Study On Medications Prediction Algorithms Based On The Mining Of Multivariate Correlations Among Heterogeneous Electronic Medical Records Data

Breakthroughs in key technologies of artificial intelligence,development of data storage technology,and support from national policies provide a solid foundation for medical and health big data mining.Among them,intelligent treatment medication decision support system can assist doctors to more efficiently select and make the best treatment plan and treatment medication combination that is beneficial to patients,so as to better alleviate the current situation of inconsistent medical resources and realistic needs.However,there are still many challenging problems to be solved due to the particularity of clinical actual scene requirements and the multiple and complex correlation of multi-source heterogeneous clinical electronic medical record data.Including how to deal effectively with time-varying dynamic interactions among heterogeneous sequences,how to build the correlations among hierarchical tasks in multi-task scene,how to model the cascade causal correlations among hierarchical sequences in heterogeneous electronic medical records data,how to effectively mine the multivariate correlations among medical codes from multi-sourced medical knowledge.In view of the above problems and challenges,this dissertation focus on the demands for treatment medications decision-making in different clinical scenarios during patients’ whole medical cycle,and explored effective mining methods for complex multivariate correlation between multi-source heterogeneous clinical electronic medical record data.The main content of this dissertation can be summarized as follows:1.Firstly,considering the mining of time-varying dynamic interactions among heterogeneous sequences in electronic medical records data,a medication prediction algorithm with dual adaptive sequential networks is proposed for the clinical scenario in which first-time patients need dynamic treatment intervention.Based on the long and short memory neural network,firstly the decomposed adaptive LSTM network utilizes the context fusion module to construct the auxiliary input for the interactive sequence,and then the meta-learning network is constructed to provide dynamic parameter weights for it to construct the time-varying correlation of the sequence interaction and capture the time correlation within the sequence.Finally,the multi-source data embedding representation is integrated into the patient representation according to the correlation between the multi-source data embedding representations and the treatment medications through the attentive fusion network,to predict the treatment medications in the next stage.Experimental results on open data sets show that the proposed end-to-end medications prediction algorithm can effectively mine multiple correlations between electronic medical records data in this clinical scenario,and the critical role of the algorithm substructure is verified by ablation experiments.2.Secondly,considering the mining of cascade correlations among hierarchical tasks,a medication prediction algorithm with relation augmented hierarchical multi-task learning network for the medical scenario where discharged patients need to proactively prepare drugs to deal with the risk of disease recurrence in the future.The algorithm is based on the hierarchically causal correlation between diagnosed disease and treatment medications in electronic medical records.The algorithm firstly captures the correlation between medical codes by the multi-head attention network in medical codes embedding module of attention,then mines the correlation between disease sequence and medication sequence by the relation-aware long short memory neural network in medication recommendation module,and at the same time utilizes the pseudoresidual structure to realize cross-layer transmission of information.Finally,the obtained patient representations from the disease embedding module and the medications embedding mudule are respectively used to predict future diseases and recommend treatment medications.Experimental results on open data sets show that compared with single task and traditional multi-task learning methods,the relation augmented hierarchical multi-task learning method proposed in this dissertation can effectively mine and utilize the task-level correlation to achieve decision support for active disease prevention.3.Thirdly,considering the mining of implicit cascade causal correlations among multilevel hierarchical sequences in electronic medical records data,a medication prediction algorithm with multilevel selective and interactive network is proposed for the clinical scenarios where readmission patients need to be comprehensively evaluated for medications change.The algorithm is based on the multi-level hierarchical causal structure similar to the treatment path in the electronic medical record data.Firstly,the interactive short and long-term memory network is proposed to capture not only the influence of the auxiliary sequence on the integration of the history information of the main sequence,but also the influence of the current input of the main sequence,so as to effectively enhance the interaction between the cascade sequences.Then,the selection module based on sparse attention was used to focus on the information more relevant to the prediction task,and the secondary information was attenuated or ignored to reduce the influence of irrelevant feature noise.Finally,the global selective fusion module based on attention was used to calculate the patient representation vector according to the correlation between the multi-source heterogeneous sequence embedding representation and the predicted medications to predict the current treatment of the patient.Experimental results show that the proposed algorithm is superior to the existing prediction algorithm or recommendation algorithm in many evaluation indexes,and can better capture the causal correlation of multi-level hierarchical causal structure of electronic medical record data and reduce the impact of noise.4.Fouthly,considering the mining of multivariate correlations among medical codes implicit in multi-sourced medical knowledge,a medications prediction algorithm with multi-sourced knowledge via multi-level graph contrastive learning is proposed for the clinical scenario of multi-expert joint consultation for patients.The algorithm is based on the multi-source medical knowledge from many medicine experts including medical domain knowledge and medical prior relations,and it first captures the implicit correlation between medical codes through the medical domain ontology graph with unsupervised graph contrastive learning method,then captures the explicit correlation between medical codes through the medical prior relation graph with unsupervised graph contrastive learning method.In this way,the knowledge-augmented and relationaugmented medical codes representations can be obtained respectively.Finally,the augmented medical codes representation and supervised medical codes representations are combined into the sequence learning network through the downstream treatment medications prediction network to capture the temporal correlation between medical codes sequences.The experimental results show that the proposed method can obtain better medication prediction performance than the existing ontology knowledge graph enhancement algorithm by mining the multivariate correlation between medical codes in medical domain knowledge,and can assist doctors to make clinical decision of treatment medication more effectively.For theoretical contributions,considering the mining of complex multivariate correlations such as multi-source heterogeneity,time-varying dynamic interactions and cascade causal correlations among EMR data,this thesis proposes four algorithms:dual adaptive sequential learning network,relation augmented hierarchical multi-task learning network,multilevel selective and interactive network and multi-sourced knowledge augmented network.For practical values,the proposed algorithm can provide doctors with comprehensive treatment and medication decisionmaking assistance in different clinical scenarios during patients’ whole medical cycle.