Research And Application On Key Techniques In ICU-oriented Medical Data Mining

The health care system is in the midst of enormous change in recent years,and the way about healthcare delivery has evolved dramatically,such as health care providers,hospital documents,bedsides monitoring,and information sharing.There was a geometric series increase of electronic health records(EHRs)as hospitals and ambulatory care providers transit from paper-based to electronic record-keeping systems.However,due to the EHR data does not provide any useful information itself,we cannot take value directly from raw EHRs.Thus,the "data explosion,information scarcity" is still a vital problem to solve.How to acquire knowledge from EHRs to benefit clinical,medical,healthcare,and society? Processing EHRs manually is not only time-consuming and labor-intensive,but also with low quality.Moreover,it will be impossible to make complex decisions based on the manual crafted information,and precision medicine will never become a reality nor a possibility.Data mining is a suitable technique that can help transform EHR data into information.In this thesis,we use the intensive care unit(ICU)data as an instance,concentrate on three critical problems in mining medical data: 1)disease diagnosis,2)mortality prediction 3)clinical missing data imputation.The main contributions are listed as follows:1.Proposed a deep multi-source multi-task attention model for ICU disease diagnosis.Disease diagnosis can provide essential information for clinical decisions that influence the outcome of serious illness.In this thesis,we proposed a deep multi-source multitask attention model(DMMAM)for ICU disease diagnosis.DMMAM can utilize a unified model to diagnose all common diseases with higher diagnostic accuracy.As previous works mainly focused on specific disease diagnosis,different models need to be developed for various conditions.If applied to ICU,dozens of such models need to be deployed,which will not only be time-consuming but also resource-intensive,and model compatibility is also a big barrier.By considering the integrity of the physiological features of a patient and the relationships among different diseases,DMMAM first exploits multi-sources information from multiple types of complications,clinical measurements,and medical treatments to finely describe the patient's physical state.Then,utilize this state and multi-task theory to diagnose ICU common diseases.Technically,to improve the diagnosis accuracy,DMMAM makes full use of the interdependence between primary disease and their complications,as well as the temporal correlations in the diagnosis process.In this thesis,we use a real-world dataset,which contains 167884 patients,to validate the effectiveness of the proposed method.The experimental results show that DMMAM achieved state-of-the-art performance.2.Proposed an interpretable model for ICU mortality prediction.Estimating the mortality of patients plays a fundamental role in ICU and is usually used to assess the severity of a patient's condition.In a typical ICU scenario,scoring systems are adopted to evaluate mortality.As scoring systems use a relatively small group of physiological features,the accuracy of the assessment cannot be guaranteed.Recently,researchers have begun to use deep learning-based methods to predict mortality and improved evaluation performance considerably.However,the "black box" problem produced by the deep learning models has not been solved yet.So,the prospects of clinical adoption are not promising.To address the above issues,we proposed an interpretable deep learning model to predict the mortality for ICU patients.The proposed method fully considered the uniformity of the patient's physiological features and the temporal information between measurements and treatments.The proposed model can fine-tune the parameters separately according to the properties of different diseases.We add an interpreter to enhance the model's interpretability,which uses the attention mechanism and external auxiliary information to interpret the prediction results.In this thesis,we use a real-world dataset to validate the correctness and performance of the proposed model.The experimental results show that the proposed model can keep both the best prediction performance and good model interpretability.3.Proposed an ICU integrated disease diagnosis and mortality prediction model(IDDSAM).Disease diagnosis and mortality prediction are two of the essential tasks in medical practice.To rescue lives,expeditious diagnosis and patient mortality assessment are required in ICU.Unlike the existing approaches where diagnosis and severity assessment are studied separately,in this thesis,we treat these actions as two tasks in an integrated procedure that clinicians must be able to quickly and accurately conduct such that patients are given the best possible chance for therapeutic success.The model first utilizes a variety of data sources to fully describe the patient's physiological state at each time windows and then uses a window alignment operation to handle the time delay among different data sources.Next,connect these physiological states along with time windows in a sequential manner,and use a recurrent neural network to model the data.Finally,by using the multi-task mechanism and hidden space sharing,the diagnosis results and mortality prediction results are given out,simultaneously.Besides,we solved the imbalance problem by introducing a local loss function and improves prediction accuracy.In this thesis,we use a real-world dataset,MIMIC-III,to validate the justifiability and the performance of the proposed model.The experimental results show that,compared with baselines,the proposed model achieved the state-of-the-art performance.4.Proposed a deep dynamic imputation method to solve the missing value problem in clinical time-series data.In clinical time-series data,missing values are pervasive and inevitable,which not only raises the complexity and difficulty of analyzing the data but also leads to biased results.This problem also occurred in our previous work of disease diagnosis and mortality prediction,which decreased the model performance greatly.To improve the performance of learning models and benefit related research,in this thesis,we proposed a deep learning-based missing value imputation method and integrated it into a mortality prediction process in an end-to-end manner.The proposed method makes use of the missing patterns of the clinical time-series data(such as missing value mask,time interval,burstiness,the cumulative missing rate,et al.)for imputation initialization.It then uses a gated recurrent unit to optimize and update the imputed values,iteratively.Next,we use a quality assurance mechanism to evaluate the imputed values comprehensively.After that,the fully imputed data will be sent to downstream applications,uniformly,for specific tasks.The proposed method is specifically designed to deal with the missing value problem in medical multivariate time series data.The advantages are summarized as follows: 1)it can impute missing values without any assumptions,2)with an imputation quality assurance mechanism,the imputation accuracy is high,3)it can be integrated into downstream applications endto-end,easily.Extensive experiments on a real-world ICU dataset demonstrate that: 1)the proposed model outperforms the baselines in terms of imputation quality and prediction accuracy,2)with the same conditions,the higher accuracy of the imputation,the better performance of downstream applications.