Using Adaptive Principal Component Analysis (APCA) and Machine Learning to characterise paediatric asthma in data collected by Structured Light Plethysmography (SLP)

98 children from 1 month to 18 years old with a diagnosis of Wheezing, Asthma or Bronchiolitis were assessed using SLP. Age-matched normal data was provided from the PneumaCare Ltd. database. SLP is a non-invasive light-based method that captures tidal breathing movement by reconstructing the surface of the anterior chest and abdominal wall. We adaptively analyse a moving surface over time by decomposing the data via APCA into its most important modes. We extract indices from the modes and their behaviour, for example Thoraco-Abdominal Asynchrony Fig. 1 and amplitudes of successive modes relative to the primary mode, providing a measure of shape complexity Fig. 2. Additional indices, such as the presence of harmonics, the position in expiration of maximum flow, and the relative inspiratory-expiratory flow (IE50) are also measured. An ensemble of Bagged Decision Trees achieves over 90% classification accuracy on normal vs abnormal. Work is underway to attempt classification of specific diagnoses and extend the study to adult cohorts.