Table 3.
Frequently selected metabolites in feature selection models by fivefold cross-validation and their P values and regression coefficients in multiple regression analyses.
| CES-D score | Binary CES-D traits | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| HSIC Lasso + KR | Lasso | Multiple linear regression <0.05 | P valuea | RCa | HSIC Lasso + SVM | Lasso | Multiple logistic regression <0.05 | P valuea | RCa | |
| Number of selected metabolites for prediction (mean ± standard deviation) | 9.6 ± 2.8 | 13.4 ± 6.4 | 16.0 ± 4.0 | NA | NA | 9.4 ± 6.2 | 23.4 ± 6.7 | 26.4 ± 8.0 | NA | NA |
| Metabolites selected in both nonlinear and linear models | ||||||||||
| 3-Hydroxyisobutyrate | 5/5 | 5/5 | 5/5 | 6.67 × 10−4 | −0.832 | 5/5 | 5/5 | 5/5 | 1.78 × 10−4 | −0.320 |
| NMR | ||||||||||
| Gamma-linolenyl carnitine | 4/5 | 5/5 | 5/5 | 1.44 × 10−3 | −0.758 | 5/5 | 5/5 | 5/5 | 5.20 × 10−3 | −0.243 |
| MS C18 | ||||||||||
| l-leucine | 4/5 | 3/5 | 4/5 | 3.52 × 10−3 | −0.753 | 5/5 | 5/5 | 5/5 | 1.06 × 10−3 | −0.286 |
| MS C18 | ||||||||||
| Metabolites selected only in linear models | ||||||||||
| Uric acid | 0/5 | 5/5 | 5/5 | 3.03 × 10−4 | −1.024 | 0/5 | 5/5 | 4/5 | 9.61 × 10−3 | −0.239 |
| MS C18 | ||||||||||
| Metabolites selected only by P values from multiple regression | ||||||||||
| l-gamma-glutamyl-l-leucineb | 1/5 | 1/5 | 5/5 | 9.19 × 10−3 | −0.679 | 0/5 | 0/5 | 5/5 | 1.82 × 10−2 | −0.206 |
| MS C18 | ||||||||||
Abbreviations: CES-D Center for Epidemiologic Studies-Depression Scale, HSIC Lasso Hilbert–Schmidt independence criterion lasso, KR kernel regression, RC regression coefficients, SVM support vector machine, NMR nuclear magnetic resonance spectroscopy, MS C18 mass spectrometry in C18 mode.
The frequencies are shown as n/5, which means that the metabolites were utilized for prediction n times out of five replicated feature selections.
aP value and regression coefficients are adjusted by sex, age, body mass index, marital status, the degree of damage from the Great East Japan Earthquake, antidepressant use, Lubben Social Network Scale 6, and social capital scale in multiple linear/logistic regression.
bIn the jMorp metabolomic database, l-gamma-glutamyl-l-leucine and l-gamma-glutamyl-l-isoleucine were not differentiated for one of the features selected by P values from multiple regression, and standard reagents, i.e., H-Glu(Leu-OH)-OH (BACHEM, Budendorf, Switzerland) and l-gamma-glutamyl-l-isoleucine (Santa Cruz Biotechnology, Heidelberg, Germany), were utilized to determine l-gamma-glutamyl-l-leucine as the detected metabolite for the feature in the current study.