Respiratory Failure Prediction Based on Geometric Modeling of Pediatric CCCU Monitor Data

The goal of this project is to use data from a CCCU monitor in order to predict respiratory failure events in pediatric ICU patients, and to predict the probability of successful extubation of pediatric patients from mechanical ventilation.
The decision whether a patient is capable of autonomous respiration is made today primarily based on the doctor's intuition and experience, or based on measurements obtained by special equipment. The extubation, however, might fail after a few hours, resulting in reintubation of the patient.
The project's data was obtained by low frequency sampling of CCCU monitors, which monitor various sensors (e.g. pulse, blood pressure, etc.), of patients at the ICU at SickKids Hospital in Toronto, Canada.
The signals are processed using geometric modeling methods based on Riemannian distances and Diffusion maps. We assume that there exists a latent signal that reflects the patient's distress and is reflected in trends in the signals measured and in the connections between them. The methods used are described in this report. The method of "Connectivity" has produced the best results for predicting successful extubation from mechanical ventilation.

The method of Connectivity has produced the best results for predicting successful detachment from artificial respiration.

Respiratory Failure Prediction Based on Geometric Modeling of Pediatric CCCU Monitor Data
Respiratory Failure Prediction Based on Geometric Modeling of Pediatric CCCU Monitor Data
Collaboration:

Dr. Danny Eytan