Machine learning of metabolic fingerprints
for AECOPD diagnosis
and biomarker discovery. (a) Workflow for the diagnosis of AECOPD
by machine learning. The discovery cohort included 205 samples (103/102,
SCOPD/AECOPD) for parameter tuning and model construction. The optimized
model was tested in an independent validation cohort with 41 individuals
(19/22, COPD/AECOPD). No statistically significant differences in
age and gender between SCOPD and AECOPD in the discovery cohort (p > 0.05). (b) ROC curves differentiate SCOPD from AECOPD
for the discovery (blue) and validation (red) cohorts. (c) Venn diagram
of 8 m/z features screened as the
metabolic signature panel with frequency ≥90%, p < 0.05, abundance >500, and AUC of single feature >0.7.
Scatter
diagram of three key differential features for SCOPD and AECOPD, including
(d) lactic acid (Laa), (e) uric acid (Ura), and (f) malondialdehyde
(Mal). **** is represented by p < 0.0001. (g)
Fold change of four up-regulated metabolites (Laa, Ura, Mal, and 3-hydroxybutyric
acid (Hya) with magenta color) and four down-regulated metabolites
(creatine (Cre), dimethylglycine (Dim), threonine (Thr), and fucose
(Fuc) with cyan color) in AECOPD patients compared with SCOPD. (h)
The heat map of the discovery cohort, including SCOPD and AECOPD patients,
is constructed by eight metabolic biomarkers as potential signatures
for AECOPD diagnosis. (i) PCA analysis showed a clear discrimination
between SCOPD and AECOD patients based on eight metabolic biomarkers.