Dynamical monitoring and causal inference for respiratory metabolic indicators of septic patients

Background: The objective of this study is to develop monitoring and predictive models for respiratory dynamics of sepsis to improve the emergency medical care of septic shock patients. Methods: We develop two patient-specific models to recognize and forecast the respiratory dynamics of a septic shock patient using the patient’s longitudinal data of three respiratory metabolic indicators PO2, PCO2, and SpO2, obtained from the arterial blood gas analysis over 8 days. The first is based on the neural dynamical system architecture while the second is on the long and short-term memory (LSTM) recurrent neural network (RNN) architecture. The causal relations among the indicators are inferred via information flow theory from the dynamical system models. Results: The models recognize the respiratory dynamics of the septic patient very well and can make short-term predictions with clinically acceptable relative errors of less than 5.2% in the L1 and L2 norm and less than 8.2% in the L norm, attesting to the effectiveness of the models. The subsequent causal analysis shows that SpO2 or PO2 is, respectively, the cause of PCO2, while there exist mutually causal relationships between SpO2 and PO2, consistent with the clinical experience. Conclusions: These models provide useful quantitative tools for physicians to make critical diagnostic and treatment decisions for septic shock patients in emergency situations.