摘要：The progression of complex human diseases is associated with critical transitions across dynamical regimes. These transitions often spawn early-warning signals and provide insights into the underlying disease-driving mechanisms. In this paper, we propose a computational method based on surprise loss (SL) to discover data-driven indicators of such transitions in a multivariate time series dataset of septic shock and non-sepsis patient cohorts (MIMIC-III database). The core idea of SL is to train a mathematical model on time series in an unsupervised fashion and to quantify the deterioration of the model's forecast (out-of-sample) performance relative to its past (in-sample) performance. Considering the highest value of the moving average of SL as a critical transition, our retrospective analysis revealed that critical transitions occurred at a median of over 35 hours before the onset of septic shock, which suggests the applicability of our method as an early-warning indicator. Furthermore, we show that clinical variables at critical-transition regions are significantly different between septic shock and non-sepsis cohorts. Therefore, our paper contributes a critical-transition-based data-sampling strategy that can be utilized for further analysis, such as patient classification. Moreover, our method outperformed other indicators of critical transition in complex systems, such as temporal autocorrelation and variance.
