of the doctor in fig. 1) and a data storage and analysis module (such as a cloud storage service center). When a patient goes to a doctor, the related information of the patient can be stored by the data storage and analysis module to travel an electronic medical record of the patient, a doctor examines the patient according to the electronic medical record to obtain examination data of the patient and feeds back the examination data to the data storage and analysis module for storage and analysis, the doctor sends a medication decision plan to the medication decision module before giving a diagnosis result to the patient, the medication decision module feeds back a medication decision result to the doctor according to the related information of the patient stored by the storage and analysis module and stores the decision diagnosis result in the data storage and analysis module, and the doctor gives the diagnosis result of the patient according to own experience and the medication decision result and feeds back the diagnosis result of the patient to the data storage and analysis module for storage. The medication decision is to screen out a personalized medication decision suitable for the patient from the medication scheme set stored in the data storage and analysis module, so that a systematic decision scheme is provided for a doctor, and the doctor is assisted in making a secondary diagnosis. The doctor gives the analysis result of the secondary diagnosis and gives decision suggestion feedback to form a medical record storage data storage analysis module so as to provide a new sample for the medication decision module, help to train the model well and improve the prediction accuracy step by step in a circulating manner.
The medication decision system provided by the embodiment of the present invention mainly includes steps 1 to 3, but the sequence of each step is not limited to the steps shown in this embodiment.
Step 1: patient-related information, medication regimen-related information, and actual efficacy assessment information of the medication regimen on the actual efficacy effect of the patient are collected to obtain an analysis dataset.
In the embodiment of the invention, the patient is a non-small cell lung cancer patient, and the medication scheme is a targeted medication scheme. The relationship among patients, the feedback condition of the patients after using the targeted drug regimen (the actual curative effect of the targeted drug regimen on the patients) and the relationship among the targeted drug regimens all have important influence on the decision of the drug regimen, so that the analysis data set can be obtained by collecting the relevant information of the patients, the relevant information of the targeted drug regimen and the actual curative effect evaluation information of the patients after using the targeted drug regimen for the subsequent curative effect prediction model to predict the predicted curative effect evaluation information of the therapeutic effect of each targeted drug patient on each patient according to the analysis data.
A targeted medication decision system based on data and characteristic analysis is established by collecting historical diagnosis and treatment records of patients, and sorting medication records of targeted drugs for each period of cancer patients, characteristic records of the targeted drugs and the like.
The relevant information of the patient collected by the data center comprises data such as patient diagnosis, disease, operation, nursing plan, drug selection and the like, patient image information, detection information, diagnosis records, drug use condition records and the like, and the data is sorted and classified to provide comprehensive analysis data information for drug use decision. In addition, the patient-related information includes patient auxiliary-related information such as medication, family history, treatment measures, height, weight, age, sex, lifestyle habits, symptoms and side effects, in addition to those mainly related to the patient's condition, thereby facilitating analysis of similarities between different patients. For example, if the patient is a non-small cell lung cancer patient, the patient auxiliary relevant information further includes a tumor mutation gene, a lung cancer stage and the like. The medication scheme related information mainly comprises various drug combination combinations with certain curative effect on the diseases of patients, and also comprises drug scheme auxiliary related information, wherein the auxiliary related information comprises the molecular structures, the chemical structures and the side effects of various drugs with different drugs and the relationship between corresponding targets (corresponding to the patients with the non-small cell lung cancer). Medication recommendations similar to the medication may be obtained from the medication-associated information to the patient. Each of the medication regimens is a set of regimens for providing treatment to a patient, for example, for a patient with non-small cell lung cancer, the medication regimens are targeted drug combination regimens, and the targeted drug combination regimens include information about the type and amount of the drugs used. The similarity information of the patient and the medication scheme can help us to predict the unknown relationship between the patient and the medication scheme, and if the historical medication record of the patient A does not record the medication scheme B, but the patient C with high similarity to the patient A uses the medication scheme B and has good practical curative effect, the patient A can recommend the medication scheme B to the patient A. Therefore, according to the auxiliary related information of the patient and the auxiliary related information of the medication scheme, the doctor can be helped to recommend the appropriate related medication scheme to the patient, and how to recommend the medication scheme is specifically carried out, which will be described in detail later. The data information of the medication analysis collected in this step is processed correspondingly to obtain each data set for predictive analysis, such as a patient set, a medication scheme set, and an actual curative effect evaluation value set of the medication scheme on the patient. The patient collection includes a plurality of patients, and the medication regimen collection includes a plurality of related medication regimens capable of treating a disease in the patient.
Step 2: and constructing a curative effect prediction model to analyze the relationship among the patients and the relationship among the medication schemes according to the analysis data set, and predicting the predicted curative effect evaluation information of the medication schemes on the treatment effects of different patients by combining the actual curative effect evaluation information.
In the curative effect prediction model, hidden feature representation of patient related information, medication scheme related information and relationship related information between patients and medication schemes is extracted, and then the extracted hidden feature information is integrated through a neural network to train the curative effect prediction model so as to predict the predicted curative effect evaluation information.
Specifically, as shown in fig. 2, in the present embodiment, the efficacy prediction model mainly includes an input layer, a first SDAE self-encoder, a second SDAE self-encoder, and a neural network layer. Step 2 may further include step 21 to step 25, but the order relationship between step 21 to step 25 is not limited to that shown in the present embodiment.
Step 21: acquiring a relation matrix representing the relation between the patient and the medication scheme, wherein relation elements in the relation matrix represent the relation between the patient and the medication scheme, if the patient uses the medication scheme, the corresponding relation elements represent the actual curative effect evaluation value of the medication scheme on the patient, and if the patient does not use the medication scheme, the corresponding relation elements are set as a first fixed value.
Let us assume that the set of patients obtained in step 1 is pat, the set of medication regimens is med, and the actual efficacy evaluation value of the medication regimen medj on the patient Pati (patient ith in the patient set) after the medication regimen medj (patient jth in the patient set) in the patient set is eij. The input layer in FIG. 2 is based on the patient set as Pat, the regimen set as med and the actual efficacy assessment as eijThe relationship matrix E1,. In this embodiment, each row of the relationship matrix E1 represents information related to each patient in the patient set Pat, and each column represents information related to each medication in the medication schedule set med, and it should be noted that the row name and the column name in fig. 2 simply indicate the patient set and the medication schedule set by the patient and the medication, respectively. In the relation matrix E1, the elements corresponding to the patient Pati and the medication order medj are Eij，eijThe evaluation of the actual effect of the patient's patii after the selection of the drug regimen medj during the treatment session is shown. If the patient's patii has used the drug combination medj in a history and has a record of its efficacy, eijThe value of (a) is the recorded estimate of actual efficacy, which means that the patient's pati is observed after the administration of the drug regimen medj. E if the patient's behavior of the patii after the administration of the regimen medj is not observed or the patient's patii has not been recorded of the administration of the regimen medj, eijSet to a first fixed value, e.g., 0, the patients pati andthere is no historical correspondence between medication regimens medj based on efficacy. Therefore, the relationship matrix E1 is a sparse matrix, and not all patients and each medication regimen in the relationship matrix E1 have records of corresponding actual efficacy assessment values, so that the subsequent therapeutic effect prediction model needs to predict the relationship between the patients and the medication regimens without history (no records of actual efficacy assessment values) to obtain the predicted therapeutic effect assessment information of each medication regimen for each patient. It should be noted that in other embodiments, the rows of the relationship matrix may also represent information related to the respective medication regimens listed in the information representative of the respective patients.
Step 22: acquiring patient potential vectors and the drug potential vectors according to the relation matrix, wherein each patient potential vector represents the relation between one patient and different medication schemes, and each drug potential vector represents the relation between one medication scheme and different patients.
To predict the curative effect evaluation information of each medication for each patient, we need to extract the relationship feature information between the patient pati in the patient set pat and each medication from the relationship matrix E1, and extract the relationship feature information between the medication medj in the medication set med and each patient. Specifically, in this embodiment, the step 22 may further include the following steps:
step 221: setting all values in the relationship matrix E1 that are not the first fixed value to a second fixed value to obtain a similar relationship matrix E2.
In the present embodiment, the non-first fixed value, i.e. the non-0 value, in the relationship matrix E1 facilitates extracting the relationship similarity feature between the patient and the medication, and the second fixed value is set to 1. The similar relation matrix E2 is obtained according to the relation matrix E1 and is also realized in an input layer of a curative effect prediction model, all elements of the corresponding relation matrix E1 of the relation history record of the patient and the medication scheme are set to be 1, and the history record relation exists between the patient corresponding to the elements and the medication scheme. The similarity matrix E2 is therefore characterized by the similarity relationship between the patient and the medication regimen
Step 222: according to the similarity relation matrix, obtaining each patient label vector representing each patient and each medicine label vector representing each medicine scheme, wherein except the ith element, the rest elements are set as the first fixed value, and except the jth element, the jth medicine scheme label vector in the medicine scheme set is set as the second fixed value, and the rest elements are set as the first fixed value.
In order to extract the potential characteristic information of the patient and the potential characteristic information of the medication scheme, each patient and each medication scheme are required to be provided with a label representing the patient and the medication scheme. We obtain the patient label vector Opat and the medication scheme label vector Omed from the similarity relation matrix E2 by a one-hot encoding method. The ith of the elements of the patient's pati corresponding to this row in the similarity relationship matrix E2 is set to 1, and the rest are set to 0 to obtain the patient label vector opti corresponding to the patient's pati. Similarly, after the medication scheme medj in the similarity relation matrix E2 is transposed into a row corresponding to the column, the jth element in the row is set to be 1, and the rest elements are set to be 0, so as to obtain the medication scheme label vector Omedj corresponding to the medication scheme medj.
Step 223: extracting a patient-medication relationship feature vector of the relationship between the ith patient and each medication in the set of medications from the similarity relationship matrix, and extracting a patient-medication relationship feature vector of the relationship between the first j of the medication and each patient in the set of patients from the similarity relationship matrix.
In the similarity relationship matrix E2, the row corresponding to the patient's patii represents the relationship between the patient's patii and each medication in the medication collection med, so we can obtain the patient-medication relationship of the patient's patii according to the row in the similarity relationship matrix E2Characteristic vector RepatiThe extraction formula is shown as formula (1):
Repati＝[rei*]1′|med|(1)
in the similarity relationship matrix E2, a column corresponding to the medication scheme medj represents the relationship between the patient medj and each patient in the medication scheme set pat, so we can obtain the medication scheme-patient relationship feature vector Re of the medication scheme medj according to the column in the similarity relationship matrix E2medjThe extraction formula is shown as formula (2):
Remedj＝[re*j T]1'|pat|(2)
step 224: connecting the patient-medication scheme relational feature vector with the ith patient label vector to obtain the patient potential vector corresponding to the ith patient, and connecting the medication scheme-patient relational feature vector with the jth patient label vector to obtain the drug potential vector corresponding to the jth medication scheme.
A patient potential vector S obtained according to the patient-medication scheme relation characteristic vector and the patient label vectorpatiThe calculation formula (2) is shown in formula (3):
Spati＝[Repati,opati](3)
a drug-using potential vector S obtained according to the drug-using scheme-patient relation feature vector and the patient label vectormedjThe calculation formula (2) is shown in formula (4):
Smedj＝[Remedj,omedj](4)
in the above formula, r is a conversion coefficient factor, and (2) denotes the number of each of the medications in the med set, and (3) denotes the number of each patient in the pat set. Because our relationship matrix E1 is a sparse matrix, when the training set of the therapeutic prediction model is large, and the number and kinds of the patient potential vector and the drug regimen set med are too large, the dimensions of the patient potential vector and the drug potential vector are too large to be calculated. Each of the potential vectors obtained is a discrete representation that does not reflect the relationship between different patients and the relationship between different drug regimens. Therefore, as shown in fig. 2, i also add an embedding layer in the efficacy prediction model, by which the patient potential vector and the drug are transformed from a discrete vector representation to a continuous vector representation in a vector. Specifically, each element in the patient potential vector and the drug potential vector can be embedded into the embedding layer in a binary coding manner to obtain an embedded matrix representing the relationship between the patient and the medication.
Step 23: and respectively acquiring a patient auxiliary information matrix and a medication auxiliary information matrix according to the patient-related information and the medication-related information, wherein elements in the patient auxiliary information matrix represent information related to the patient, and elements in the medication auxiliary information matrix represent information related to the medication.
In order to add additional information of patients and drugs to assist in extracting more accurate similarity characteristic information of patients and drug administration schemes, taking the decision of non-small cell lung cancer patients as an example, the non-small cell lung cancer patients-related information collected in step 1 further includes such patient auxiliary information as tumor mutation genes, staging, drug administration, family history, treatment measures, height, weight, age, sex, living habits, symptoms and side effects. The collected medication scheme related information also comprises medication scheme auxiliary information of molecular structures, chemical structures and side effects of the drugs of the targeted drug combination and the relationship between corresponding targets. The auxiliary information can help the curative effect prediction model to acquire similar information among patients and similar information among dosage schemes, so that unknown relation between the patients and the dosage schemes can be predicted. Therefore, in the input layer of fig. 2, the patient side information matrix Fp and the medication auxiliary information matrix Fd are also obtained according to the patient-related information and the medication-related information.
Step 24: extracting, by an SDAE autoencoder, a patient latent factor vector and a medication latent factor vector from the patient side information matrix and the medication side information matrix, respectively, the patient latent factor vector and the medication latent factor vector being hidden layer output information of the autoencoder.
As shown in fig. 2, we use a first SDAE self-encoder and a second SDAE self-encoder to extract the patient latent factor vector and the medication latent factor vector from the patient side information matrix Fp and the medication side information matrix Fd, respectively. The SDAE autoencoder is a neural network that learns specific features through inputs and outputs. By extracting features to generate an output close to the input, we can reflect the original input content by the extracted features. The SDAE autoencoder is a deep network model formed by stacking multiple DAE. While DAE is a variant of AE that typically uses a corrupt version to reconstruct the original input, more applicable to complex cases than the original self-encoder. Noise is added in the training process, so that the trained model can extract features and has good effect under the condition that the features are polluted by the noise. Thus, we can randomly set the elements in the patient side information matrix and the drug regimen side information matrix to the first fixed value (e.g., 0) to obtain a noisy patient side information matrix
Figure BDA0002539787500000091
And noise adding medication scheme auxiliary information matrix
Figure BDA0002539787500000101
Then, the patient latent factor vector and the medication latent factor vector are extracted by the SDAE self-encoder respectively.
Taking the first SDAE self-encoder to extract the patient latent factor vector as an example, the working process of the SDAE self-encoder is specifically described. The SDAE autoencoder is symmetrical in structure, and for a four-layer SDAE autoencoder, the training process from layer 0 to layer 2 can be regarded as an encoding process which abstracts the noisy patient side information matrix input into a more implicit expression Fp2. Likewise, the layer 2 to layer 4 process can be considered as a decoding process, which expresses the more implicit expression Fp2Reconstituting F from original inputplAnd (6) outputting. Said more implicit expression F of the intermediate layer we trained from the encoderp2Is an abstract representation extracted from the original input data. We add noise to the original input of the autoencoder to train in order to make the trained SDAE autoencoder more robust, improve its generalization performance, learn more stable feature representation, and have good results under sparse and dirty data conditions. After all, the feature information we have collected is not sufficiently complete and normalized. Therefore, the accuracy of recommending the medication to the patient under the condition of sparse medication evaluation is improved by respectively extracting the patient latent factor vector and the medication latent factor vector through the SDAE self-encoder.
Specifically, we randomly set the elements in the originally input patient assistance information matrix Fp to zero, and construct Fp as a noisy patient assistance information matrix containing noise
Figure BDA0002539787500000102
And will be noisy
Figure BDA0002539787500000103
The first SDAE self-encoder is input. We assume that the first SDAE self-encoder has a total of L layers, then FplRepresenting the output of its l layers. During training, the weight, deviation and representation output of each layer are initialized firstly. In particular, the weight parameter
Figure BDA0002539787500000104
By
Figure BDA0002539787500000105
Generating, deviation parameters
Figure BDA0002539787500000106
By
Figure BDA0002539787500000107
Fp generated and outputlBy passing
Figure BDA0002539787500000108
And (4) generating.
Wherein,
Figure BDA0002539787500000109
and
Figure BDA00025397875000001010
is a hyper-parameter for the initialization of the parameter,
Figure BDA00025397875000001011
and
Figure BDA00025397875000001012
weights and biases from the encoder l-layer, which represent patient side information feature extraction, (-) represent sigmoid functions, and I is a unit vector. We set the front L/2 layer to the encoder portion of the first SDAE auto-encoder and the back L/2 to the decoder portion of the first SDAE auto-encoder. The auxiliary information features of the patient are expressed by the patient latent factor vector, and the extraction of the patient latent factor vector is to learn two mapping functions, an encoding function and a decoding function fpat(Wpat,bpat) And gpat(Wp'at,b'pat) Wherein the coding function maps the patient's side information to a hidden layer, and the decoder reconstructs the hidden representation into the patient's side information. Wpat,bpatRespectively weight matrix and deviation matrix of the encoder, Wp'at,b'patRespectively, the weight matrix and the bias matrix of the decoder. Two mapping functions are learned by the deviation of the reconstructed information and the original input information. Feature extraction is then accomplished with a trained encoder. Patient side information matrix for noise
Figure BDA00025397875000001013
Hidden layer output for an encoderIs composed of
Figure BDA00025397875000001014
The calculation formula is as formula (5):
Figure BDA0002539787500000111
the decoder being aimed at deriving from noisy data
Figure BDA0002539787500000112
The original data Fp is reconstructed and the output of the decoder L layer is represented by equation (6):
Figure BDA0002539787500000113
the process by which the second SDAE extracts the drug regimen latent factor vector is similar to that described above for the first SDAE auto-encoder and will not be described again here. And hidden layer outputs of the intermediate output layers of the first SDAE self-encoder and the second SDAE self-encoder are the patient latent factor vector and the medication latent factor vector.
We learn the patient latent factor vector H by training the SDAE autoencoder to minimize reconstruction errorspatAnd the drug regimen potential vector Hmed. Also taking the reconstruction error of the extracted feature of the first SDAE from the encoder as an example, the definition of the reconstruction error is shown in equation (7):
Figure BDA0002539787500000114
wherein gamma ispat＝Wpat+W'patThe parameters that represent the model are then used,
Figure BDA0002539787500000115
is its regularization parameter, FpLIs the output of the L layer, and similarly we can define the reconstruction error L of the second SDAE self-encoder extracted featuremedIt is expressed as shown in equation (8):
Figure BDA0002539787500000116
step 25: and performing collaborative filtering processing on the drug potential vector, the drug scheme potential vector, the patient factor vector and the drug scheme factor vector through a neural network, so as to obtain the predicted curative effect evaluation information.
In fig. 2, the patient-with-regimen feature vector and the medication-with-patient feature vector are converted into continuous vector representations by the embedding layer to output corresponding Z's respectivelypatAnd ZmedWherein Z ispatTransforming vectors for patient-using plan relational features transformed by the embedding layer, ZmedVectors are transformed for patient-use plan relationship features transformed by the embedding layer.
The processing process of the collaborative filtering is as follows:
first, we will refer to the patient latent factor vector HpatAnd the patient-medication regime relationship characteristic switched vector ZpatConnected to obtain a patient potential representation XpatAnd the medication potential factor vector HmedAnd the medication plan-patient relationship feature transformation vector ZmedLinked to obtain a potential representation of the regimen YmedThen Xpat＝Hpat T·Zpat；Ymed＝Hmed T·Zmed。
Then the potential representation of the patient XpatThe element and the medication plan in (1) potentially represent YmedMultiplying the medium elements to obtain a potential relationship representation of the patient medication scheme, which is shown in equation (9):
sin(Xpat,Ymed)＝Xpat·Ymed(9)
finally, the curative effect prediction evaluation value is obtained according to the prescription of the patient
Figure RE-GDA0002564822460000121
The obtaining formula is shown as formula (10):
Figure RE-GDA0002564822460000122
σ (-) denotes the activation function,
Figure RE-GDA0002564822460000123
representing the transpose of the weight matrix of the output layer,
Figure RE-GDA0002564822460000124
patient pat representationiUsing dosing regimen medjPredictive assessment of the efficacy of treatment. σ (-) can be a nonlinear activation function and w can be learned from training data.
Therefore, the curative effect prediction model provided by the invention is used for predicting the curative effect evaluation value of each medication scheme on different patients based on the historical record relation between the patients and the medication schemes and combining the patient auxiliary related information and the medication scheme auxiliary related information. Even if there is no historical medication record between a certain patient and a medication, that is, there is no record of an actual efficacy evaluation value, the efficacy prediction model may extract information on another patient highly similar to the certain patient by using the patient assistance-related information and the medication assistance-related information, and recommend an appropriate medication to the certain patient from the medication record of the other patient. Therefore, the invention can help the curative effect prediction model to realize better performance by extracting the potential factor representation by utilizing the auxiliary information and extracting the potential representation of the patient and the medication scheme by combining the potential characteristics of the medication relation of the patient to predict the curative effect of the medication scheme.
In the embodiment of the present invention, we also define a loss function of the efficacy prediction model, so as to train the efficacy prediction model through the loss function. The loss function is composed of a minimum reconstruction error of the SDAE self-encoder in the process of feature extraction and a prediction error of the curative effect prediction model for predicting the curative effect, and the prediction error is determined according to the predicted curative effect evaluation information and the actual curative effect evaluation information.
In general, the loss function is composed of an auxiliary information reconstruction error when the self-encoder reconstructs auxiliary information and a prediction error of the therapeutic effect when the therapeutic effect prediction model predicts the therapeutic effect. The side information reconstruction error contains the error (i.e., loss) of the extraction of the side information features of the patient and the medication, and we define the reconstruction error of the patient side information from the previous analysis as follows:
Figure BDA0002539787500000125
wherein
Figure BDA0002539787500000126
The regularization parameters are represented.
Figure BDA0002539787500000127
A matrix of auxiliary information of the patient after the noise is added is shown,
Figure BDA0002539787500000128
and outputting the patient auxiliary information after feature extraction reconstruction through the first SDAE self-encoder. Similarly, the reconstruction error of the medication scheme auxiliary information is as follows:
Figure BDA0002539787500000129
where μ represents the second SDAE autoencoder model parameters and regularization term parameters that extract the medication latent factor vector.
Figure BDA00025397875000001210
A matrix representing the noise-adding-scheme-side information,
Figure BDA00025397875000001211
for output after feature extraction reconstruction of the recipe auxiliary information by the second SDAE self-encoder。
The prediction error of the curative effect represents the error between the curative effect evaluation value estimated by the curative effect prediction model and the actually recorded curative effect evaluation value. Specifically, the error between the predicted efficacy assessment value and the actual calculated efficacy assessment value is obtained through collaborative filtering by the neural network. Specifically, for each (pat)i,medj) Output of pairs predicts efficacy assessment
Figure RE-GDA0002564822460000131
The curative effect evaluation of the patient can be regarded as a label of the curative effect of the patient using the medication scheme, if the patient uses the medication scheme, the condition of the patient after using the medicine is recorded, based on the record, the evaluation of the improvement degree of the condition can be obtained, and the actually recorded curative effect evaluation value e is obtainedij. Therefore, the efficacy evaluation value predicted based on the efficacy prediction model
Figure RE-GDA0002564822460000132
Can be considered as a prediction of the degree of improvement in the patient's treatment effect in relation to the regimen. We will correlate the efficacy assessment
Figure RE-GDA0002564822460000133
Limited to [ -1,1 [ ]]So that we can distinguish the effect of the therapeutic effect, and therefore can be implemented using the tanh activation function, the prediction error of the therapeutic effect prediction model is defined as shown in equation (11):
Figure RE-GDA0002564822460000134
wherein R is+Represents a group of therapeutically effective examples, and R-A group of treatment-ineffective cases are shown, which are cases of patient and regimen relationships not tested or cases of patient no improvement or worsening after treatment. Lambda [ alpha ]preAnd
Figure BDA0002539787500000135
representing regularization term and model parameters, respectivelytherefore, the loss function of the efficient prediction model is defined as shown in equation (12), where α and β represent the hyperparameters of the loss function:
Ldrug_pre＝Lprediction+αLpat+βLmed(12)
and step 3: recommending the medication scheme suitable for the individualized requirement of the patient to the patient according to the predicted curative effect evaluation information.
And sorting the predicted curative effect evaluation values in the predicted curative effect evaluation information from large to small, and selecting the medicine schemes corresponding to the first K predicted curative effect evaluation values in the sorting to recommend to the patient.
The embodiment of the invention takes the medication decision of a patient with non-small cell lung cancer as an example, the medication scheme of the patient with non-small cell lung cancer is the medication scheme related to a targeting drug, and the molecular targeting drug has the action characteristic of accurately striking a key target point in the tumor growth process so as to not or rarely damage normal cells. When the patient is effectively treated by means of operation, chemotherapy or radiotherapy and the like, the serum tumor marker level is reduced compared with that before treatment, and obviously increased again when the patient relapses, so that the serum tumor marker level is prompted to be helpful for evaluating the treatment curative effect and predicting the tumor relapse.
The treatment condition of the tumor can be evaluated by judging the tumor area and the reduction level of the tumor marker through CT to evaluate the treatment effect. We hypothesize that the tumor area before treatment, including the tumor area in the spreading region, is
Figure BDA0002539787500000136
Tumor marker mar before treatmentiIs horizontally as
Figure BDA0002539787500000137
The total tumor area after treatment is
Figure BDA0002539787500000138
Tumor marker levels of
Figure BDA0002539787500000139
We define the normal level of tumor markers as
Figure BDA0002539787500000141
Figure BDA0002539787500000142
Typically 0.
We define the extent of area reduction from the change in tumor area before and after treatment as:
Figure BDA0002539787500000143
the change of the tumor marker can also reflect the treatment condition of the disease, and the change of the marker can be used for reversing to a certain extent
The effect of the treatment i defined the degree of change in the markers as:
Figure BDA0002539787500000144
therefore, we define the therapeutic efficacy assessment of the treatment regimen as shown in equation (15), i.e. said eijIs defined by the formula:
Figure BDA0002539787500000145
wherein,
Figure BDA0002539787500000146
the important parameters of the tumor marker with high relevance and the tumor marker with low relevance of the area and the non-small cell lung cancer are respectively. imp represents a tumor marker set having a high correlation with the evaluation of non-small cell lung cancer, and low represents a tumor marker set having a low correlation with non-small cell lung cancer.
Figure BDA0002539787500000147
Indicating an assessed association with non-small cell lung cancerHigh degree of change of tumor markers,
Figure BDA0002539787500000148
indicates the degree of change in tumor markers that have a low correlation with the evaluation of non-small cell lung cancer. It should be noted that the actual curative effect evaluation method according to the present invention is not limited to the method provided by the embodiment of the present invention.
As shown in fig. 2, the present invention further provides a deep learning-based medication decision system, which is characterized by comprising: a data collection module, an efficacy prediction module, and a decision module configured to collect patient-related information, medication regimen-related information, and actual efficacy assessment information of the medication regimen on the actual efficacy effect of the patient to obtain an analysis dataset. The curative effect prediction module is configured to extract patient characteristic information and medication scheme characteristics according to the analysis data set, and train the curative effect prediction model according to the extracted characteristic information to obtain predicted curative effect evaluation information of each medication scheme on different patient treatment effects. The decision module is configured to recommend the medication scheme suitable for the individual requirement of the patient to the patient according to the predicted curative effect evaluation information. Wherein the efficacy prediction module comprises an input layer, an embedded layer, a neural network layer, a first SDAE self-encoder and a second SDAE self-encoder. The input layer is configured to obtain a relationship matrix representing a relationship between the patient and the medication scheme according to the analysis data, wherein relationship elements in the relationship matrix represent the relationship between the patient and the medication scheme, if the patient uses the medication scheme, the corresponding relationship elements represent the actual curative effect evaluation value of the medication scheme on the patient, if the patient does not use the medication scheme, the corresponding relationship elements are set to a first fixed value, patient potential vectors and the medication potential vectors are obtained according to the relationship matrix, each patient potential vector represents the relationship between one patient and a different medication scheme, each medication potential vector represents the relationship between one medication scheme and a different patient, and patient auxiliary information is respectively obtained according to the patient related information and the medication scheme related information An information matrix and a medication scheme assistance information matrix, wherein elements in the patient assistance information matrix represent information comprising the patient and elements in the medication scheme assistance information matrix represent information relating to the medication scheme. The embedding layer configured to convert the patient potential vector and the prescription vector embedding from discrete data to a continuous vector representation input to the neural network layer. The first SDAE self-encoder is configured to extract a patient latent factor vector from the patient assistance information matrix, the patient latent factor vector being hidden layer output information of the first SDAE self-encoder. The second SDAE self-encoder is configured to extract a medication potential factor vector from the medication side information matrix, the medication potential factor vector being hidden layer output information of the second SDAE self-encoder. The neural network layer is configured to perform collaborative filtering processing on the drug potential vector, the drug regimen potential vector, the patient factor vector, and the medication regimen factor vector, thereby obtaining the predicted curative effect evaluation information.
To evaluate the performance of the prediction method, we used an 8-fold cross-validation method, since this method can provide a sufficiently accurate estimate of the true error rate. 80% of all data we collected were used as training set and 20% as testing set. And then, testing different characteristic parameters and evaluation curative effect parameters in the model, and respectively calculating the average error of the test experiment. In this way, the test and training will be repeated eight times to ensure the accuracy of the experiment. Generally, we set several basic indicators to evaluate the performance of the classification of the efficacy prediction model.
TP (true positive): the patient uses the drug regimen and produces better results consistent with the predicted classification results.
TN (true negative): the patient who used the regimen and produced better results is contrary to the predicted classification results.
FP (false positive): the patient used the drug regimen and produced a worse outcome as opposed to the predicted classification results.
TN (false negative): patients who have used this drug regimen and produced worse results consistent with the predicted classification results.
The case in which the predicted treatment effect evaluation value of the patient includes the medication schemes that have been used by the patient and have produced better effects among the K medication schemes ranked in the front is determined as PT. Similarly, if the first K predicted medication regimens for a patient include drugs that have been used by the patient and produced a worse outcome, then the case with the worse classification outcome is designated as PN.
We define the following as an evaluation to measure the prediction:
Figure BDA0002539787500000151
Figure BDA0002539787500000152
Figure BDA0002539787500000153
Figure BDA0002539787500000161
Figure BDA0002539787500000162
to facilitate a standardized metric, the number of instances result can be converted to a ratio between 0 and 1, defined by the several indices above. The relation between the selected K value and the prediction accuracy Ks and the error rate Kn is analyzed through the result of verifying the predictive performance of the model by the test set. As shown in fig. 3, the case where the predicted case with the medication efficacy better for the patient uses the case with the correct classification of the case is shown as the case with the correct classification of the case. It can be seen that the predicted effect will be better and better with increasing selected K values, and increase faster, indicating that the predicted regimen will be more biased towards the patient's clinically selected regimen. The other relation curve shows that the predicted example taking the medication scheme as a candidate under the condition that the medication curative effect of the patient is better occupies the condition that the proportion Kn of the example with the correct classification of the medication scheme is changed along with the change of the K value. Likewise, as the selected K value increases, the number of predicted erroneous samples increases, but slowly. When the K value is smaller, the number of counterexamples is less, and the value of K influences the accuracy of the medication scheme recommendation.
Fig. 4 shows the relationship between the importance parameter of the efficacy evaluation value and the importance parameter of the auxiliary information and the evaluation index Acc. It can be seen that when the area parameter in the evaluation value calculation has a large proportion, the accuracy is relatively improved, which may be because the change of the tumor marker is not stable, and there is a certain error in determining the efficacy evaluation. And the enlargement of the tumor area means the aggravation of the disease condition of the cancer, and the reduction of the area also means the improvement of the disease condition. Another set of parameters is the weights of the patient and drug side information in the training model, and it can be seen that the side information is important to improve the performance of the model. The auxiliary information of the medicine is more beneficial to improving the accuracy of the classification of the curative effect of the medicine.
Fig. 5 shows the relationship between the importance parameter of the efficacy evaluation value and the importance parameter of the auxiliary information and the evaluation index Sen. It can be seen that the sensitivity parameter decreases with increasing marker trade-off, and the magnitude of the increase and decrease is greater. This is because the area of cancer cells may not change much in some cases, and it is judged that positive cases are likely to be classified as negative cases only by the area. Similarly, a marker is less sensitive, and when its trade-off is too large, this will also be the case. Another set of parameters is the weights of the patient and drug assistance information in the training model, and it can be seen that the patient and drug regimen assistance information have a comparable impact on sensitivity.
FIG. 6 shows the relationship between the importance parameter of the efficacy evaluation value and the importance parameter of the auxiliary information and the evaluation index Spec. It can be seen that as the trade-off of the tumor marker increases, the specificity value increases after decreasing, and the increase is faster. It is likely that the area is more specific for the assessment modality and the specificity of tumor markers is limited. The other group of parameters is the proportion of the weight of the patient auxiliary information and the medication scheme auxiliary information in the training model to the proportion of the medication scheme which is actually not suitable for the patient classified into the suitable medicine. This may be because the patient's information cannot be better considered with a small trade-off of the patient's auxiliary information, thereby administering the medication to the patient. Selecting a suitable dosage regimen for the patient is helpful in improving the performance of the model.
Fig. 7 and 8 illustrate the change in sampling index in the case of three-line treatment of adenocarcinoma and squamous carcinoma patients, respectively. In order to facilitate response to the medication, we visually reflect the treatment condition of the patient through the CEA change and the curative effect evaluation value, so as to evaluate the recommended performance of the medication scheme selected by the medication decision system provided by the invention under different stages. In the figure, each node represents the response to the efficacy of the regimen during this dosing period, and we use the estimates to reflect the efficacy of the regimen. When CEA is elevated, we believe that the disease is repeated and it may be that the drug in the regimen develops resistance. Thus, we reselect the medication regimen for the patient's current condition. Performance under different probability parameter control and continuous drug selection. Similarly, for squamous carcinoma patients, we understand that squamous carcinoma patients are more sensitive to SCC-Ag. Similarly, we reflect the treatment status of patients after drug administration by the change of SCC-Ag and the evaluation of curative effect. In fig. 8, we see that the initial case node fluctuated more, the estimates were unstable and lower, and we believe that the patient just started to have no significant response to the dosing regimen. After the adaptive medication scheme of the patient is reselected through system adjustment, the detection index of the marker begins to decline in the next medication period, and the evaluation value also rises. After the disease condition is repeated, we can select the second line treatment and the back line treatment scheme to show that the effect is more stable. The medicine taking scheme has various types and choices, and the medicine taking scheme which is more suitable for the patient is selected for the patient, so that the medicine taking scheme at each stage recommended by the medicine taking decision system has a good improvement effect, and a doctor is helped to provide the treatment service of the patient in an all-round life cycle. Meanwhile, the method helps patients to know their own conditions in time and promotes early diagnosis of NSCLC.
Figure 9 shows the accuracy of the medication decision system of the present invention. Based on the records of NSCLC patients, we can calculate the accuracy of physician diagnostic decision medication and the accuracy of systematic prediction. As can be seen from the figure, the physician's decision is very accurate. In small sample data, the accuracy is as high as 98%. In the data of a large sample, the accuracy can reach 89%. Although the system has a lower accuracy in small sample data, it is between 0.32 and 0.47. This is because the sample size is small, the training data is small, the model generalization capability is poor, the conclusion of the system model is not enough to support the doctor's decision, and the efficiency of the doctor cannot be effectively improved. In large sample data, training data is increased, and the generalization capability of the model is enhanced. When the sample data reaches 4000, the accuracy can be improved to more than 80%.
While embodiments in accordance with the invention have been described above, these embodiments are not intended to be exhaustive or to limit the invention to the precise embodiments described. Many modifications and variations are possible in light of the above teaching. The embodiments were chosen and described in order to best explain the principles of the invention and the practical application, to thereby enable others skilled in the art to best utilize the invention and various embodiments with various modifications as are suited to the particular use contemplated. The invention is limited only by the claims and their full scope and equivalents.

Claims (10)

1. A medication decision method based on deep learning is characterized by comprising the following steps:
collecting patient-related information, medication regimen-related information, and actual efficacy assessment information of the medication regimen on the actual efficacy effect of the patient to obtain an analysis dataset,
constructing a therapeutic effect prediction model to analyze the relationship between the patients and the relationship between the medication options according to the analysis data set, and predicting therapeutic effect evaluation information of the medication options for the therapeutic effect of different patients by combining the actual therapeutic effect evaluation information,
recommending the medication scheme suitable for the individualized requirement of the patient to the patient according to the predicted curative effect evaluation information.
2. The deep learning-based medication decision method according to claim 1, wherein the step of the prediction model obtaining the predicted curative effect evaluation information comprises:
acquiring a relation matrix representing the relation between the patient and the medication scheme, wherein relation elements in the relation matrix represent the relation between the patient and the medication scheme, if the patient uses the medication scheme, the corresponding relation elements represent the actual curative effect evaluation value of the medication scheme on the patient, and if the patient does not use the medication scheme, the corresponding relation elements are set as a first fixed value,
obtaining patient potential vectors and the drug potential vectors from the relationship matrix, each of the patient potential vectors representing a relationship between one patient and a different medication regimen, each of the drug potential vectors representing a relationship between one medication regimen and a different patient,
respectively acquiring a patient auxiliary information matrix and a medication auxiliary information matrix according to the patient-related information and the medication-related information, wherein elements in the patient auxiliary information matrix represent information related to the patient, elements in the medication auxiliary information matrix represent information related to the medication,
extracting, by an SDAE autoencoder, a patient latent factor vector and a medication latent factor vector from the patient side information matrix and the medication side information matrix, respectively, the patient latent factor vector and the medication latent factor vector being hidden layer output information of the autoencoder,
and cooperatively filtering the drug potential vector, the drug regimen potential vector, the patient factor vector and the medication regimen factor vector through a neural network so as to learn the similarity between the patients and the similarity between the medication regimens and predict the treatment effect of each medication regimen on different patients by combining the relationship between the patients and the historical medication regimens of the patients so as to obtain the predicted treatment effect evaluation information.
3. The deep learning based medication decision method of claim 2, wherein the analysis dataset comprises a patient set and a medication scheme set, and the step of obtaining a patient potential vector and the medication potential vector from the relationship matrix comprises:
setting all non-first fixed values in the relationship matrix to a second fixed value to obtain a similarity relationship matrix,
according to the similarity relation matrix, obtaining each patient label vector representing each patient and each medicine label vector representing each medicine scheme, wherein except the ith element, the ith patient label vector in the patient set is the second fixed value, the rest elements are set as the first fixed value, except the jth element, the jth medicine scheme label vector in the medicine scheme set is the second fixed value, and the rest elements are set as the first fixed value,
extracting a patient-medication relation feature vector of the relation between the ith patient and each medication in the set of medications from the similarity relation matrix,
extracting a patient-medication regimen-patient relationship feature vector for the relationship between the jth of the medication regimen and each patient in the set of patients from the similarity relationship matrix,
concatenating the patient-medication regimen relationship feature vector with the ith patient label vector to obtain the patient potential vector for the ith patient,
and connecting the medication scheme-patient relation feature vector with the jth patient label vector to obtain the drug potential vector corresponding to the jth medication scheme.
4. The deep learning-based medication decision method according to claim 3, wherein after the elements in the patient side information matrix and the medication regimen side information matrix are randomly set to the first fixed values, respectively, the patient latent factor vector and the medication regimen latent factor vector are extracted by the SDAE self-encoder.
5. The deep learning-based medication decision method of claim 4, wherein the SDAE self-encoder is trained to learn the patient latent factor vector and the medication regimen latent vector by minimizing reconstruction errors.
6. The deep learning-based medication decision method of claim 5, wherein the loss function of the efficacy prediction model is composed of a minimized reconstruction error of the SDAE self-encoder during feature extraction and a prediction error of the efficacy prediction model for the efficacy prediction,
the prediction error is determined based on the predicted efficacy assessment information and the actual efficacy assessment information.
7. The deep learning-based medication decision method of claim 1, wherein recommending the medication regimen to the patient that is appropriate for the patient's personalized requirements based on the predicted efficacy assessment information comprises:
and sorting the predicted curative effect evaluation values in the predicted curative effect evaluation information from large to small, and selecting the medicine schemes corresponding to the first K predicted curative effect evaluation values in the sorting to recommend to the patient.
8. The deep learning based medication decision method of claim 3, wherein discrete patient potential vectors and medication regimen potential vectors are converted to continuous vector representations by an embedding layer for input to the neural network.
9. A deep learning based medication decision system, comprising: a data collection module, an efficacy prediction module, and a decision module configured to collect patient-related information, medication regimen-related information, and actual efficacy assessment information of the medication regimen on the actual efficacy effect of the patient, to obtain an analysis dataset,
the curative effect prediction module is configured to extract patient characteristic information and medication scheme characteristics according to the analysis data set, train the curative effect prediction model according to the extracted characteristic information to obtain predicted curative effect evaluation information of each medication scheme on different patient treatment effects,
the decision module is configured to recommend the medication scheme suitable for the individualized requirement of the patient to the patient according to the predicted curative effect evaluation information.
10. The deep learning-based medication decision method of claim 9, wherein the efficacy prediction module comprises:
an input layer configured to obtain a relationship matrix representing a relationship between the patient and the medication scheme according to the analysis data, wherein relationship elements in the relationship matrix represent the relationship between the patient and the medication scheme, if the patient uses the medication scheme, the corresponding relationship elements represent the actual efficacy evaluation value of the medication scheme on the patient, if the patient does not use the medication scheme, the corresponding relationship elements are set to a first fixed value, patient potential vectors and the medication potential vectors are obtained according to the relationship matrix, each of the patient potential vectors represents a relationship between one patient and a different medication scheme, each of the medication potential vectors represents a relationship between one medication scheme and a different patient, and a patient auxiliary information matrix and a medication scheme related information are respectively obtained according to the patient related information and the medication scheme related information An auxiliary information matrix, elements of the patient auxiliary information matrix representing information including the patient-related information, elements of the medication regime auxiliary information matrix representing the medication regime-related information,
an embedding layer configured to convert the patient potential vector and the medication regimen vector embedding from discrete data to a continuous vector representation input to the neural network layer,
a first SDAE self-encoder configured to extract a patient latent factor vector from the patient assistance information matrix, the patient latent factor vector being hidden layer output information of the first SDAE self-encoder,
a second SDAE self-encoder configured to extract a medication potential factor vector from the medication side information matrix, the medication potential factor vector being hidden layer output information of the second SDAE self-encoder,
a neural network layer configured to perform collaborative filtering processing on the drug potential vector, the drug regimen potential vector, the patient factor vector, and the medication regimen factor vector, thereby obtaining the predicted efficacy evaluation information.
CN202010543306.7A 2020-06-15 2020-06-15 Drug administration decision method and system based on deep learning Active CN111696678B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010543306.7A CN111696678B (en) 2020-06-15 2020-06-15 Drug administration decision method and system based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010543306.7A CN111696678B (en) 2020-06-15 2020-06-15 Drug administration decision method and system based on deep learning

Publications (2)

Publication Number Publication Date
CN111696678A true CN111696678A (en) 2020-09-22
CN111696678B CN111696678B (en) 2023-05-26

Family

ID=72481084

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010543306.7A Active CN111696678B (en) 2020-06-15 2020-06-15 Drug administration decision method and system based on deep learning

Country Status (1)

Country Link
CN (1) CN111696678B (en)

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113113117A (en) * 2021-03-25 2021-07-13 肖尧 Traditional Chinese medicine diagnosis and treatment system
CN113284627A (en) * 2021-04-15 2021-08-20 北京交通大学 Medication recommendation method based on patient characterization learning
CN114283949A (en) * 2021-11-08 2022-04-05 湖南云视数据科技有限责任公司 ADR (adaptive Doppler ratio) aided decision making system based on deep learning
CN115050451A (en) * 2022-08-17 2022-09-13 合肥工业大学 Automatic generation system for clinical sepsis medication scheme
CN115376706A (en) * 2022-10-26 2022-11-22 杭州艾名医学科技有限公司 Prediction model-based breast cancer drug scheme prediction method and device
CN116631578A (en) * 2023-07-25 2023-08-22 山东硕杰医疗科技有限公司 Lung cancer network comprehensive management information platform
CN117334347A (en) * 2023-12-01 2024-01-02 北京大学 Method, device, equipment and storage medium for evaluating treatment effect
CN117690597A (en) * 2023-11-29 2024-03-12 重庆爱永星辰标准医药服务有限公司 Biological medicine information clustering analysis method based on Internet

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2012123837A (en) * 2005-11-29 2012-06-28 Children's Hospital Medical Center Optimization and individualization of medicine selection and medication
CN107092797A (en) * 2017-04-26 2017-08-25 广东亿荣电子商务有限公司 A kind of medicine proposed algorithm based on deep learning
CN108062970A (en) * 2017-12-15 2018-05-22 泰康保险集团股份有限公司 Drug recommends method and device
CN108109700A (en) * 2017-12-19 2018-06-01 中国科学院深圳先进技术研究院 A kind of chronic disease Drug efficacy evaluation method and apparatus
US20190198142A1 (en) * 2017-12-26 2019-06-27 International Business Machines Corporation Medication decision support
CN110349681A (en) * 2019-07-11 2019-10-18 中南大学 Drug combination recommended method, system and device for non-small cell lung cancer
CN110827951A (en) * 2019-12-24 2020-02-21 陕西省中医医院 Clinical intelligent decision platform

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2012123837A (en) * 2005-11-29 2012-06-28 Children's Hospital Medical Center Optimization and individualization of medicine selection and medication
CN107092797A (en) * 2017-04-26 2017-08-25 广东亿荣电子商务有限公司 A kind of medicine proposed algorithm based on deep learning
CN108062970A (en) * 2017-12-15 2018-05-22 泰康保险集团股份有限公司 Drug recommends method and device
CN108109700A (en) * 2017-12-19 2018-06-01 中国科学院深圳先进技术研究院 A kind of chronic disease Drug efficacy evaluation method and apparatus
US20190198142A1 (en) * 2017-12-26 2019-06-27 International Business Machines Corporation Medication decision support
CN110349681A (en) * 2019-07-11 2019-10-18 中南大学 Drug combination recommended method, system and device for non-small cell lung cancer
CN110827951A (en) * 2019-12-24 2020-02-21 陕西省中医医院 Clinical intelligent decision platform

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113113117A (en) * 2021-03-25 2021-07-13 肖尧 Traditional Chinese medicine diagnosis and treatment system
CN113284627A (en) * 2021-04-15 2021-08-20 北京交通大学 Medication recommendation method based on patient characterization learning
CN113284627B (en) * 2021-04-15 2024-05-17 北京交通大学 Medication recommendation method based on patient characterization learning
CN114283949A (en) * 2021-11-08 2022-04-05 湖南云视数据科技有限责任公司 ADR (adaptive Doppler ratio) aided decision making system based on deep learning
CN115050451A (en) * 2022-08-17 2022-09-13 合肥工业大学 Automatic generation system for clinical sepsis medication scheme
CN115376706A (en) * 2022-10-26 2022-11-22 杭州艾名医学科技有限公司 Prediction model-based breast cancer drug scheme prediction method and device
CN116631578A (en) * 2023-07-25 2023-08-22 山东硕杰医疗科技有限公司 Lung cancer network comprehensive management information platform
CN116631578B (en) * 2023-07-25 2023-10-13 山东硕杰医疗科技有限公司 Lung cancer network comprehensive management information platform
CN117690597A (en) * 2023-11-29 2024-03-12 重庆爱永星辰标准医药服务有限公司 Biological medicine information clustering analysis method based on Internet
CN117334347A (en) * 2023-12-01 2024-01-02 北京大学 Method, device, equipment and storage medium for evaluating treatment effect
CN117334347B (en) * 2023-12-01 2024-03-22 北京大学 Method, device, equipment and storage medium for evaluating treatment effect

Also Published As

Publication number Publication date
CN111696678B (en) 2023-05-26

Similar Documents

Publication Publication Date Title
CN111696678B (en) Drug administration decision method and system based on deep learning
WO2021120936A1 (en) Chronic disease prediction system based on multi-task learning model
Spooner et al. A comparison of machine learning methods for survival analysis of high-dimensional clinical data for dementia prediction
Young et al. Accurate multimodal probabilistic prediction of conversion to Alzheimer's disease in patients with mild cognitive impairment
Yu et al. Medical decision support system for cancer treatment in precision medicine in developing countries
Afsaneh et al. Recent applications of machine learning and deep learning models in the prediction, diagnosis, and management of diabetes: a comprehensive review
Nsugbe Toward a self-supervised architecture for semen quality prediction using environmental and lifestyle factors
CN111798954A (en) Drug combination recommendation method based on time attention mechanism and graph convolution network
Ejiyi et al. A robust predictive diagnosis model for diabetes mellitus using Shapley-incorporated machine learning algorithms
CN113113130A (en) Tumor individualized diagnosis and treatment scheme recommendation method
Yu et al. A diagnostic prediction framework on auxiliary medical system for breast cancer in developing countries
Han et al. Risk prediction of diabetes and pre-diabetes based on physical examination data
Maheshwari et al. Quantum machine learning applied to electronic healthcare records for ischemic heart disease classification
Huang et al. Regularized continuous-time Markov model via elastic net
Oliullah et al. Analyzing the effectiveness of several machine learning methods for heart attack prediction
Chu et al. Machine learning algorithms for predicting the risk of fracture in patients with diabetes in china
US20240303544A1 (en) Graph database techniques for machine learning
Qin A Prediction Model of Diabetes Based on Ensemble Learning
Sonia et al. An empirical evaluation of benchmark machine learning classifiers for risk prediction of cardiovascular disease in diabetic males
Saju et al. Heart Disease Prediction Model using Machine Learning
Wang et al. Prediction of target range of intact parathyroid hormone in hemodialysis patients with artificial neural network
Yadav et al. Comparison of machine learning techniques for precision in measurement of glucose level in artificial pancreas
Mohamed et al. Predicting The Heart Attack From Accessible Patients Medical Datasets Using Data Mining Technique
KR102691998B1 (en) Method for diabetes prediction by using genetic data applying race-specific frequency and clinical data and system thereof
Vieira On the Short-Term Prediction of Multiple Sclerosis Disease Progression

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant