FIGURE 1 Feature extraction pipeline. The four-wavelength fNIRS data is collected from the subjects and several pre-processing methods are applied before extracting the features.
Total of 209 features were extracted from fNIRS and fNIRS derived chromophore relative concentration signals [20]. The differential path length factor (DPF) used in MBBL was 5.97 and the extinction coefficients were based on the work of Cope [26]. The summary of the features is shown in Table 1. Some of the features have been described in the latest fNIRS, EEG and fMRI NDD studies while some are common features, such as signal standard deviation and gender of the subject. The frequency bandwidths of respiratory (0.1 - 0.6 Hz), cardiac (0.6 - 5 Hz) and full (0.008 - 5 Hz) were utilized when extracting fNIRS features. In addition to noise attenuation, the bandwidth specific features are assumed to contain information related to different physiological events and could contain more predictive power in the brain aging assessment.
Spectral entropy, fractional amplitude of physiological fluctuations (fAPF) and relative power derived features have been used in fNIRS and EEG studies related to aging and AD research [18,27–29]. Furthermore, pulse shape derived from fNIRS has been used to distinguish AD subjects from healthy controls [11]. Thus, the features were selected as they are assumed to have potential predictive power in the age group classification task. In addition, the features of differential phase transfer entropy (dPhaseTE) [30,31], sample entropy (sampEn) [32,33], and correlation between HbT and H2O [34] were utilized. Unique feature of envelope-signal spectra correlation (ESSC) was also used. The feature is calculated by computing the correlation between the power spectral density (PSD) of the signal and its envelope. The feature describes the relationship between the spectral power of the original signal and its low frequency component.
TABLE 1 List of all features extracted from the fNIRS and derived relative chromophore concentration signals used in the study.