Irregular breathing detection in CPAP assisted patients using hierarchical temporal memory

Continuous Positive Airway Pressure (CPAP) is a treatment that applies mild air pressure to retain the airways open. CPAP typically is applied to patients with breathing problems, such as sleep apnea. CPAP also is widely used to treat preterm infants whose lungs have not fully developed. Anomaly detection, which is the practice of finding patterns in data that do not conform to expected behavior, is critical for regulating CPAP and assisting physicians in monitoring or diagnosis. Motivated in applying machine learning to CPAP data, we used in this experiment the Numenta Anomaly Benchmark (NAB), which is an open source repository created by Numenta that provides anomaly detectors using an Hierarchical Temporal Memory (HTM) based detector. HTM is a cortically inspired learning algorithm and is well adapted to anomaly detection. The NAB was used to test multiple detectors on breathing patterns from an infant respiratory complexity clinical study gathered at the American University of Beirut Medical Center (AUBMD). We also designed, implemented and tested a heuristically driven batch detector for comparison purposes and to complement HTM and the set of detectors provided by NAB. Testing NAB, HTM and the custom proposed detector reveals good anomaly prediction results and motivates follow on research.