Demographic and follow-up information of the participants included in this study are summarized in Table 1 and Additional file 1: Table S1. Longitudinal samples from the BABRI included 65 sNC to build structural covariance networks and 28 sNC, 18 pNC, 20 non-rMCI (containing only the MCI who maintained MCI), and 24 rMCI to conduct predictive analyses. The NC and MCI participants were followed up with a mean duration of 37 and 27 months, respectively.

Another longitudinal sample from the ADNI included 47 sNC to build structural covariance networks and 47 sNC, 54 pNC, 65 non-rMCI, and 73 rMCI to conduct predictive analyses. The NC and MCI participants were followed up with a mean duration of 44 and 49 months, respectively.

Overall, there were no significant differences in age, sex, or education levels between participants with sNC and pNC in both samples, while non-rMCI was relatively older (p = 0.030 in BABRI and p = 0.099 in ADNI) and more homozygous carriers of ε4 Allele of Apolipoprotein E (APOE4) (p = 0.001 in ADNI) than rMCI. For general cognitive function as measured by MMSE at baseline, while sNC participants from BABRI presented better cognition than pNC (p < 0.001), the pNC showed quite similar cognition with sNC in ADNI (p = 0.213). rMCI in both samples were cognitively better than non-rMCI (p = 0.066 in the BABRI and p < 0.001 in the ADNI). For vascular risk factors (i.e. hypertension, diabetes, hyperlipidemia, smoking history, and BMI-measured obesity), rMCI was more likely to occur in non-smoking individuals in the ADNI sample (p =0.009, Additional file 1: Table S2), and no difference in vascular risk factors was found between pNC and sNC.

Seed-PLS analyses were performed on the independent sNC data within the BABRI and ADNI (Fig. 2, Additional file 1: Tables S3-S4). In both the BABRI and ADNI samples, the GM density of the DMN seed regions was mainly covaried with the extended posterior cingulate cortex, superior and middle temporal lobe, middle frontal lobe, and insula. The GM density of the FPN seed regions was mainly covaried with the middle frontal lobe, middle occipital lobe, middle cingulum gyrus, middle temporal lobe, and cuneus. The GM density of the HN seed regions covaried with the extended hippocampal lobe, middle and superior frontal lobe, middle temporal lobe, and other regions.

Structural covariance network based on independent cognitively normal elderly from the BABRI. BSR, bootstrap ratio, representing the covariance degree with the seed regions

In the BABRI sample, compared with sNC, pNC had a lower composite score of structural covariance of the FPN (scFN, p = 0.030, Table 2) and the HN (scHN, p = 0.035), and compared with non-rMCI, rMCI had a higher composite score of scFN (p = 0.018) and scHN (p = 0.007). The composite score of the structural covariance of the DMN (scDN) also showed a higher trend in rMCI, but it was not significant (p = 0.061).

In the ADNI sample, the baseline scores of scDN, scFN, and scHN of sNC and rMCI were higher than those of pNC and non-rMCI (all p ≤ 0.001, Table 2). Of note, 34 of 65 (63%) non-rMCI were progressed to AD (pMCI) during the follow-up, and pMCI had lower structural covariation scores than rMCI (Fig. S1, p < 0.001) and non-rMCI who maintained MCI (sMCI, p < 0.05), adjusted for age, sex, education level, and TIV.

We also examined whether the differences in structural covariation scores were associated with APOE4 carrier status, the strongest known genetic risk factor for late-onset AD cases [42]. We found that APOE4 homozygotes had lower structural covariation scores than non-carriers (Fig. S2, scDN, p = 0.017; scFN, p = 0.056; scHN, p = 0.009) and APOE4 heterozygotes (scDN, p = 0.03; scFN, p = 0.063; scHN, p = 0.025). In addition, the effect of disease duration on MCI reversion was explored in the supplementary analysis (Fig. S3). The longer the disease course of rMCI, the lower the composite score of the structural covariant network (scDN, r = − 0.555, p = 0.049; scFN, r = − 0.513, p = 0.073; scHN, r = − 0.475, p = 0.101).

The above analyses indicated the possibility to use GM covariance patterns to predict MCI development and reversion several years later at the group level. To investigate whether GM covariance patterns could be useful for predicting MCI development and reversion at the individual level, random forest models were used to construct predictive models for changes in future clinical status.

For N-t-M progression, all structural covariance networks were able to classify cognitively normal elderly people into sNC and pNC based on the BABRI sample (AUC = 0.692–0.792, Table 3). It is worth noting that the hippocampal covariance network achieved the best performance (AUC = 0.792). In addition, based on the ADNI sample, the baseline scores of scDN (AUC = 0.766), scFN (AUC=0.765), and scHN (AUC = 0.785) could also accurately distinguish pNC from sNC (Table 3).

For M-t-N reversion, the baseline scores of scDN, scFN, and scHN also showed good predictive performance based on the BABRI sample (Table 3, AUC = 0.722–0.745). In the ADNI sample, the baseline scores of scHN achieved the best prediction effect (AUC = 0.809), and the AUCs of the other two prediction models were above 0.701 (Table 3).

To further identify the brain regions that play a key role in predicting N-t-M progression and M-t-N reversion, feature weight distributions were depicted in Fig. 3A for the BABRI sample and Fig. S4A for the ADNI sample. We found that for both samples, the superior temporal gyrus of the DMN, and the middle frontal gyrus of the FPN and HN, played a key role in the N-t-M prediction. Notably, the hippocampus and parahippocampal regions played a key role in the M-t-N prediction of all three networks (Fig. 3B, Fig. S4B).

Feature weight distribution of the random forest prediction model of the BABRI sample. The contribution of voxels from the default network, frontoparietal network, and hippocampal network for prediction of the progression of normal cognition (A) and the reversion of mild cognitive impairment (B). The bar diagram on the right is the cluster with the top 10 feature weights in each network, and the horizontal axis is the weight value. Abbreviations: L, left; R, right; Mid, middle; Sup, superior; Ant, anterior; Inf, inferior; Orb, orbital

Supplementary analyses were also conducted to clarify the potential impacts of field strength and scanner sites for the ADNI sample, and it turned out that, by including the two factors as covariate variables in the predictive models or only using data with single field strength, all three networks maintained good performance for predictive N-to-M progression and M-to-N reversion (Additional file 1: Tables S6-S9). The prediction of MCI progression and reversion in ADNI samples was also conducted (Additional file 1: Table S5). For MCI progression and reversion, both the structural covariant network and the hippocampus and parahippocampal region showed excellent predictive performance (all AUC > 0.8), especially for scDN (AUC = 0.874) and scHN (AUC = 0.868). More details could be found in the supplementary materials.

To explore which of the structural covariance networks and the key brain regions were more sensitive to the development and reversion of MCI, we evaluated the predictive accuracy of key brain regions, that is, regions that emerged as having key roles in the structural covariance networks and hippocampal region (Table 4).

For N-to-M progression, the predictive accuracy of the middle frontal gyrus, superior temporal gyrus, and hippocampus (AUC = 0.607–0.717) was lower than that of the hippocampus covariance network (AUC = 0.792) based on the BABRI sample, while the predictive accuracy of the hippocampus (AUC = 0.803) was similar to that of the covariance networks based on the ADNI sample.

For M-t-N reversion, the predictive accuracy of individual hippocampal regions (AUC = 0.759) was better than that of the covariance networks based on the BABRI sample, while the predictive accuracy of the hippocampal region (AUC = 0.780) and parahippocampal region (AUC = 0.769) was worse than that of the hippocampus covariance network (AUC = 0.809) based on the ADNI sample.

In addition, we evaluated the GM covariance relationship between these key brain regions. By regressing out the effects of age, sex, education, and TIV, we found that for the N-to-M progression, sNC from BABRI presented a marginally significant positive correlation between the GM density of the superior temporal gyrus and that of the hippocampus (r = 0.400, p = 0.053), while this correlation disappeared in pNC (r = 0.218, p = 0.455) in the BABRI sample. Similar results were also identified in the ADNI sample, with sNC (r = 0.503, p = 0.001), not pNC (r = 0.225, p = 0.117), having significant GM covariation between the middle frontal gyrus and hippocampus.

And for M-t-N reversion, the GM density of the hippocampal and parahippocampal regions was correlated in both sNC and pNC (p < 0.01) in BABRI and ADNI samples.