Detection of Breath Nitric Oxide at Ppb Level Based on Multiperiodic Spectral Reconstruction Neural Network

As breath nitric oxide (NO) is a biomarker of respiratory inflammation, reliable techniques for the online detection of ppb-level NO in exhaled breath are essential for the noninvasive diagnosis of respiratory inflammation. Here, we report a breath NO sensor based on the multiperiodic spectral reconstruction neural network. First, a spectral reconstruction method that transforms a spectrum from the wavelength domain to the intensity domain is proposed to remove noise and interference signals from the spectrum. Different from the traditional spectral processing method based on the wavelength domain, the method enhances the absorption characteristics of a target gas in the intensity domain, while discretizing noise and interference signals. This facilitates the extraction of the target gas spectrum. Then, a neural network is built to detect the concentration of breath NO. Laboratory-based results show that the sensor enables online detection of NO (1.63–846.68 ppb) with mean absolute error (MAE), mean absolute percentage error (MAPE), and detection accuracy of 0.31 ppb, 0.96% and 0.63%, respectively. Furthermore, an actual exhalation experiment proved that the sensor is capable of distinguishing breath NO of healthy people from that of simulated patients, which provides a reliable way to realize exhaled breath detection based on optical methods in the medical field.