Identifying Causal Effects from Data for the Clinical Ventilation Process Modelling

Proper modeling of the ventilation process is crucial to the effective operation of computerized ventilator management systems. We aim to develop a ventilation modeling technique, which depends less on lung dynamics assumptions, is able to describe the ventilation process quantitatively, and includes only clinically available parameters. We propose a Granger-causality based technique to identify causal relationships among ventilation variables, as the structural constraints typically provided by the subjective theory, controlled experiments or directed acyclic graphs (DAGs) are not available. We examine the performance of the proposed modeling methodology from different perspectives with real data. Domain knowledge confirmed and experiments show that the model outperforms the Vector Autoregression (VAR) and Neural Network methods. The proposed method provides initial insights into the data based ventilation process modeling.