Wearable Sensors and VOC Monitoring: Data Fusion and Machine Learning for Non-Invasive Disease Detection.

Abstract

Emerging technologies that detect volatile organic compounds (VOCs) emanating from the body enable rapid and non-invasive detection of diseases. Respiratory illnesses, in particular, can be identified through measurement of VOCs in exhaled breath. However, the low specificity of some methods leaves room for improvement. By combining information from multiple data streams, we can potentially improve on the results obtained from using each data stream individually. Here we evaluate a novel data fusion approach combining a wearable vitals monitor and exhaled breath VOCs for the classification of asthma patients. The wearable vitals monitor or "e-glove" measures key information such as skin temperature, trans-epidermal water loss (TEWL), heart rate, and galvanic skin response (GSR), while the breath VOC data (measured by gold standard gas chromatography-mass spectrometry, GC-MS) provide insight into the biochemical and metabolic state of the body. Data fusion is a process used in many applications for combining information from disparate sources. We leverage data fusion in this case to achieve superior classification accuracy using the integrated data streams. The low-level data fusion achieved a higher ROC AUC score (0.88) in asthma classification than the breath VOC-only approach (0.72, p<0.001). These results are promising for the future data analysis workflows for non-invasive diagnostic and health monitoring.