Using Features Extracted from Vital Time Series for Early Prediction of Sepsis

Abstract

To get early prediction of sepsis, we propose to extract more time-dependent characteristics that retain the temporal evolvement information of the underlying biomedical dynamic system, including differential, integration, time-dependent statistics, variations and convolutions. Considering that two categories are unbalanced in the training set, we employed easy ensemble algorithm to get multiple base learners. As for the base learner, we tried three models: random forest, XGBoost and LightGBM. By boosting the results of multiple base learners, we constructed our ensemble model. Our team which name is njuedu ranked 25th in the official test and scored 0.282 in full test set.Since the submitted model version only used training set A to train our model, the model had a higher score of 0.401 in test set A, and 0.278 in test set B, and only-0.207 points in test set C. 1. Introduction Prediction is of great importance in biomedical field, e.g. early goal-directed therapy provides significant benefits with respect to outcome in patients with severe sep-sis and septic shock[1-3]. To achieve the real-meaning 'pre'diction, it means the model has to rely on the history information to get a current prediction. Therefore, specific models with memory have been developed, such as hidden Markov model[4], long short-term memory recurrent neu-ral networks[5-7], etc. However, memory-units also introduce dynamical complexity, and might impose the model on a risk of instability. There have been several studies about early prediction of sepsis. Desautels and colleagues[8] tried to use the Insight model for severe sepsis detection and got an AUC of 0.75. Mao and colleagues[9] validated InSight based on a retrospective dataset from the mixed ward of the University of California, San Francisco (UCSF) Medical Center (San Francisco, Calif.) to detect and predict three gold standards associated with sepsis and achieve AUC of 0.92 and 0.87 on sepsis and severe sepsis respectively.Kam and Kim[10] used a deep learning model to create an early sep-sis prediction system and validated its feature extraction capabilities. The best result they got was an AUC of 0.929 using the LSTM variant. They followed the feature extraction steps of [11]. In this manuscript, we tried to extract various time-dependent characteristics from the time series, and use the derived features as the input of regular machine learning model like random forest, XGBoost[12] and LightGB-M[13] etc. 2. Methods We use random forest, XGBoost and LightGBM as prediction models. However, most efforts have been paid to the pre-processing and time-dependent characteristics extraction. And all data descriptions can be found in [14]. 2.1. Pre-processing 2.1.1. NAN replacement Since we will focus on the temporal evolvement of the biomedical indices, it is necessary to replace the NAN in the original data with meaningful values in order to facilitate extraction of certain time dependent characteristics.We can see from Figure 1 that the data is missing very badly. We use the NAN replacement rules as the following: if at = N AN, then at = a i i = max(Φ), Φ = µa Φ = (1) in which a t represents the value of characteristic a taken at time t, µ a is the arithmetic mean of characteristic a in training set, Φ := {j|a j = N AN, j ∈ [max(1, t − 3), t)}, and denotes the empty set. This is also a combination of the forward-fill and mean-fill method. 2.1.2. Feature extraction We will treat every record as an independent sample in the task. That means we will lose most evolvement in