Online Supplement A

ONLINE SUPPLEMENT A

1. The Mel-scale spectrum is basically the Discrete Cosine Transform (DCT) of the logarithmic energy spectrum of a signal, where the spectral energies are calculated using logarithmically spaced filters with increasing bandwidth (mel-filters). In more details, the mel- frequency ceptral coefficients, MFCCs, were derived as follows: First, the spectral representation of each 2-sec segment was passed through a bank of band-pass filters. The filters had a triangular-shape frequency response. Their central frequencies (CF) were linearly spaced in the frequency axis for low frequencies but logarithmically spaced for high frequencies, forming the so-called mel-scale filter bank. The logarithm of the resulting mel-spectrum is calculated and transformed via the DCT, yielding an MFCC sequence for each corresponding CF 1.
2. The adjusted R-square (coefficient of determination) is defined as, where n is the number of responses and d the degrees of the polynomial. measures the total sum of squares: , with the mean of the observed data. measures the residual sum of squares:. Note that SSTotal and SSResid are indicators of the sample variances of the dependent variable and the estimated residuals respectively. Adjusted R-square is a refinement of the R-square statistic accounting for the degrees of freedom, and thus making model fits comparable.
3. Point-wise (or non-simultaneous) prediction boundsmeasure the confidence that a new observation will lie within the bound interval, given the predictor values. These bounds are given by where s2 is the mean squared error, t is the inverse of Student's-t cumulative distribution with respect to the corresponding confidence level, and S is the covariance matrix,(, of the estimates .

REFERENCES

1. Davis S, Mermelstein P. Comparison of parametric representations for monosyllabic word recognition in continuously spoken sentences [Internet]. IEEE; 1980. Available from:

ONLINE SUPPLEMENTB

E-Table 1.Linear Regression Line fits of extracted features (rows) with respect to individual subject‘s characteristic Age, Height and Weight (columns). The adjusted R-square is also shown.

E-Table 2.Linear Regression Line fits of extracted features (rows) with respect to the combined subject‘s characteristic Age, Height and Weight (columns). The adjusted R−square is also shown.

ONLINE SUPPLEMENTC

E-Figure 1: Linear fit (solid line) for each feature (rows, y axis) with respect to patient characteristics (columns, × axis). Point-wise prediction bounds with 95% confidence level are also shown with dashed lines. Inset: R2a, the adjusted coefficient of determination of the quadratic fit; r, the linear correlation coefficient only if significant correlation was achieved.

E-Table 3Linear Regression Line fits of extracted features (rows) with respect to individual subject’s z-scores Weight-for-Height WHZ, Weight-for-Age WAZ, Height-for-Age HAZ and Body mass index-for-Age BAZ. The adjusted R-square is also shown.

E-Table 4.Linear Regression Line fits of extracted features (rows) with respect to the combined subject’s z-scores Weight-for-Height (WHZ), Weight-for-Age (WAZ), Height-for-Age (HAZ) and Body mass index-for-Age (BAZ). The adjusted R-square is also shown.