Introduction: Respiratory Distress (RD) is typically associated with many critical conditions, particularly in the aged population. A coarse characterization of RD is laboured breathing or signs that the patient is not getting enough oxygen. Still, it is difficult to gauge the acuteness of RD by short term trend or value of a single parameter. The present-day monitoring solutions do not track the severity of RD distinctly and this obscures accurate prognosis.
Aim: The goal of work is to quantify RD condition to a severity index on a scale of 0 to 1. The method should use three non-invasively acquired vitals as inputs viz. respiratory rate (RR), peripheral capillary oxygen saturation (SpO2) and heart rate (HR).
Methods: The proposed method tracks the pattern of RR, SpO2 and HR over time using Convolutional Neural Network (CNN) based model for multivariate time series segmentation. The segments extracted from 180 records of MIMIC-III Clinical and Waveform Database were clinically annotated for the study. The 2812 annotated segments were split into training, validation and test sets for applying the CNN technique. This method was then compared with other two different approaches viz. Long-Short Term Memory (LSTM) model and Bayesian Inference model based on Symbolic Aggregate approXimation (SAX).
Result: The CNN, LSTM and SAX models achieved root mean square error of 0.288, 0.301 and 0.305 respectively with respect to the clinical annotations on the test set. All methods achieved an Area Under the Receiver Operating Characteristics (AUROC) close to 0.95. Though the performance of all the models was almost at par, CNN model performed marginally better.
Conclusion: The CNN approach of quantifying severity RD can be integrated into a real time streaming setup and can be effectively used for continuous tracking of intensity of respiratory abnormality.