Machine Learning Approaches: From Theory to Application in Schizophrenia

Table 1

Performance evaluation.

Author	Sample size	MRI technique	SVM input	Number of ROIs	Best accuracy (%)
[29]	SCZ = 64; HC = 60	Structural T1-w	Features vector	1 (left Amy)	86.13
[30]	SCZ = 64; HC = 60	Structural T1-w	Pairwise dissimilarities	7l + 7r (Amy, DLPFC, EC, HG, Hipp, STG, and Tha)	79
[31]	SCZ = 30; HC = 30	Structural T1-w	Features vector	1 (left Tha)	83.33
[32]	SCZ = 59; HC = 55	Structural T1-w; DWI	Pairwise dissimilarities	7l + 7r (Amy, DLPFC, EC, HG, Hipp, STG, and Tha)	86.84
[33]	SCZ = 42; HC = 40	Structural T1-w	Features vector	3l + 3r (DLPFC, EC, and Tha)	90.24
[28]	SCZ = 54; HC = 54	Structural T1-w	Features vector	1 (DLPFC)	84.09
[34]	SCZ = 50; HC = 50	Structural T1-w	Features vector	4l + 4r (Amy, EC, STG, and Tha)	84

Each method compared subjects affected by schizophrenia (SCZ) with healthy controls (HC). All methods focused on specific regions of interest (amygdala (Amy), dorsolateral prefrontal cortex (DLPFC), entorhinal cortex (EC), Heschl’s gyrus (HG), hippocampus (Hipp), superior temporal gyrus (STG), and thalamus (Tha)). The last column shows the performance of each algorithm in terms of accuracy, which is the overall proportion of correct classification (i.e., the number of correctly classified subjects divided by the number of all subjects).