Table 2.

Factors that influence the ML effect size and classification accuracy.

Source	Relevant factor	Acting on	Effect in study:
			Training + LOOCV	Train → test
Reference: homogeneous sample (“0”)	Separation strength, Δy0 Spread, σy0	–	dML0 = Δy0/σy0 acc0 = Φ(dML0/2)	–
Heterogeneity of the diseasea	f	Separation strength, Δy → dML	dML = √[(1 + f)/2] dML0	dML = fdML0
Heterogeneity: biological variation	σBIOL	broadening, σy → dML	dML = (σy0/σy)dML0	dMLT = (σy/σyT)dML
Measurement noise	σEXP	broadening, σy → dML	dML = (σy0/σy) dML0	dMLT = (σy/σyT)dML
Sampling effects (finite N)	N, σy	uncertainty in accuracy	≤ SD(acc) in train/test case	SD(acc) = √(acctrue × (100%−acctrue)/N)
Gold → silver standard	Intra-class kappa, κ	ceiling of accuracy		acc = κ × acc0 +(1−κ)/2

^ai.e., related to the prediction model.

f = cos(θ), the relative amount of heterogeneity;σ²_y = σ²_BIOL + σ²_EXP; acc = accuracy; subscripts “T” refer to values in the Test sample.