| With the rapid accumulation of electronic health records(EHRs)and the significant progress of intelligent information technology,the medical big data task for EHRs has been widely studied and applied.Among them,deep learning models have made remarkable achievements.These models can provide auxiliary information for doctors’ decision-making and help doctors conduct clinical analysis,which is of great significance to promote the development of intelligent medicine.However,the existing methods still have limitations in EHRs based time series medical data representation learning:(1)EHRs data is characterized by irregular time interval and large difference in time granularity.Existing models are difficult to effectively use irregular time interval information and integrate time-series medical data with different time granularity,which makes it difficult for models to accurately capture complex time-series medical models;(2)EHRs data contains rich medical project data.The existing time series model usually treats the medical projects in a single access record as a whole,which makes it difficult to capture the time series association relationship of medical projects in the time axis direction,resulting in the model is difficult to accurately capture the time series pattern contained in medical projects;(3)In time series medical data mining tasks,the existing time series models only rely on the information of current patients to make predictions,and fail to explicitly consider the impact of similar patients’ behaviors on current patients,which makes it difficult for the models to make full use of the medical mode information contained in similar patients;(4)In the task of time series medical data mining,the interpretability of time series models has a significance that cannot be ignored.The existing models are either lack of interpretability mechanisms or rely on well-designed structures.The lack of interpretability will lead to the model being difficult to apply to actual clinical decisions.The well-designed interpretability structures limit the flexibility of the model itself in reverse,resulting in the model’s expression learning ability being restricted.Therefore,based on EHRs data,the temporal pattern mining method and model interpretability method are studied.The specific research contents are as follows:(1)To solve the problem that the existing models are difficult to effectively use the irregular time interval information and integrate the time series data with different time granularity,a representation learning model based on Irregular Time interval awareness and Multi Granularity Time series fusion(ITMGT)is proposed.This model processes different types of temporal medical data based on multiple sequence inputs,so as to achieve effective embedding of different types of temporal medical data;Multi sequence medical data are fused based on different time granularity to capture the interaction between multi-sequence input data;Finally,based on the attention mechanism oriented to multi granularity time series,the weight distribution of time series medical data in the time axis direction is realized,so that the model can focus on important data and predict according to the multi granularity time series pattern.(2)To solve the problem that existing temporal pattern mining methods are difficult to capture the temporal association between complex medical items,the Fine-Grained Temporal Association(FGTA)representation learning model based on information dissemination is proposed.Firstly,the medical items of two adjacent medical records are formed into a bipartite graph.The edge of the graph is from the medical items of the previous medical record to the medical items of the next medical record,representing the temporal correlation between adjacent medical items;Then,the neural network model based on information dissemination is used to capture this fine-grained temporal correlation;Finally,it is combined with the existing time series model to achieve fine-grained time series relationship representation learning and medical item representation learning.(3)To solve the problem that the existing time series models fail to explicitly consider the impact of similar patients’ behaviors on current patients,the Event Graph Neural Network(EGNN)representation learning model for similar patients’ perception is proposed.First,the EHRs data is modeled as a time series medical graph,in which the patient and patient obtain the association relationship through the information of their disease,history hospital,etc.,and the patient’s medical time series information obtains the association relationship through the event node;Then,based on graph neural network representation learning method,the model performs attention based aggregation and gating mechanism based aggregation for conventional type neighbors and event type neighbors respectively,so that when learning the timing information of patients,the perceptual domain of graph neural network is used to obtain behavior information of similar patients,and the model’s representation learning ability is enhanced.(4)To solve the problem that the existing time series model is difficult to obtain interpretability,the model-agnostic Personal Risk Factor Evaluation(PRFE)method for deep learning models is proposed to provide interpretability for intelligent medical models.This method regards the risk prediction model based on deep learning as a multivariable function,generates the fastest path of risk change through iteration,and calculates the cumulative risk of each feature based on this path to provide interpretability for the time series model.In summary,a variety of time series pattern mining methods is proposed in this paper to accurately mine the hidden information in the EHRs’ time series data,and a personalized interpretable analysis for the prediction results is provided.In this paper,we selected the typical survival prediction tasks and disease prediction tasks in clinical applications and conducted experiments based on MIMIC-Ⅲ and MIMIC-Ⅳ data sets. |
PDF Full Text Request