Research On Medical-Event Prediction Methods From Clinical Electronic Health Record

Medical event prediction from clinical electronic health records aims to analyze patient electronic health record data and build computational models to predict future medical events such as medication utilization,clinical diagnosis,and risk of mortality.This allows for a systematic study of patient electronic health records,overcomes the limitations of existing prediction models,and meets the needs for medical event prevention and management.Clinical electronic health records are temporal and have multiple sources.How to build computational models for medical event prediction based on these data characteristics has become a key focus of clinical decision-making research.Currently,many computational methods have been proposed to help clinical medical event predictions based on patient electronic health records.However,these methods generally have the following limitations: ignoring the diferences and correlations between medical events,lacking fexibility in predicting diferent types of medical events,lacking efective methods for modeling complementary information between structured and unstructured records in multi-source electronic health records,and ignoring medical domain knowledge.To address these issues mentioned above,this dissertation proposes multiple medical event prediction methods based on patient clinical electronic health records,aiming to fully exploit the temporal and multiple-source representation patterns of electronic health records and improve the accuracy of medical event prediction.The main research content and innovation points of this paper including:(1)This dissertation proposes a medical event trend prediction method based on multi-type event interactions to address the problem that the diferences among multiple medical events are easily ignored in structured electronic medical event prediction.The global multi-type event interaction-based medical event prediction method uses the temporal information of patients’ historical visit records to construct a medical event-level temporal relationship graph.Patient representations are modeled from the medical event temporal relationship graph and the visit sequence.The global representation of the medical events is obtained based on the medical event temporal relationship graph representation and combined with the temporal sequence representation of the patient’s visit level as the fnal representation of the patient.The local multi-type event interaction-based medical event trend prediction method is based on the patient’s current visit status and multi-type medical event interaction.Experimental results demonstrate that the proposed medical event trend prediction methods can efectively learn the interaction patterns between a patient’s multiple types of medical events and improve the performance of medical event trend prediction models in structured electronic health records.(2)This dissertation proposes a medical event prediction method based on a crossevent attention mechanism to address the lack of fexibility in multi-type medical event prediction models in structured electronic health records.The method designs two taskaware cross-event attention based on the medical event prediction task type.This method encodes the representations of diferent types of medical events in patient records in a unifed way,thereby improving the fexibility of multi-type medical event prediction.Experimental results demonstrate that the proposed method can efectively solve the fexibility problem in multi-type medical event prediction,enhance patient representation ability,and improve the performance of multi-type medical event prediction models.(3)This dissertation proposes a pre-training-based medical event prediction method using multi-modal electronic health records to address the insufcient modeling of complementary information between structured and unstructured records in multi-source electronic health records.The method constructs a pre-training paradigm that combines structured and unstructured data in electronic health records.This pre-training model captures the complementary semantic properties of structured and unstructured data through crossmodal attention interaction and designs two cross-modal pre-training tasks as optimization objectives,fne-tuned on three clinical medical event prediction tasks.Experimental results show that the proposed multi-modal electronic health record pre-training method can efectively capture the complementary semantic properties between structured and unstructured electronic health records.The model performs excellently on multiple downstream tasks and exhibits stable performance on small-sample datasets,demonstrating the learning ability of the proposed multi-modal electronic health record pre-training model.In addition,the proposed multi-modal electronic health record pre-training model also shows signifcant performance improvements when applied to tasks using structured or unstructured data separately.(4)This dissertation proposes a domain knowledge-enhanced method for clinical medical event prediction to address the problem of insufcient utilization of medical domain knowledge to enrich the representation of patient electronic health records in clinical event prediction tasks.The method frst extracts relation triples from medical domain knowledge to construct a medical domain knowledge subgraph.Then it uses knowledge aggregation modules based on experience and electronic health records to obtain patient domain knowledge-enhanced representations.This dissertation adopts a method of minimizing the upper bound of mutual information to constrain the patient’s original representation and the domain knowledge-enhanced representation to ensure that the knowledgeenhanced method obtains as much domain knowledge as possible.Experimental results show that the proposed method can efectively integrate medical domain knowledge and patient electronic health records to improve the performance of medical event prediction models.