We collected the data from two independent cohorts. Data used for the preparation of the current study were obtained from the ADNIMERGE dataset of the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu) (n = 407) and from the Samsung Medical Center (SMC) (n = 285).

The ADNI was launched in 2003 as a public-private partnership, led by principal investigator Michael W. Weiner, MD. The primary purpose of the ADNI is to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical or neuropsychological assessments, can be combined to evaluate the progression of MCI and early AD. For up-to-date information, see www.adni-info.org. Participants from the ADNI-1 to ADNI-3 and ADNI GO were included in the current study if (1) their baseline diagnosis was CU or subjective memory complaints, (2) they had valid cognitive assessments, and (3) they had at least two follow-up assessments of cognitive function. The baseline diagnoses were determined using the standard criteria described in the ADNI procedure manuals.¹ As a result, a total of 407 individuals met these qualifications to be included in the current study. All individuals had usable neuropsychological data, and 352 individuals had hippocampal volumetric data.

In addition, 285 CU individuals were recruited from the SMC. These patients also had usable neuropsychological data, and 231 individuals had hippocampal volumetric data. They underwent amyloid PET scans at the SMC between September 2015 and December 2021, and were followed up at least twice through clinical interviews and thorough cognitive tests until December 2021. The following were used to establish the baseline diagnostic criteria for CU: (i) the Korean Mini-Mental State Examination (K-MMSE) ≥24 or above −1.5 SD from the age-, gender-, and education-adjusted norms if education level was <9 years; (ii) above −1 SD from the age-, gender-, and education-adjusted norms on the delayed recall of the Seoul Verbal Learning Test-Individuals version (SVLT-E); (iii) above −2 SD from the age-, gender-, and education-adjusted norms on the Korean version of the Boston Naming Test (K-BNT), the Rey-Osterrieth Complex Figure Test (RCFT) copy, and the Korean Color Word Stroop Test (K-CWST) color reading; and (iv) absence of other neurological diseases.

The authors obtained approval from the ADNI Data Sharing and Publications Committee for data use and publication. No Institutional Review Board (IRB) review approval was required to use de-identified ADNI data that is available for download. All methods were carried out in accordance with the approved guidelines. At the SMC, the IRB approved the use of SMC data for this study, and all of the methods used were carried out in compliance with the approved standards.

The current study employed the following neuroimaging data from the ADNIMERGE dataset: average AV45 standardized uptake value ratios (SUVRs) of the frontal cortex, parietal cortex, anterior cingulate cortex and precuneus relative to the cerebellum, and the hippocampal volume (HV). The detailed protocols for image acquisition have been described in previous studies (Jagust et al., 2010) and the ADNI database.² Cortical reconstruction and volumetric segmentation were performed using the FreeSurfer image analysis program (Hartig et al., 2014).³

For SMC data, all individuals underwent either FBB or FMM PET scans at the SMC using a Discovery STe PET/CT scanner (GE Medical Systems, Milwaukee, WI, USA) in 3D scanning mode to examine 47 slices of 3.3-mm thickness spanning the entire brain (Kim et al., 2018; Jang et al., 2019). In our previous study, we used a direct comparison Centiloid units (dcCL) conversion equation to directly convert the SUVR values of the FBB or FMM cortical target volume of interest (CTX VOI) into dcCL units (Klunk et al., 2015; Cho S. H. et al., 2020). The conversion was performed using equations for the FBB (dcCL_FBB = 151.42 × dcSUVR_FBB − 142.24); and the FMM (dcCL_FMM = 148.52 × dcSUVR_FMM − 137.09) PET. All individuals also underwent 3D T1-weighted turbo field echo MRI at SMC using a 3.0-T MRI scanner (Philips 3.0T Achieva; Philips Healthcare, Andover, MA, USA) as previously described (Kim H. J. et al., 2016). Images were processed using the CIVET anatomical image-processing pipeline (version 2.1.0) (Zijdenbos et al., 2002). We calculated the intracranial volume (ICV) by measuring the total volume of the voxels within the brain mask (Smith, 2002). The FMRIB (Functional Magnetic Resonance Imaging of the Brain) Software Library (FSL) method was used to create brain masks. Since cortical surface models were extracted from brain MRI volumes transformed into stereotaxic space, cortical thickness was assessed in the native space by applying an inverse transformation matrix to the cortical surfaces and rebuilding them in native space (Im et al., 2006). To measure HV, we employed a computerized hippocampus segmentation technique that combined graph-cut optimization, atlas-based segmentation, and morphological opening (Kwak et al., 2013).

The current study used the ADNI-modified Preclinical Alzheimer Cognitive Composite with the Trail-Making Test, Part B time to completion (mPACCtrt) as the cognitive endpoint from the original ADNIMERGE dataset. The PACC was developed as an outcome measure of cognitive changes in preclinical AD (Donohue et al., 2014). The original version includes (a) the total recall score of the Free and Cued Selective Reminding Test (FCSRT), (b) the delayed recall score of the Logical Memory IIa (LM), (c) the Digit Symbol Substitution score from the Wechsler Adult Intelligence Scale revised version (DSST), (d) Mini-Mental State Examination (MMSE) total score. Because the FCSRT was not included in the ADNI battery, the FCSRT was replaced by the delayed recall of the Alzheimer’s Disease Assessment Scale (ADAS) in the mPACCtrt. Also, we used the Trail-Making Test, Part B (TMT-B) instead of the DSST (Donohue et al., 2014). As a result, the mPACCtrt consisted of (a) ADAS-cognitive subscale delayed word recall, (b) logical memory delayed recall, (c) the MMSE total score, and (d) (log-transformed) trail-making test time to completion. The composite score was the sum of the z-scores of each constituent test, which were based upon the mean and standard deviations of the baseline scores of the CU individuals in the ADNI (Donohue et al., 2014, 2017).

For the SMC dataset, the Longitudinal Amyloid Cognitive Composite in Preclinical AD (LACPA) was used. The LACPA was developed for longitudinal tracking of amyloid-β (Aβ)-related cognitive decline in CU individuals using well-characterized and relatively large Korean CU cohorts (Kim Y. J. et al., 2022). The LACPA equation is as follows:

where IR is the immediate recall, DR is the delayed recall, K-TMT-B is the Korean Trail-Making Test Part B and Z is the z-score.

We used mixture modeling to test for distinct growth patterns in the cognitive trajectories of the CU individuals. Mixture modeling generally uses categorical latent variables representing the composition of a subpopulation, in which case the members of the subpopulation are unknown and inferred from the data (Wang and Bodner, 2007). In mixture modeling using longitudinal data, unperceived heterogeneity is captured through categorical and continuous latent variables (Wang and Bodner, 2007). We used mixture modeling with GMM and the LCGA for the ADNI dataset and SMC datasets, respectively. The LCGA is a special type of GMM that limits the variance and covariance estimates for growth factors within each class to zero (Jung and Wickrama, 2008). To determine the adequate number of latent classes, we compared the following four methods: Bayesian information criterion (BIC) (Schwarz, 1978); Akaike information criterion (AIC) (Akaike, 1974); Lo, Mendell, Rubin test (LMR) (Lo et al., 2001); and parametric bootstrapped likelihood ratio test (BLRT) (McLachlan and Peel, 2000). After determining the number of latent classes, the baseline characteristics of the latent classes were analyzed using the independent t-test and the Chi-square test.

To investigate the effects of variables on the cognitive trajectory group based on the mPACCtrt score derived from the ADNI dataset, we performed multivariable logistic regression analyses, which include AV45 SUVR, baseline mPACCtrt score, and HV associated with covariates including age, gender, education level, APOEε4. In addition, to determine the effects of variables on the cognitive trajectory group based on the LACPA score derived from the SMC dataset, we conducted multivariable logistic regression analyses, which include dcCL, baseline LACPA score, and HV associated with covariates including age, gender, education level, and APOE ε4 (Figure 1). The HV was used as the value divided by the ICV (HV/ICV) to consider the cranial cavity. The corrected HV was multiplied by 1,000, because of the relatively small scale. The dcCL was used as the value divided by 100.

After developing the predictive models, we constructed a nomogram using the dataset. We assigned a point value to each variable using the beta coefficients from the logistic regression model. Then, the most powerful variable was set at 100 points, while other variables were assigned between 0 and 100 points based on their proportions. The total points summed by the scores corresponding to each variable were immediately converted to risk probability. Finally, the predictive accuracy of the nomogram was verified by discrimination (C-index) and calibration.

Mplus version 8.3 was used for the GMM, and the LCGA (Muthén and Muthén, 2019). Full information maximum likelihood (FIML) estimation was employed for the missing values found in the longitudinal data. R 4.0.3 (Vienna, Austria⁴) was used for the logistic regression analyses and to develop the nomogram.

The datasets used and/or analyzed during the present study are available from the corresponding authors on reasonable request.

The baseline demographic characteristics of the CU individuals included in the ADNI and SMC datasets are shown in Table 1. The mean age of the individuals was 73.2 years in the ADNI dataset and 70.7 years in the SMC dataset (p < 0.001). The proportion of APOE ε4 carriers was 30.2 and 29.8% in the ADNI dataset and the SMC dataset, respectively. The proportion of females was 53.8% in the ADNI dataset and 61.8% in the SMC dataset (p = 0.038).

For the ADNI dataset, to explore the number of latent class, BIC, adjusted BIC and AIC values were compared as we increased the number of classes to determine the appropriate model. Then, we compared the models using the LMR and BLRT (Table 2). Although the three-class model exhibited smaller AIC and adjusted BIC values, the two-class model had the smallest BIC value. Moreover, the difference between these values when obtained from the two- and three-class models was minimal compared to the difference for the one- and two-class models. This finding indicated that the two-class model was an appropriate model. To determine whether the two- or three-class model was better, we also used the LMR and BLRT. The LMR comparing the two- and three-class models indicated the two-class model was better (p = 0.014), and the BLRT result also supported the two-class model (p < 0.001). Considering these results comprehensively, we decided that the two-class model would be adequate for the final model. Its entropy (the quality of the classification) was high at 0.79. For the SMC dataset, the LMR comparing the one-, two-, and three-class models indicated that the two-class model was better (p = 0.018), and the BLRT result also supported the two-class model (p < 0.001). Considering these results comprehensively, we again decided that the two-class model would be adequate for the final model. Its entropy, the quality of the classification, was high at 0.861.

Figure 2 shows the cognitive trajectories of the two-class model using the PACC and LACPA for the ADNI and SMC datasets, respectively. In the ADNI dataset, Class 1, in which the estimated class percentage was 86.2%, was a group whose cognitive function remained stable for 7 years (intercept = 0.479, SE = 0.148, p = 0.001; slope = −0.012, SE = 0.027, p = 0.649). Class 2, in which the estimated class percentage was 13.8%, showed cognitive decline during the same period (intercept = −1.058, SE = 0.534, p = 0.047; slope = −1.262, SE = 0.088, p < 0.001). Thus, we identified Class 1 a “stable group” and Class 2 as a “declining group” (Figure 2A). The same trajectory subgroups were also identified in CU individuals of the SMC dataset. Class 1, in which the estimated class percentage was 87.0%, was a group whose cognitive function remained stable for 7 years (intercept = 0.257, SE = 0.045, p < 0.001; slope = −0.060, SE = 0.023, p = 0.008). Class 2, in which the estimated class percentage was 13.0%, showed cognitive decline during the same period (intercept = −0.734, SE = 0.180, p < 0.001; slope = −0.424, SE = 0.075, p < 0.001) (Figure 2B).

The baseline characteristics of the two trajectory groups in the ADNI dataset (351 stable and 56 decliners), and the SMC dataset (248 stable and 37 decliners) are shown in Table 3. The data analyses indicated there were significant predictive factors among the demographic variables and biomarkers at baseline identifying individuals who were at risk of cognitive decline among the two cohorts. The mean age was higher in the decliners than in the stable groups of both datasets (t-test; ADNI p < 0.001; SMC p = 0.032). The years of education and gender distribution did not significantly differ between the two groups. The presence of APOEε4 was more frequent among decliners than in the stable group (Chi-square test; ADNI p = 0.028; SMC p = 0.021). In a comparison of biomarkers including Aβ and HV, Aβ uptake was higher in the decliners than in the stable group (t-test; ADNI p < 0.001; SMC p < 0.001). In terms of the HV, it was lower in the decliners than in the stable group in ADNI, whereas there was no significant difference between the two groups in the SMC dataset (t-test; ADNI p < 0.001; SMC p = 0.063). The baseline neuropsychological composite score was lower in the decliners than in the stable group in both datasets (t-test; ADNI p < 0.001; SMC p < 0.001).

Multivariable logistic regression was used to analyze the associations between cognitive decline and potential predictors. Table 4 shows the significant measures predicting cognitive decline. In the model using the ADNI dataset, the multivariable logistic regression model showed that Aβ uptake, baseline neuropsychological composite score and HV were predictive of being included in the cognitive decline group. Increased Aβ uptake (β [SE]: 4.852 [0.862], p < 0.001), lower baseline neuropsychological composite score (β [SE]: −0.274 [0.070], p < 0.001), and lower HV (β [SE]: −0.952 [0.302], p = 0.002) were significantly correlated with greater likelihood of being included in the cognitive decline group. The predictive performance (C-index) in the ADNI model was 0.85 (Table 4). In the model using the SMC dataset increased Aβ uptake (β [SE]: 2.007 [0.549], p = 0.007) and lower baseline neuropsychological composite scores (β [SE]: −4.464 [0.758] p < 0.001) were significantly correlated with greater risk probability. The C-index in the SMC model was 0.94. In addition, gender (β [SE]: 0.517 [0.386], p = 0.180) in ADNI data, age (β [SE]: 0.081 [0.039], p = 0.038) and gender (β [SE]: 1.093 [0.532], p = 0.040) in SMC data were also statistically significant variables when the p-value is based on 0.2, although the statistical significance is lower than other variables (Table 4).

We then constructed nomograms based on the multivariable logistic regression results (Figure 3). The total points were calculated from the sum of each point, and the risk probability was calculated from the total points. The probability in our nomograms referred to risk probability of being assigned to a cognitive decline group. The bias-corrected calibrated values were generated from validations based on 1,000 bootstrap resamples. The non-parametric calibration curves revealed that the bias-corrected calibration plots were close to the 45° line, indicating the nomograms are well calibrated (Supplementary Figure 1).

Finally, we visualized the risk score of each predictor and the predicted probability of belonging to a trajectory group using the predictive model (Figure 4). Specifically, for a woman with an AV45 SUVR of 1.5, HV/ICV of 0.004 and an mPACCtrt of −2, the risk scores of all predictors are 142, so the probability of belonging to the declining group was estimated to be 77.2%.