Data completion: We fill in the first row’s NaNs per patient using the medians (grouped by 7 age groups, gender and Unit1) summarized from the cohort, and replace subsequent NaNs with the previous non-missing values. We clip each column with (.5%, 99.5%) quantiles and do min-max normalizations. Methods: We explore two basic methods during this unofficial phase: 1) Train machine learning (ML) methods, including Logistic Regression, AdaBoost and Random Forest, with handcraft features. That is, in addition to the 34 variables, we compute 2 extra ratios of HR/SBP and BUN/Creatinine and derive 11 additional features per row per variable, e.g., if a value is newly measured, the mean/min/max/(max-min) in the past 6 hrs, if the value increases/decreases in the past 6 (or 1) hrs, etc. 2) Train LSTM per row on the past 6-hr observations (zero padded) with no handcraft features. During training, for making early predictions, we relabel each row as 1 if sepsis happens in the next 6 (or 12) hours and 0 otherwise. We randomly select 400 patients with original labels from set A and B, holding them out for testing. The best held-out utility score we get from ML methods (Random Forest) is 0.25 (AUC 0.76) on the 6-hr earlier labeled data; the higher score we get from 1-layer LSTM is 0.39 (AUC 0.84) on the 12-hr earlier labeled data. We submitted the latter one and obtained utility score of 0.32. Plan for official phase: We will continue working on both lines of methods, one focusing on refining better handcraft features and one focusing on tuning and constructing more complex models (e.g., LSTM with more layers, attention, bidirectional, etc). Two methods can be possibly merged at the end. Besides, utility scores will also be taken into account in terms of loss optimization and model selection.