First, participants were recruited within the Alzheimer’s Disease Neuroimaging Initiative (ADNI 3, ClinicalTrials.gov ID: NCT02854033). Inclusion criteria for AD, amnestic mild cognitive impairment (aMCI), and cognitively unimpaired subjects (CU) followed the protocol of ADNI 3 (Number: ATRI-001). We only included amyloid-positive subjects who had all three imaging modalities [tau (¹⁸F-AV-1451 PET), amyloid PET (¹⁸F-AV-45 or ¹⁸F-florbetaben PET), and T1-weighted magnetic resonance imaging (MRI)]. Amyloid (Aβ) status was determined by the global standard uptake value ratio (SUVR) > 1.11 (¹⁸F-AV-45) or > 1.08 (¹⁸F-florbetaben). At baseline, the final participants comprised 25 AD, 42 prodromal AD [Aβ (+) aMCI], and 24 preclinical AD [Aβ (+) CU]. Subsequently, 77 subjects (19 AD, 36 MCI, and 22 CU) were available in the longitudinal analysis.

Second, subjects were recruited at the memory disorder clinic in the Department of Neurology at the Asan Medical Center (AMC) and the Samsung Medical Center (SMC) in Seoul, South Korea. We obtained all three modalities for each subject at baseline (tau PET [¹⁸THK-5351], Aβ PET [¹⁸F-florbetaben], and T1 MRI). All AD subjects fulfilled the clinical diagnostic criteria for AD according to the NINCDS/ADRDA, and those with aMCI met the Petersen’s criteria. Subjects with AD and aMCI were Aβ-positive as determined by brain amyloid plaque load (BAPL score) ≥ 2. CU was defined as being elderly and free of neurological disease, Clinical Dementia Rating (CDR) 0 and Mini-Mental State Examination (MMSE) > 27. At baseline, 87 participants [21 AD, 32 prodromal AD, and 34 Aβ (-) CU] were included. At 4-year follow-up, 61 and 43 subjects had available longitudinal data for the MMSE and other cognitive scores, respectively.

For dataset 1, ethics approval was obtained by the ADNI investigators. All study participants provided written informed consent. ADNI 3 is listed in the ClinicalTrials.gov registry (identifier: NCT02854033). For dataset 2, the Institutional Review Board of both hospitals approved the study, and all subjects provided written informed consent.

In dataset 1, we utilized composite scores of memory (ADNI-MEM), executive function (ADNI-EF), language function (ADNI-LAN), and visuospatial function (ADNI-VIS) in ADNI 3 to construct CR marker. These were composite z-scores from the ADNI neuropsychological battery of test. We applied the total summation of four domain scores as a global cognitive composite score. Other cognitive measures, including the Alzheimer’s Disease Assessment Scale-Cognitive Subscale (ADAS-cog) 11 and MMSE, were used for further validation. In dataset 2, we measured the global composite score for the CR marker based on previous studies (Jahng et al., 2015; Lee et al., 2019). The global composite score was obtained from the average of five domains containing 14 neuropsychological tests (e.g., attention, visuospatial, language, executive, and memory function). The tests were composed of Seoul Verbal Learning Test (SVLT-E) and Rey Complex Figure Test (RCFT) in memory domain contrasting program, Go-No-Go test and Controlled Oral Ward Association Test (COWAT) in executive function, forward and backward Digit Span Test (DST) in Attention, RCFT copying task in Visuospatial function, and Korean-Boston Naming Test (K-BNT) in language function.

We used years of education as a CR proxy. In dataset 1, we correlated the CR marker with years of education. To further validate our CR marker, education was applied to identify CR-related regions in Aβ pathology and neurodegeneration.

In ADNI dataset, MRI scans were acquired on scanners from different manufacturers (Philips, Best, Netherlands; GE, Cleveland, OH; and Siemens, Malvern, PA, United States) using harmonized protocols. We obtained 3D T1-weighted magnetization-prepared rapid gradient echo (MPRAGE) sequence, with 1 mm isotropic voxel resolution and repetition time (TR) = 2,300 ms. Aβ PET scans were recorded for 20 min at 50 min after 370 MBq ± 10% injection (¹⁸F-AV-45) or 90 min after 300 MBq ± 10% injection (¹⁸F-florbetaben) of the tracer. Tau PET scans were acquired for 30 min at 75–105 min after 370 MBq ± 10% tracer injection. In dataset 2, we obtained 3D T1-weighted MPRAGE with voxel size (AMC: 1.11 × 1.11 × 1.2 mm³, SMC: 1.0 × 1.0 × 0.5 mm³) and a TR (AMC: 6.8 ms, SMC: 9.9 ms) from the manufacturer Philips (Eindhoven, Netherlands). Tau PET scans were acquired for 20 min, commencing 50 min after the injection of 185 MBq ± 10% of ¹⁸F-THK5351. Aβ PET scans were acquired for 20 min, commencing 90 min after the injection of 300 MBq ± 10% of the tracer. Hoffman phantom-based PET harmonization was applied in datasets 1 and 2.

In both datasets, individual tau and Aβ PET scans were co-registered onto the individual T1 image and normalized into Montreal Neurological Institute (MNI) standard space. Preprocessed images were smoothed (6 mm). SUVR images were calculated for all individuals using the cerebellar gray matter as a reference. All preprocessing was conducted using SPM12 (Wellcome Trust Centre for Neuroimaging, University College London) and MATLAB R2014b (The Mathworks, Natick, MA, United States). For cortical thickness, T1-weighted images underwent preprocessing steps with Freesurfer 6.0¹. The total intracranial volume (TIV) was obtained by summing the volumes of gray matter, white matter, and CSF.

We conceptualized CR as the residual of actual cognitive performance and estimated performance; the latter was estimated from AD neuropathology, demographics, a genetic factor (ApoE ε4), and TIV. The “residual” concept for quantifying CR has been applied previously (Reed et al., 2010; Habeck et al., 2017; van Loenhoud et al., 2017). While these studies captured only the structural aspects, our method had an advantage in that the overall AD neuropathology was reflected in the model as follows (A-T-N: primary components of AD biomarkers):

Cognitive function_estimated = β₀ + β₁ × X_Tau + β₂ × X_Aβ + β₃ × X_Thickness + β₄ × X_Age + β₅ × X_Sex + β₆ × X_{ApoEε 4} + β₇ × X_TIV

(+ β₈ × X_{Aβ tracer type}: In dataset 1, Aβ tracer type was included; X_Tau, X_Aβ, and X_Thickness: Global value obtained from each imaging modality)

Cognitive function: Cognitive composite score using multiple domains

Cognitive reserve (CR) = Cognitive function_observed – Cognitive function_estimated

As a global value for each AD pathology, we extracted the global extent of tau, Aβ, and cortical thickness per subject using multimodal imaging. In tau PET, we obtained images with GM probability > 0.5 and calculated the average tau SUVR value per subject in Braak regions of interest (ROIs) (Hoenig et al., 2017). To avoid off-target effects of the tracer, we excluded ROIs in the basal ganglia and thalamus (Harada et al., 2016). In Aβ PET, we obtained images with GM probability > 0.5 and extracted a global Aβ SUVR value in combined ROIs (Jack et al., 2010). Similarly, we measured the average thickness value among 68 Desikan ROIs. Age (Guerreiro and Bras, 2015) and sex (Mielke, 2018) are demographic risk factors for AD, and ApoE allele 4 (Liu et al., 2013; Safieh et al., 2019) is considered a representative genetic factor of AD. TIV was added as a covariate for thickness and an estimate of brain reserve (Tate et al., 2011). We subsequently performed multiple linear regression with the dependent variable of cognitive function; predictors of global tau, Aβ deposition, and thickness (AD neuropathology); and covariates of age, sex, ApoE ε4 status, and TIV. The beta coefficients of these variables were applied to estimate the cognitive function. Finally, we calculated the CR marker as the residual between the actual and estimated score. According to the equation, higher CR marker value denoted greater CR as it indicated that relatively high cognitive function was maintained at a given level of AD neuropathology in the population.

For dataset 1, we calculated Pearson’s correlation between the CR marker and a conventional CR proxy, years of education. We also performed multiple linear regression using education as an outcome; CR marker as a predictor; and age, sex, and TIV as covariates.

Fundamentally, we investigated whether the CR marker could modulate the effect of brain pathology on cognitive function across the AD spectrum. To identify regions for which greater education enabled subjects to tolerate greater Aβ burden (CR-related regions), we first performed voxel-wise multiple linear regression of Aβ SUVR with education as the predictor, adjusting for age, sex, Aβ scanner type, and MMSE. The resulting t map was thresholded at the voxel level at α = 0.01 and corrected at the cluster level at FWE α < 0.05. We then extracted the average Aβ SUVR within the CR-related regions for each subject. Finally, we tested the interaction of the CR marker × Aβ SUVR on cognition, controlling for age and sex across the AD spectrum. We conducted similar procedures with cortical thickness in ROI-wise, adjusting for age, sex, MMSE, and TIV among the AD spectrum (FDR p < 0.05). We measured the atrophy value as the reciprocal of thickness value within significant regions. We tested the interaction of the CR marker × atrophy value on cognitive function, controlling for age, sex, and TIV across the AD spectrum.

For dataset 2, we performed longitudinal analysis to verify the effect of the CR marker on clinical progression. For a 4-year period, we examined the conversion rate and change in cognitive scores and disease severity across the entire group. We performed Cox-proportional hazard regression among prodromal AD (n = 32) and CU (n = 34) to investigate the relationship between the CR marker and disease conversion. In this case, we calculated the CR marker without AD subjects at baseline. Diagnostic changes to more severe stages were only considered as a disease conversion. The model contained conversion with time as an outcome measure; CR marker (continuous or binary) as a predictor; and age, sex, ApoE, and TIV as covariates. We validated the effect of the CR marker by confirming hazard ratio (HR), p-value, and 95% confidence intervals (CIs) of HR in the model. Finally, we conducted a likelihood ratio test to evaluate the goodness of model fit between two competing models with and without the CR marker, respectively. We also assessed the Akaike Information Criterion (AIC) to estimate the relative quality of statistical models for a given data. AIC deals with the tradeoff between the goodness of model fit and complexity of the data; smaller AIC indicates higher quality (Vrieze, 2012).

We conducted linear mixed models with cognitive composite score, memory score, and MMSE as outcomes; the CR marker at baseline, time, and CR marker × time as predictors; and adjusted for age, sex, ApoE, and TIV. We repeated a linear mixed model with CDR-SB as an outcome with the same predictors and covariates. Our interest was the interaction of the CR marker and time, as this referred to the effect of the CR marker on clinical progression.

These analyses were divided into AD spectrum and CU groups, as CR behavioral pattern would exert differently according to the disease status. The 95% CIs of beta coefficient of the CR marker, time, and their interaction were estimated. Finally, we performed a likelihood ratio test and AIC to estimate the additional value of the model with the interaction.

In order to estimate more reliable CIs for beta values of the interaction in clinical progression, semi-parametric bootstrapping of the linear mixed model was performed. For this process, we made 1,000 bootstrap samples by resampling the original data with replacement and calculated 95% CIs of the bootstrapped coefficients.

For dataset 1, we repeated linear mixed models with ADAS-cog 11 as outcomes; the CR marker at baseline, time, and CR marker × time as predictors; and adjusted for age, sex, ApoE, and TIV. The analyses were also divided into AD spectrum and CU groups. We conducted a likelihood ratio test and AIC to estimate the additional value of the model with the interaction.

In a longitudinal analysis of dataset 2, to figure out the difference between using continuous Aβ SUVR value and using visual diagnosis of Aβ, first we applied binary value (positivity, negativity) instead of continuous value in the calculation of the CR marker. Then, we repeated linear mixed model analysis and compared the results with the model constructed from continuous Aβ SUVR.

Second, we investigated the effect of education on clinical progression in dataset 2. Therefore, we conducted linear mixed model analysis by replacing the CR marker with education into the model.

Finally, we added the global pathological burden (global tau, Aβ, and thickness value) in the linear mixed model as covariates to partly account for the explanation that the CR marker is only a derivative of pathological measurement in relation to cognitive function.

Statistical analyses were conducted in SPSS 18 (Chicago, IL, United States). To compare participant characteristics, chi-square tests and ANOVA tests were performed. In the longitudinal analysis, linear mixed model analysis and semi-parametric bootstrapping were conducted using R².

As shown in Table 1, we observed significant differences among the three groups in global tau, Aβ deposition, cortical thickness, and cognitive scores. In dataset 1, we investigated the A/T/N classification status using CSF information. Among the 91 subjects, 62 subjects (16 AD, 26 MCI, 20 CU) have baseline CSF information of p-tau181 and t-tau. Unlike the pathological burden obtained from imaging modalities, CSF levels of p-tau and t-tau did not differ significantly among the three groups (Table 1). Abnormal levels of CSF biomarkers were defined as p-tau > 27 pg/ml (T +) and t-tau > 300 pg/ml (N +) (Blennow et al., 2019). In the AD subjects, A/T/N (+/+/+) group accounted for 68.8% (11/16) and A/T/N (+/−/−) group accounted for the rest. Among the MCI subjects, A/T/N (+/+/+) group accounted for 42.3% (11/26), A/T/N (+/−/−) group accounted for 50% (13/26), and A/T/N (+/+/−) group accounted for the rest. In the CU subjects, each A/T/N (+/+/+) and A/T/N (+/−/−) group accounted for 45% of the total, and A/T/N (+/+/−) group accounted for the rest (10%). Therefore, it was confirmed that the AD group showed a relatively large distribution of A/T/N (+/+/+). The average follow-up period was 26 months. In total, 77 subjects (19 AD, 36 MCI, and 22 CU) were available in the longitudinal analysis.

In dataset 2, the mean age was lower in the AD group than that in the prodromal AD and CU groups, attributing to the inclusion of 15 early-onset AD. The portion of ApoE ε4 carriers was higher in the AD spectrum than that in the CU. We could not investigate the A/T/N classification of the respective group because the dataset did not contain CSF information. The average follow-up period was 27 months. A total of 61 participants (18 AD, 28 prodromal AD, and 15 CU) had longitudinal MMSE, and 43 subjects (11 AD, 20 prodromal AD, and 12 CU) had follow-up composite score, memory score, and CDR-SB. The total incidence of diagnostic conversion was 28.8%. The conversion rate of prodromal to AD was 53.1%, and the conversion rate of CU to prodromal or CU to AD was 2.9%, respectively.

In dataset 1, we confirmed the relationship between cognitive performance and each predictor via beta values. The linear regression model revealed that composite score was negatively associated with global tau (β_tau = −6.35) and Aβ deposition (β_Aβ = −1.80). The composite score was positively associated with the global cortical thickness (β_thickness = 12.85). The R² of the model was 0.52 (F-test, p = 1.49^*e^–10, adjusted R² = 0.48). There was no multicollinearity among variables (maximum VIF < 1.85). As a result, the estimated composite z-scores were significantly correlated with the actual composite z-scores using Pearson’s correlation (r = 0.72, p = 6.51^*e^–16). Then, we dichotomized the actual z-score using median value (median = 0.819) and investigated whether the estimated z-scores predicted actual z-scores well. In a discrete ROC curve analysis, the estimated composite z-score classified high and low actual z-scores well (AUC = 0.83, 95% CI = 0.75–0.92, accuracy = 0.77, sensitivity = 0.83, specificity = 0.71). Through this analysis, we can conclude that the estimated cognitive function from the biological markers, demographics, and genetic factors would be a reasonable estimate of the actual cognitive function. Finally, we calculated the CR marker using the differences between the actual and estimated composite scores.

Using dataset 2, we already validated the suitability of the model (Lee et al., 2019). In brief, the constructed model demonstrated that the composite score was significantly associated with each global AD neuropathology (β_tau = −18.86, β_Aβ = −32.05, and β_thickness = 34.11). The R² of the model was 0.57 (F-test, p < 0.00001). There was no multicollinearity within variables (maximum VIF < 1.9). Estimated composite scores were significantly correlated with actual composite scores (r = 0.76, p = 2.21^*e^–17). Also, we dichotomized the actual composite score using median value (median = 47.2) and investigated whether the estimated composite scores predicted high and low composite scores well. In a discrete ROC curve analysis, the estimated composite score classified high and low actual composite scores reasonably well (AUC = 0.92, 95% CI = 0.87–0.98, accuracy = 0.82, sensitivity = 0.77, specificity = 0.86). Then, we calculated the CR marker from the residuals between the actual and estimated composite scores.

In dataset 1, the CR marker was significantly correlated with years of education (r = 0.29, p = 0.005). Subjects with more years of education had greater CR than individuals with fewer years of education. The positive association between the CR marker and years of education remained even after adjusting for age, sex, and TIV (t = 2.94, p = 0.004).

Fundamentally, we tested whether the CR marker could modulate the association between AD pathological burden and cognitive function across the AD spectrum. In Figures 1A,B, for Aβ pathology, higher level of education was associated with greater Aβ deposition within the left superior and middle frontal regions, after adjusting for age, sex, Aβ scanner type, and MMSE (t = 2.37, p < 0.01). We examined whether the CR marker modified the effect of left frontal Aβ burden on cognition in the AD spectrum, controlling for age, sex, and ApoE. There was a significant interaction effect of the CR marker × left frontal Aβ SUVR on ADAS-cog 11 (t-stat = 2.17, p = 0.034). The interaction effect of CR marker × left frontal Aβ SUVR on MMSE was marginal (t-stat = −1.93, p = 0.058). As shown in Figures 1C,D, higher education was associated with greater cortical atrophy within the left superior temporal gyrus and supramarginal gyrus after controlling for age, sex, MMSE, and TIV. The interaction effect of the CR marker × left temporoparietal atrophy on ADAS-cog 11 was confirmed to be significant (t-stat = 2.29, p = 0.026).

Cox regression analysis revealed that in prodromal AD and CU groups, the CR marker was negatively associated with conversion rate, indicating that higher CR was related to lower conversion risk (continuous: HR = 0.57, 95% CI: 0.34 ∼ 0.97, β = −0.56, p = 0.037; binary: HR = 0.33, 95% CI: 0.11∼0.97, β = −1.10, p = 0.044). The model with the CR marker showed a better model fit than that of the model without the CR marker in a likelihood ratio test (continuous: p = 0.031; binary: p = 0.037). The model with the CR marker had lower AIC value than that of the model without the CR marker [without CR marker: 115.75, with CR marker: 113.12 (continuous)/113.40 (binary)].

In the analysis of cognitive decline among AD spectrum, clear cognitive decline was observed across time in all cognitive scores. Particularly, higher CR marker was related to more exacerbated decline in cognitive performance in the AD spectrum (MMSE: p = 0.026, composite score: p = 0.045, memory score: p = 0.036) throughout Table 2 and Figure 2. In CU, cognitive scores decreased over time, with the exception of MMSE. The CR marker also modified the relationship between cognitive function and time. However, the pattern modulated by the CR marker was different from that of the AD spectrum, such that subjects with high CR exhibited attenuated cognitive decline (MMSE: p = 0.017, composite score: p < 0.001, memory score: p = 0.003) throughout Table 2 and Figure 2. The likelihood ratio test revealed that the model with the interaction showed a better model fit than that of the model without interaction in all cognitive scores among CU and AD spectrum. In the aspects of AIC, the model with interaction also had lower AIC values in both AD spectrum and CU as shown in Table 3. AD spectrum participants with high CR exhibited more drastic alterations in disease severity in Table 2. The model with the interaction showed a better model fit than that of the model without interaction among AD spectrum in Table 3.

More reliable results were obtained from 95% CIs derived from bootstrapping, indicating that the behavioral pattern of CR differed depending on disease stage and was related to accelerated cognitive decline in AD spectrum subjects but with alleviated cognitive decline in CU subjects as shown in Table 4.

In the analysis of ADAS-cog 11 among AD spectrum in dataset 1, the CR marker modified the relationship between cognitive decline and time, such that subjects with higher CR marker exhibited worse cognitive decline (β = 0.07; p = 0.026; Supplementary Table 1). Since the higher score in ADAS-cog 11 indicates greater cognitive impairment, AD spectrum subjects with higher CR marker showed more steeper increase in the score. However, in the CU group, the CR marker showed a tendency to modulate the relationship between cognitive function and time, but did not reach a significant level (β = −0.02; p = 0.20; Supplementary Table 1). The likelihood ratio test also revealed that the model with the interaction showed a better model fit than that of the model without the interaction in the AD spectrum (Supplementary Table 1).

The constructed model using visual diagnosis of Aβ demonstrated that the composite score was significantly associated with each global AD neuropathology (β_tau = −18.36, β_Aβ = −17.42, β_thickness = 22.41). The R² of the model was 0.62 (F-test, p < 0.00001, adjusted R² = 0.59). We calculated the CR marker from the residuals between the actual and estimated composite scores. The CR marker correlated well with years of education (r = 0.49, p < 0.00001). In the analysis of cognitive decline among AD spectrum, the CR marker showed a tendency to modulate the relationship between cognitive function and time, but did not reach a significant level (MMSE: p = 0.08, composite score: p = 0.13, memory score: p = 0.26, Supplementary Table 2). However, subjects with high CR exhibited more drastic alterations in disease severity (CDR-SB: p = 0.046). In the longitudinal analysis of CU subjects, the CR marker modulated the association between cognitive performance and time (MMSE: p = 0.001, composite score: p < 0.001, memory score: p = 0.027, Supplementary Table 2). In case of CDR-SB, CU subjects with high CR showed more delayed alterations in disease severity (CDR-SB: p = 0.024). Therefore, it was found that although the CR marker using binarization of amyloid PET tended to modulate the relationship between cognition and time to some extent, the CR marker using continuous SUVR value reflects the properties of CR better.

In case of education, years of education was not associated with accelerated cognitive decline (memory score: p = 0.12, MMSE: p = 0.19, composite score: p = 0.78). In the CU group, education modified the relationship between cognitive function and time, such that subjects with higher education exhibited attenuated cognitive decline (memory score: p = 0.003, MMSE: p = 0.004, composite score: p < 0.001, Supplementary Table 3).

We verified whether the CR marker had an independent effect on clinical progression by adjusting global pathological burden as covariates. Even though the values of global pathology were added as covariates, the CR marker still had an effect on clinical progression among AD spectrum in MMSE, memory score, but not composite score (MMSE: p = 0.023, memory score: p = 0.047, composite score: p = 0.06, Supplementary Table 4). The subjects with high CR showed more exacerbated cognitive decline in the AD spectrum. In the analysis of CU subjects, the CR marker also affected the relationship between cognitive function and time (MMSE: p = 0.006, memory score: p = 0.003, composite score: p < 0.001, Supplementary Table 4). In case of CDR-SB, CU subjects with high CR showed more delayed alterations in disease severity (CDR-SB: p = 0.047). Through this subsequent analysis, it was determined at least in part that the CR marker could have an additional effect on clinical progression, not simply a derivative of global pathology.