P125 Using adaptive principal component analysis and age-varying kernel distributions to characterise COPD in data collected by structured light plethysmography (SLP)

Structured Light Plethysmography (SLP) is a non-invasive, light-based method that enables reconstruction of a patient’s anterior chest and abdominal wall. Samples are taken at 30Hz for several minutes for each patient. 30 datasets obtained from patients diagnosed with COPD (labelled abnormal) and 33 without (labelled normal) are available for characterisation. We demonstrate how a classifier may be trained to characterise COPD based on existing samples. Method SLP data was collected using the Thora-3Di by PneumaCare Ltd, made available from the Pneumacare database. Time-varying surfaces are decomposed into their constituent modes via adaptive principal component analysis. This method extracts a mean surface shape and motion modes. We extract measurement indices from the decomposition to classify between normal and abnormal. Indices found to be useful include Peak Expiration Width (PEW), (the fraction of the expiration time that is spent at greater than 60% of the maximum expiration rate), Component Ratio 1 (CR1), (the amplitude of the second motion mode relative to the first, indicating complexity of the breathing pattern), and Displacement at Maximum Flow (DMF), (the fraction of expiration that has occurred at the instant of peak expiration rate). Two-Dimensional Gaussian Kernel distributions are constructed using a training set of normal and abnormal patient samples for each measurement index, with age as the second dimension. While typical Gaussian Kernel distributions would centre a distribution component on each patient, we place distributions over each pair of patients with the same classification, which corrects for non-uniformity of the distribution of patient ages. Results Distributions for PEW and CR1 are shown in the figure. Light regions indicate high COPD likelihood. CR1 is typically lower for normal patients; PEW has a normal region, above and below which indicates a higher likelihood of COPD. Best performance is achieved with a voting scheme, where each measurement index distribution votes once. Classification accuracy is 86% using 5-fold cross-validation. Conclusion We have presented a non-invasive characterisation method for COPD. The method may be performed on captured SLP data to provide additional decision making information for clinicians.