Table 7.

Ablation studies on models pre-trained representations and downstream tasks with frozen encoder.

Methods	Age prediction			NC vs. AD			sMCI vs. pMCI			no-to-low vs. moderate-to-heavy drinkers
	RMSE	SC	RVRC	BACC	SC	RVRC	BACC	SC	RVRC	BACC	SC	RVRC
LNE* (Ouyang et al., 2021)	10.3	0.72†	88.9†	81.9†	0.76†	98.4†	70.6†	0.78†	101.2†	62.8†	0.70†	86.5†
LNE*+Proto	10.1	0.74†	91.2†	82.0	0.79†	102.3†	70.8†	0.80†	103.7†	63.2†	0.72†	89.3†
LNE*+SS	10.0	0.79	99.6	82.2	0.83†	104.1†	70.9	0.85†	106.8†	63.4†	0.78†	100.5†
LNE	10.0	0.81	103.2	82.1	0.87	106.5	71.1	0.88	108.2	63.7	0.81	102.3

Silhouette Coefficient (SC) and revised Variation Ratio Criterion (RVRC) were computed on the pre-trained representations to evaluate the quality of the neighborhood. +Proto means adding the age-consistent neighborhood regularization to LNE*, and +SS denotes the mini-batch sampling strategy.

^†Significant lower scores compared to LNE, which includes both Proto and SS.