High Confidence Clinical Posture Classification and Non-Clinical Posture Detection

摘要

Pressure injuries are a major driver of increased morbidity and healthcare costs globally. A growing problem with an increasing aging population. Prevention requires extensive surveillance and adherence to frequent pressure offloading of at-risk patients. In this research, we explore a range of techniques to provide high reliability results for posture detection irrespective of quality of images. The foundational neural networks of this research is a range of low resource neural network classifiers, up to three layers deep. We then combine the neural networks to improve performance through the use majority rules and max confidence rule. Based on the shortcomings of the previous techniques we explore a more robust technique for novelty detection of image outside the known clinical posture training set. Done using a neural network leveraging the soft max values of the posture classification neural networks to detect images as novel classes to enable accurate tracking of patient mobility.