Patient prognosis from vital sign time series: Combining convolutional neural networks with a dynamical systems approach

摘要: In this work, we propose a stacked switching vector-autoregressive (SVAR)-CNN architecture to model the changing dynamics in physiological time series for patient prognosis. The SVAR-layer extracts dynamical features (or modes) from time-series, which are then fed into CNN-layer extract higher-level representative of transition patterns among modes. We evaluate our approach using 8-hours minute-by-minute mean arterial blood pressure (BP) over 450 patients MIMIC-II database. modeled time-series third-order SVAR process with 20 modes, resulting first-level size 20×480 per patient. A fully connected CNN is used learn hierarchical these inputs, and predict hospital mortality. combined CNN/SVAR BP achieved median interquartile-range AUC 0.74 [0.69, 0.75], significantly outperforming CNN-alone (0.54 [0.46, 0.59]), SVAR-alone logistic regression (0.69 [0.65, 0.72]). Our results indicate that including an layer improves ability CNNs classify nonlinear nonstationary time-series.