Research Of Disease Progression Prediction Based On Irregular Visit Interval

Benefiting from the explosive growth of data and the rapid development of computing power,artificial intelligence technology represented by deep neural networks has developed rapidly.In the medical and health field,training deep learning models based on historically accumulated data to conduct intelligent diagnosis and auxiliary decision-making for specific tasks can overcome the limitations of doctors’ individual experience to a certain extent and help ordinary doctors to diagnose and make decisions at an expert level.In particular,the application of deep learning technology to disease progression prediction,early warning and personalized early intervention for patients’ disease development is of great significance,and is a hot issue that many scholars at home and abroad pay attention to.Since the historical data accumulated by patients can be organized into sequences in chronological order,many existing studies build disease progression prediction models based on recurrent neural networks.However,the above methods organize historical diagnosis and treatment data by sequential positions,ignoring the timestamp and time interval of the data.In the medical field,patient visits have certain irregularities,resulting in serious time interval irregularities in the data.This paper explores the impact of the irregular time interval nature of the data on the model and proposes a disease progression prediction model based on irregular visit intervals.On this basis,this paper studies the application of the proposed model in the federated learning framework,and proposes a corresponding solution to the non-IID problem of federated learning data.The main contributions of this paper are as follows:（1）Analyze the impact of irregular time intervals on prediction performance.This paper analyzes the influence of irregular time interval on the recurrent neural network by comparing the prediction performance of the recurrent neural network constructed on the sequence data with regular time interval and irregular time interval.In this paper,a method for generating irregularly spaced datasets based on custom time series is proposed,and a comparative analysis is carried out on this basis.Through experiments,it is found that in addition to the irregularity of the time interval（input interval）of the input sequence,the irregularity of the time interval（prediction interval）between the last time point of the input sequence and the predicted time point also has a negative impact on the model prediction performance.（2）Establish a disease progression prediction model based on irregular visit intervals.In this paper,a disease progression prediction model TA2D-LSTM is proposed,which comprehensively considers the influence of irregular characteristics of different time intervals on the model and jointly considers the contribution difference of different medical treatment points.The model consists of two parts,input sequence modeling and predicted point modeling.In the first part,a time-aware TLSTM unit is used to deal with irregular input intervals,an attention mechanism is used to learn the different contributions of all input visits,and the input sequence is jointly modeled.In the second part,the outputs of the first part are adjusted using two information decay modules DECAY,to deal with the irregularity of the prediction interval,and the adjusted outputs are combined as the predicted point modeling for the final prediction.Finally,experiments are performed on real disease datasets and artificially generated datasets to verify the effectiveness of the proposed model.（3）Propose solutions based on federated learning application scenarios.The TA2D-LSTM model proposed in this paper is applied to the federated learning framework,the impact of data nonIID on the model is analyzed,and the shortcomings of two representative processing methods,FedShare and FedProx-D,are verified.On this basis,the fusion method of the two is applied,which alleviates the lack of privacy protection of FedShare and obtains better performance,which provides an effective way for the application of the proposed model under the framework of federated learning.