A slow feature based LSTM network for susceptibility assessment of acute mountain sickness with heterogeneous data

Objective: Acute mountain sickness (AMS) is a syndrome that occurs when an individual rapidly rises to a high altitude and fails to adapt to acute hypobaric hypoxia physiologically. The aim of this paper is to develop an intelligent approach for the individual susceptibility assessment of AMS based on dynamic heterogeneous data monitored by multiple wearable devices. Methods: In this paper, the adaptive domain of hypoxia tolerance (ADHT) is established based on k-means clustering and mutual information (MI). Furthermore, a slow feature based long short-term memory (LSTM) learner is proposed to evaluate an individual's ability to tolerate hypoxia, which is used as the susceptibility evaluation of AMS. Results: The proposed method's performance is evaluated by using the heterogeneous physiological data of 18 subjects, augmented to 396 samples. The maximum MI value (0.3946) between cluster results and the lake louise score is retained to establish ADHT. The classification accuracy of the slow feature based LSTM method reaches 85.71% and the area under the ROC curve reaches 0.925. Comparing with other benchmark and deep learning approaches, the proposed method perform best in term of accuracy, precision, specificity and Matthews correlation coefficient. Conclusion: The results show that the proposed method is feasible in classifying individual hypoxia tolerance and evaluating AMS susceptibility. The system takes full advantage of dynamic heterogeneous data during offline modeling, and only needs the IHT data fed back by wearable devices during online monitoring. Significance: The method improves the convenience of susceptibility assessment of AMS.