Objectives: Early detection of sepsis is critical in clinical practice since each hour of delayed treatment has been associated with an increase in mortality due to irreversible organ damage. This study aimed to develop an explainable artificial intelligence model for early predicting sepsis by analyzing the electronic health record data from ICU provided by the PhysioNet/Computing in Cardiology Challenge 2019. Design: Retrospective observational study. Setting: We developed our model on the shared ICUs publicly data and verified on the full hidden populations for challenge scoring. Patients: Public database included 40,336 patients’ electronic health records sourced from Beth Israel Deaconess Medical Center (hospital system A) and Emory University Hospital (hospital system B). A total of 24,819 patients from hospital systems A, B, and C (an unidentified hospital system) were sequestered as full hidden test sets. Interventions: None. Measurements and Main Results: A total of 168 features were extracted on hourly basis. Explainable artificial intelligence sepsis predictor model was trained to predict sepsis in real time. Impact of each feature on hourly sepsis prediction was explored in-depth to show the interpretability. The algorithm demonstrated the final clinical utility score of 0.364 in this challenge when tested on the full hidden test sets, and the scores on three separate test sets were 0.430, 0.422, and –0.048, respectively. Conclusions: Explainable artificial intelligence sepsis predictor model achieves superior performance for predicting sepsis risk in a real-time way and provides interpretable information for understanding sepsis risk in ICU.

Early prediction of sepsis is critical in clinical practicesince each hour of delayed treatment has been associated with an increase in mortality due to irreversible organ damage. This study aimed to develop an algorithm for accurately predicting the onset of sepsis in the proceeding of six hours. Firstly, we selected 37 available variates features after data pre-processing, and then extracted three kinds of features as well in this paper, including 62 missing value features, 8 scoring quantified features and 61 time series features. After that, a multi-feature fusion based XGBoost classification model was developed and was further improved by a Bayesian optimizer and an ensemble learning framework. Analysis was performed on the PhysioNet/Computing in Cardiology Challenge 2019, which provided a publicly available sepsis data sourced from 40,336 ICU patients. Finally, after searching an optimized predicted risk threshold of 0.522 through the official submissions, our team "SailOcean" applied the developed model on the full hidden test set of 24,819 ICU patients from three hospital systems and obtained a final Unormalized score (U-Score) defined by the organizers of 0.364, which was the highest unofficial score.

How the complexity or irregularity of heart rate variability (HRV) changes across different sleep stages and the importance of these features in sleep staging are not fully understood. This study aimed to investigate the complexity or irregularity of the RR interval time series in different sleep stages and explore their values in sleep staging. We performed approximate entropy (ApEn), sample entropy (SampEn), fuzzy entropy (FuzzyEn), distribution entropy (DistEn), conditional entropy (CE), and permutation entropy (PermEn) analyses on RR interval time series extracted from epochs that were constructed based on two methods: (1) 270-s epoch length and (2) 300-s epoch length. To test whether adding the entropy measures can improve the accuracy of sleep staging using linear HRV indices, XGBoost was used to examine the abilities to differentiate among: (i) 5 classes [Wake (W), non-rapid-eye-movement (NREM), which can be divide into 3 sub-stages: stage N1, stage N2, and stage N3, and rapid-eye-movement (REM)]; (ii) 4 classes [W, light sleep (combined N1 and N2), deep sleep (N3), and REM]; and (iii) 3 classes: (W, NREM, and REM). SampEn, FuzzyEn, and CE significantly increased from W to N3 and decreased in REM. DistEn increased from W to N1, decreased in N2, and further decreased in N3; it increased in REM. The average accuracy of the three tasks using linear and entropy features were 42.1%, 59.1%, and 60.8%, respectively, based on 270-s epoch length; all were significantly lower than the performance based on 300-s epoch length (i.e., 54.3%, 63.1%, and 67.5%, respectively). Adding entropy measures to the XGBoost model of linear parameters did not significantly improve the classification performance. However, entropy measures, especially PermEn, DistEn, and FuzzyEn, demonstrated greater importance than most of the linear parameters in the XGBoost model.300-s270-s.

No abstract