Gaussian Process Regression For EHR Time Series Interpolation And Its Application In AKI Prediction

OBJECTIVE:Acute Kidney Injury(AKI)is a common clinical event associated with increased risk of hospital mortality and additional healthcare-related expenditures.Early prediction of AKI can help clinicians to identify patient' risk of deterioration earlier and intervene in early stage of disease progression to prevent further damage.The development of EHR systems and the application of data mining and machine learning methods in healthcare can provide us diverse approaches to explore the trajectory of diseases and also provide us an opportunity for identifying AKI risk at early stage.However,clinical observations are always sampled irregularly,with rates varying between patients,variables,and even through time.The strategy of dealing with missing data is worth discussing in the secondary analysis of EHR.This study discussed how to choose interpolation methods for medical time series,and developed models for AKI prediction using single time point value and time series based on machine learning and deep learning methods which could provide support for clinical decision.METHODS:This study ?.)evaluated common interpolation methods for physiological data missingness including Gaussian Process(GP),linear regression and spine regression.?.)Predictive models for AKI were also developed based on LightGBM,considering the first physiological and lab measurements within 24h of each patient as predictors to predict AKI in their stay.?.)Time series prediction models utilizing the multivariate attention LSTM model(MALSTM-FCN)were developed to predict AKI risk at 2h,6h,12h and 24h after ICU admission.Four interpolation approaches for EHR were compared via model performance.RESULTS:?.)GPR achieved the lowest MSE among three methods in an experiment of heart rate missingness interpolation,showing its superiority in handling physiological time-series data.?.)There are 21722 patients in MIMIC-? included in our cohort,8694 of them developed AKI in their stay(40.0%).The performance of LightGBM showed strong predictive power with AUC of 0.93 and ACC of 0.87,while AUC for random forest and logistic regression is 0.90 and 0.74,ACC 0.86 and 0.75 respectively.?.)Results of the MALSTM-FCN time series prediction models and LightGBM model showed that the AUC of 2h and 6h prediction models are 0.908(LightGBM-Linear Interpolation Model)and 0.902(MALSTM-GP Model and LightGBM-Linear Interpolation Model),the best AUC for 12h and 24h models are 0.890,0.877(MALSTM-GP Interpolation)respectively.MALSTM-GP Interpolation Model achieved best ACC of 0.819,0.818,0.806 and 0.798 for 2h,6h,12h and 24h model respectivelyCONCLUTION:Interpolation based on GP is superior in dealing with physiological signal missingness and machine learning methods provide effective approaches for developing disease prediction models.The proposed predicting model using LightGBM-Linear Interpolation model achieved best AUC of 0.908 in 2h,while MALSTM-GP achieved the best performance for predicting AKI in 6h,12h and 24h.