Large-Scale Methodological Comparison of Acute Hypotensive Episode Forecasting Using MIMIC2 Physiological Waveforms

We compare the dynamic Bayesian network and k-nearest neighbor-based predictors for the occurrence of acute hypotensive episodes (AHE) with respect to various data conditions (size, class balance ratio) and problem definition settings (lag, lead time). From our dataset extracted from the large ICU physiological waveform repository of MIMIC2 database, we find that both models are effective for predicting AHE and their performances improve with increasing training dataset size. We also empirically demonstrate that the nearest neighbor method has a better performance for larger datasets in terms of both prediction result and computational time, but it severely degrades for class imbalanced data while the Bayesian network remains robust.