Comparative Prognostic Performance of Hippocampal Volume and Plasma Biomarkers in Amyloid-Negative Subjective Cognitive Decline: A Four-Year Follow-up Authors: Bora Yoon, Hyuk-je Lee, Yun Jeong Hong, Kee Hyung Park, SangYun Kim, Min Jeong Wang, Seong Hye Choi, Sungmin Kang, Jee Hyang Jeong, Dong Won Yang Published: 2026 ABSTRACT AbstractBackground and PurposeSubjective cognitive decline (SCD) may precede mild cognitive impairment (MCI) and dementia. However, the prognosis of patients with amyloid-negative (A-) SCD remains unclear. We investigated whether baseline plasma and imaging biomarkers predict four-year clinical progression in A-SCD.MethodsSixty-five individuals with A-SCD were followed for four years. Baseline plasma biomarkers-phosphorylated tau 181 (pTau181), glial fibrillary acidic protein, neurofilament light chain (NfL), and pTau181/Aβ42 ratio-and MRI markers (hippocampal volume and white matter hyperintensity [WMH] grade) were assessed. Outcomes included MCI conversion and cognitive progression defined as ≥1-point (Prog-1) or ≥2-point (Prog-2) decline in Korean Mini-Mental State Examination. Predictive performance was evaluated using multivariable logistic regression and bootstrap-validated receiver operating characteristic analyses.ResultsFourteen participants (21.5%) converted to MCI, and 24 (36.9%) were classified as Prog-1. Plasma biomarkers did not show consistent predictive performance across predefined endpoints (all p>0.05), although NfL demonstrated modest discrimination for MCI conversion (area under the curve [AUC] 0.686, p=0.008). Hippocampal volume demonstrated significant discriminative performance for MCI conversion (AUC 0.765, p=0.016) and Prog-1 (AUC 0.933, p<0.001). For Prog-2, hippocampal volume showed high discrimination (AUC 0.952, p<0.001). WMH grade also demonstrated significant discrimination for MCI conversion (AUC 0.835, p<0.001) and Prog-1 (AUC 0.857, p<0.001). Bootstrap analyses (1,000 iterations) tended to show higher AUC estimates for imaging markers compared with plasma biomarkers.ConclusionsIn this A-SCD cohort, hippocampal atrophy demonstrated more consistent prognostic performance than Alzheimer’s disease-focused plasma biomarkers for predicting four-year conversion to MCI and cognitive decline, suggesting a potential role for MRI-based neurodegeneration markers in clinical risk stratification.Keywords: subjective cognitive decline, hippocampus, amyloid-negative, magnetic resonance imaging, biomarkers INTRODUCTION Subjective cognitive decline (SCD) is defined as self-reported cognitive complaints without objective cognitive impairment on standardized neuropsychological assessments.1 As the earliest clinical manifestation within the Alzheimer’s disease (AD) continuum, SCD is associated with an increased risk of progression to mild cognitive impairment (MCI) and dementia, particularly in individuals with brain amyloidosis (A+).2,3 However, 75%–80% of individuals with SCD do not exhibit amyloid pathology (amyloid-negative [A-] SCD),4 and this group may include argyrophilic grain disease, primary age-related tauopathy (PART), limbic-predominant age-related TDP-43 encephalopathy (LATE), cerebrovascular disease, Lewy body pathology, or psychiatric conditions.2,5,6,7 Thus, A-SCD represents a biologically heterogeneous and comparatively understudied subgroup whose progression mechanisms may differ fundamentally from those of A+SCD. Cognitive decline in A-SCD populations correlates with markers of neurodegeneration (e.g., hippocampal atrophy and tau changes on imaging), vascular pathology, genetic risk factors (e.g., apolipoprotein E [APOE] ε4), and informant-confirmed cognitive complaints.8,9,10 Structural imaging biomarkers, such as hippocampal volume and white matter hyperintensities (WMHs), reflect neurodegeneration and cerebrovascular pathology and are increasingly recognized as contributors to cognitive impairment and dementia progression.11,12,13 Reduced hippocampal volume on MRI indicates neuronal loss and correlates with memory impairment, even without amyloid pathology, particularly in non-AD-related neurodegenerative processes.6,7 WMHs, which indicate cerebral small vessel disease, reflect cerebrovascular damage, and correlate with executive dysfunction and cognitive decline risk, independent of or with amyloid pathology.11,12 Recent advances in ultrasensitive blood-based assays have enabled the early detection of Alzheimer-related pathophysiology and risk stratification.14,15 Plasma phosphorylated tau at threonine-181 (pTau181) and threonine-217 are specific markers of AD-associated tau pathology16,17 and have been established as core diagnostic, staging, and prognostic biomarkers for AD.18 Glial fibrillary acidic protein (GFAP), which marks astrocytic activation, is elevated in individuals with cerebral amyloid deposition before the onset of clinical symptoms.19 Neurofilament light chain (NfL), which reflects neuronal axonal damage, correlates with cognitive decline in early AD14,20,21 and is now included in the revised AD criteria as a staging and prognostic indicator.18 However, the prognostic value of these plasma biomarkers in patients with A-SCD is unclear. Since A-SCD may include diverse non-AD pathologies, such as PART, hippocampal sclerosis, and LATE,22 prognostic markers likely reflect neurodegeneration rather than amyloid-driven processes. In this context, the prognostic performance of predominantly AD-oriented plasma biomarkers may be attenuated in A-SCD, where biomarker signals are more likely to reflect heterogeneous neurodegenerative processes rather than amyloid-driven pathology. This study investigated whether baseline plasma or MRI-based biomarkers could predict four-year clinical conversion or cognitive decline in A-SCD individuals, comparing the prognostic utility of plasma biomarkers and structural MRI markers, focusing on hippocampal volume and WMH grade. METHODS Study design and participants This investigation used data from a longitudinal Cohort Study to Identify Predictors for Clinical Progression to Mild Cognitive Impairment or Dementia from Subjective Cognitive Decline (CoSCo).23 Participants aged ≥60 years with SCD were recruited from dementia centers in Korea. The eligibility criteria were as follows: 1) self-reported persistent cognitive decline without a dementia or MCI diagnosis; 2) normal performance on comprehensive neuropsychological testing (Seoul Neuropsychological Screening Battery-II, SNSB-II),24 defined as cognitive scores at or above the 7th percentile, except for verbal memory scores between the 7th and 50th percentiles; 3) completion of elementary education; and 4) willingness to participate with informed consent. The exclusion criteria included neurological disorders affecting cognition (e.g., Parkinson’s disease and stroke), severe psychiatric illnesses (e.g., uncontrolled depression, schizophrenia, and substance abuse), and metabolic conditions (hypothyroidism, vitamin deficiency, and infection). Ethics approval for this study was obtained from the Institutional Review Board of Seoul St. Mary’s Hospital, The Catholic University of Korea (IRB no. KC18ONDI0394), and the study was conducted in accordance with the Declaration of Helsinki. All participants provided written informed consent. This study is part of the CoSCo study, registered at the Clinical Research Information Service (CRIS), Republic of Korea (KCT0004244). We conducted a longitudinal cohort study involving 120 individuals with SCD. The amyloid positron emission tomography (PET) status was assessed at baseline, and those confirmed as A- (n=93, 77.5%) met the criteria for the A-SCD cohort. Over four years, 28 participants (30.1%) were lost to follow-up, leaving 65 A-SCD participants with complete four-year cognitive outcome data. Clinical and neuropsychological assessment Clinical assessments at baseline included structured history-taking for medical comorbidities, including the Framingham Cardiovascular Risk Score25 for vascular risk assessment, family history of dementia, medication review, and physical and neurological examinations of the patient. Participants completed the Korean Mini-Mental State Examination-2nd edition (K-MMSE)26 and the SNSB-II battery, providing comprehensive cognitive profiling adjusted for age, sex, and education. Plasma biomarker measurements Blood samples were collected after overnight fasting and processed within one hour of collection by centrifugation. Plasma aliquots were frozen and stored at -80℃ until analysis. The plasma concentrations of pTau181, GFAP, NfL, amyloid-beta 40 (Aβ40), and amyloid-beta 42 (Aβ42) were measured using single-molecule array (Simoa) assays on the HD-X platform (Quanterix). The Neurology 4-Plex E kit (Quanterix Corp.) was used for GFAP, NfL, and Aβ peptides, whereas pTau181 levels were quantified using the pTau181 Advantage V2.1 assay. Paramagnetic bead-based enzyme-linked immunosorbent assay (ELISA) technology was used, with samples measured in duplicate at a 4-fold dilution. The lower limits of quantification of the assay were as follows: pTau181, 2.0 pg/mL; GFAP, 11.6 pg/mL; NfL, 1.6 pg/mL; Aβ40, 4.08 pg/mL; and Aβ42, 1.51 pg/mL. Amyloid PET imaging and processing Amyloid PET imaging was performed using standardized [18F]-florbetaben protocols.27 At each site, nuclear medicine professionals visually interpreted amyloid PET scans and classified them as A+ or A- using the Brain Amyloid Plaque Load (BAPL) scoring system.28 Centiloid values (CL) were calculated from global standardized uptake value ratios (SUVRs) using BTXBrain software (Brightonix Imaging Inc.). BTXBrain performs spatial normalization of amyloid PET images to the Montreal Neurological Institute (MNI) space using a deep learning model trained on 1,000 paired PET/MR datasets from multiple centers.29 The software computes whole-cortex and reference region count rates to convert global SUVRs into centiloid units using the calibration equations by Rowe et al. (2017).30 Participants classified as A- by visual PET assessment were included in the analysis, focusing on clinical progression and biomarker profiles. A sensitivity analysis was performed using a stricter threshold (CL<25) to define amyloid negativity according to validated centiloid cutoffs.31 MRI imaging and processing The MRI protocols included axial T1-weighted, T2-weighted, fluid-attenuated inversion recovery (FLAIR), and high-resolution three-dimensional T1 thin-section sequences on a 3.0T scanner. The parameters matched the Alzheimer’s Disease Neuroimaging Initiative (ADNI) protocol.32 Volumetric segmentation was performed using AQUA 2.0 software (Neurophet). The total and regional brain volumes, including the bilateral hippocampal volumes, were extracted from T1-weighted images. Hippocampal volumes were normalized for total intracranial volume (ICV), age, and sex using z-scores based on the software’s internal reference framework. WMHs were rated on FLAIR images using the Clinical Research Center for Dementia of South Korea (CREDOS) ischemia classification system (Noh et al.33 criteria): Periventricular WMHs were scored as P1 (≤5 mm), P2 (6–10 mm), and P3 (>10 mm); deep WMHs as D1 (≤10 punctate ≤10 mm), D2 (>10 punctate or early confluent ≤20 mm), and D3 (confluent >20 mm); P and D scores were summed for WMH grades of 1 (minimal), 2 (moderate), or 3 (severe). Although automated volumetric analysis was not available at the time of data analysis, the visual rating scale was selected because of its strong correlation with automated volumetric measurements (ρ=0.880, p<0.001), as well as its established clinical applicability in multicenter studies and high inter-rater reliability (Cohen’s κ=0.88).33 Lacunes were defined as 3–15 mm lesions appearing hyperintense on T2-weighted images with a perilesional halo on FLAIR.34 Cerebral cortical microbleeds were identified as round, low-signal lesions <10 mm in size in the lobar regions.34 The number of lacunes, microbleeds, and WMH grades were noted. Outcome measures and definitions At the four-year follow-up, clinical progression was assessed through clinical evaluation and cognitive testing. This study evaluated two prespecified and complementary outcomes to capture different facets of disease progression: 1) diagnostic transition and 2) longitudinal cognitive decline. First, the diagnostic transition from SCD to MCI (converter) was defined on established clinical criteria, including subjective cognitive concerns accompanied by objective impairment (test scores below the 7th percentile in at least one cognitive domain), preserved daily functioning, and the absence of dementia. Second, longitudinal cognitive decline (progressor) was defined as a ≥1-point decrease in the K-MMSE score from baseline to follow-up (Prog-1). While diagnostic conversion represents a categorical clinical milestone, this cognitive progression measure was utilized to capture subtle, dichotomized changes in global cognition. Additionally, to account for potential measurement variability and ensure the robustness of our clinical findings, a more stringent threshold of a ≥2-point decrease in the K-MMSE score (Prog-2) was established as an alternative criterion for cognitive progression in a sensitivity analysis. Statistical analyses To compare baseline demographics, clinical characteristics, and biomarker profiles among the 65 completers across different outcome groups, we used independent t-tests or Mann–Whitney U tests and χ2 tests for continuous variables and χ2 or Fisher’s exact tests for categorical variables, as appropriate. Continuous variables are expressed as mean±standard deviation (SD) or median [interquartile range] based on their distribution, and categorical variables are presented as counts (%). To assess potential selection bias due to attrition, baseline demographics and initial cognitive scores were also compared between completers (n=65) and dropouts (n=28) using the same group comparison methods described above. To evaluate the prognostic utility of biomarkers, multivariable logistic regression models were constructed for three distinct outcomes: 1) diagnostic conversion from SCD to MCI (MCI converters), 2) cognitive progression defined as any decrease in K-MMSE score (≥1 point), and 3) a more stringent threshold of cognitive progression (≥2 points decrease in K-MMSE) as a sensitivity analysis to account for potential measurement variability and practice effects. All plasma biomarkers (pTau181, GFAP, NfL, and pTau181/Aβ42 ratio) showed non-normal distributions. For comparison, we log-transformed the plasma biomarkers and reported the median raw values. Prior to modeling, all continuous biomarkers, including hippocampal volume, WMHs, and plasma markers, were standardized (z-scored) to ensure comparability of effect size across biomarkers. Models for plasma biomarkers were adjusted for age and sex, while models for neuroimaging biomarkers (hippocampal volume and WMH grade) were adjusted for age, sex, and ICV. To improve clinical interpretability, odds ratios (ORs) for continuous biomarkers were reported per one SD change (increase for plasma biomarkers, WMHs, and per one SD decrease for hippocampal volume). Model calibration was assessed using the Hosmer–Lemeshow goodness-of-fit test and Brier scores, where a higher p-value in the Hosmer–Lemeshow test and a lower Brier score indicate better calibration between predicted probabilities and observed outcomes. Receiver operating characteristic (ROC) curves were generated, and the area under the curve (AUCs) with 95% confidence intervals (CIs) and p-values (tested against a null hypothesis of AUC=0.5) were calculated using the DeLong method via the pROC package (version 1.18.0; https://CRAN.R-project.org/package=pROC). We performed bootstrap resampling with 1,000 iterations (via the boot package, version 1.3–28; https://CRAN.R-project.org/package=boot) to assess the robustness of our estimates given the limited sample size (n=65). Bootstrap resampling was conducted at the level of individual participants. In each bootstrap sample, multivariable logistic regression models were fully re-estimated and the corresponding AUCs recalculated to obtain the bootstrap distribution of performance metrics. Optimal thresholds were identified using the Youden index to determine sensitivity, specificity, positive predictive value, and negative predictive value. To directly compare the prognostic performance of imaging markers against plasma biomarkers, we calculated the differences in AUC (ΔAUC) and their corresponding 95% CIs using the same bootstrap resampling procedure with 1,000 iterations. A difference was considered statistically significant if its 95% CI did not include zero. Analyses were conducted using R version 4.2.2 (R Foundation for Statistical Computing; https://www.R-project.org/), with logistic regression performed using the base stats package. All statistical tests were two-tailed, with significance set at p<0.05. Given the exploratory nature and limited sample size (n=65), no a priori power calculation was performed; instead, the study focused on a systematic comparison of potential biomarkers, with precision evaluated through the reporting of 95% CIs and optimism-corrected effect sizes. RESULTS Participant characteristics Of the initial A-SCD participants, 28 (30.1%) were lost to follow-up. However, there were no significant differences in baseline age, sex, education, or K-MMSE scores between completers (n=65) and dropouts (n=28) (all p>0.05) (Supplementary Table 1 in the online-only Data Supplement). The demographic and clinical characteristics of the 65 participants who completed the four-year follow-up are summarized in Table 1. The mean age was 69.9±5.8 years, and 38 (58.5%) were female. Over four years, 14 people (21.5%) developed MCI (“converters”), while 51 (78.5%) stayed stable (“non-converters”). Regarding cognitive decline, 24 participants (36.9%) were classified as progressors using the any-decline criterion (≥1-point decrease in K-MMSE; Prog-1), and 16 participants (24.6%) met the more stringent criterion for clinical progression (≥2-point decrease in K-MMSE; Prog-2). Table 1Demographics and characteristics of 65 participantsCharacteristicsConverters (n=14)Non-converters (n=51) p Prog (≥1 pt) (n=24)Non-prog (<1 pt) (n=41) p Prog (≥2 pt) (n=16)Non-prog (<2 pt) (n=49) p Age (yr)70.9±6.369.6±5.70.49371.0±5.269.2±6.10.21571.6±4.969.3±6.00.177Sex, female7 (50.0)31 (60.8)0.54713 (54.2)25 (61.0)0.61210 (62.5)28 (57.1)0.932Education (yr)10.6±4.011.5±4.00.46210.7±3.911.6±4.00.3459.8±3.711.8±4.00.077APOE ε4 carrier2 (14.3)7 (13.7)>0.9994 (16.7)5 (12.2)0.7153 (18.8)6 (12.2)0.812Composite Centiloids4.0±14.42.2±11.40.672-0.6±11.04.5±12.30.088-0.6±12.43.7±11.80.220K-MMSE scoreBaseline26.5±1.727.5±2.00.07028.3±1.726.7±1.90.00128.5±1.826.9±1.90.004Four-year follow-up25.4±2.727.9±1.30.00526.2±2.228.0±1.40.00125.8±2.427.9±1.5<0.001Neuropsychological tests (percentile)DST-F47.6±33.366.5±26.50.06760.6±30.263.4±28.40.71563.8±30.561.9±28.70.829K-BNT59.1±23.262.2±28.40.68369.4±27.556.9±26.30.07874.8±24.457.2±26.90.024RCFT-copy47.9±23.458.3±22.40.15655.2±20.356.6±24.40.80353.1±23.957.0±22.60.550SVLT-DR23.4±12.928.1±13.20.23925.2±12.628.2±13.50.36725.4±13.227.6±13.20.563RCFT-DR50.6±26.048.3±23.60.76946.5±23.250.1±24.50.56244.0±19.250.3±25.20.358DSC50.1±23.268.0±23.30.01966.5±22.462.8±25.50.54072.5±18.761.4±25.40.113COWAT-p47.1±26.257.6±28.60.20660.2±26.452.5±29.30.28259.1±25.754.1±29.20.541K-TMT-B52.1±24.166.1±20.40.06160.0±21.664.9±22.00.39262.3±21.463.4±22.20.865K-CWST-CR46.4±26.659.7±23.10.10361.6±23.854.1±24.50.22664.3±22.254.4±24.70.158Data are presented as mean±standard deviation or the number (%). Cognitive progression was defined as a decrease in the K-MMSE score from baseline to the four-year follow-up. Progressors (≥1 pt) include all individuals who showed any decline in K-MMSE score. Progressors (≥2 pt) represent a subgroup with a more robust clinical decline of 2 points or more, analyzed as a sensitivity analysis to account for measurement variability. The significance of differences by group were evaluated with t-test for continuous variables and chi-squared test for categorical variables while Mann–Whitney test and Fisher’s exact test were used if appropriate. All tests were two-sided and not adjusted for covariates.APOE, apolipoprotein E; COWAT, Controlled Oral Word Association Test (phonemic); DSC, Digit Symbol Coding Test; DST-F, Digit Span Test-Forward; K-BNT, Korean-Boston Naming Test; K-CWST-CR, Korean-Stroop Test-Color Reading; K-MMSE, Korean Mini-Mental State Examination, 2nd edition; K-TMT, Korean-Trail Making Test; RCFT, Rey Complex Figure Test; RCFT-DR, Rey Complex Figure Test-delayed recall; SVLT-DR, Seoul Verbal Learning Test-delayed recall. Demographics and characteristics of 65 participants Data are presented as mean±standard deviation or the number (%). Cognitive progression was defined as a decrease in the K-MMSE score from baseline to the four-year follow-up. Progressors (≥1 pt) include all individuals who showed any decline in K-MMSE score. Progressors (≥2 pt) represent a subgroup with a more robust clinical decline of 2 points or more, analyzed as a sensitivity analysis to account for measurement variability. The significance of differences by group were evaluated with t-test for continuous variables and chi-squared test for categorical variables while Mann–Whitney test and Fisher’s exact test were used if appropriate. All tests were two-sided and not adjusted for covariates. APOE, apolipoprotein E; COWAT, Controlled Oral Word Association Test (phonemic); DSC, Digit Symbol Coding Test; DST-F, Digit Span Test-Forward; K-BNT, Korean-Boston Naming Test; K-CWST-CR, Korean-Stroop Test-Color Reading; K-MMSE, Korean Mini-Mental State Examination, 2nd edition; K-TMT, Korean-Trail Making Test; RCFT, Rey Complex Figure Test; RCFT-DR, Rey Complex Figure Test-delayed recall; SVLT-DR, Seoul Verbal Learning Test-delayed recall. There were no significant differences in age, sex, education, or APOE ε4 carrier status between converters and non-converters, or progressors and non-progressors across both definitions (Table 1). Both progressor groups (Prog-1 and Prog-2) exhibited significantly higher baseline K-MMSE scores compared to their respective non-progressor counterparts (≥1 pt: 28.3±1.7 vs. 26.7±1.9, p=0.001; ≥2 pt: 28.5±1.8 vs. 26.9±1.9, p=0.004). In neuropsychological assessments, progressors (≥2 pt) showed significantly higher baseline scores in the Korean version of the Boston Naming Test (p=0.024), but no other significant differences were observed in other cognitive domains at baseline between the groups (Table 1). Baseline amyloid-beta deposition, measured by composite Centiloids, was very low (<10) across all participants, consistent with the amyloid-negative inclusion criteria. There were no significant differences in CL between converters and non-converters (4.0±14.4 vs. 2.2±11.4; p=0.672) or between Prog-1 and non-progressors (NP)-1 (-0.6±11.0 vs. 4.5±12.3; p=0.088) (Table 1). Only one participant had a CL >25 (26.6), yet the patient remained clinically stable, with normal brain volume and plasma biomarker levels. The baseline plasma concentrations of pTau181, GFAP, NfL, and pTau181/Aβ42 ratios showed no significant differences between converters and non-converters or between the progressors and non-progressors across all decline thresholds (all p>0.05) (Table 2). Regarding neuroimaging markers, there were no significant differences in hippocampal volumes or normalized hippocampal z-scores between converters and non-converters (-0.32±1.10 vs. 0.09±0.97, p=0.222). However, the total hippocampal volumes were significantly smaller in the Prog-1 (≥1 pt: 2,326.94±196.67 mm3) than in the NP-1 group (2,436.32±226.47 mm3; p=0.006), with hippocampal z-scores tending to be lower in the Prog-1 group (-0.31±0.90 vs. 0.19±1.04; p=0.052). When applying the more stringent threshold (≥2 pt), hippocampal volumes remained significantly smaller in Prog-2 (2,130.8±196.7 mm3) compared to NP-2 (2,485.6±226.5 mm3; p=0.004), with lower hippocampal z-scores in the Prog-2 group (-1.22±1.10 vs. 0.31±0.82; p=0.004) (Table 2). Table 2Baseline biomarker values between the groupsBiomarkerConverters (n=14)Non-converters (n=51) p Prog (≥1 pt) (n=24)Non-prog (<1 pt) (n=41) p Prog (≥2 pt) (n=16)Non-prog (<2 pt) (n=49) p Plasma biomarkerspTau181 (pg/mL)25.00 [21.20–30.65]27.00 [19.55–30.30]0.70228.05 [20.52–31.50]24.10 [19.80–28.80]0.39627.55 [21.05–30.15]24.60 [19.80–30.80]0.626GFAP (pg/mL)108.40 [73.26–139.62]101.20 [74.38–138.90]0.99496.58 [76.72–144.38]105.50 [73.20–132.80]0.85498.33 [74.79–161.53]102.60 [73.20–132.80]0.621NfL (pg/mL)17.85 [14.57–20.88]20.60 [15.55–27.00]0.31519.80 [15.55–24.98]20.50 [15.50–26.90]0.95720.35 [14.95–25.82]20.50 [15.60–25.90]0.970pTau181/Aβ423.37 [2.87–4.04]3.57 [2.71–4.21]0.9683.55 [2.91–4.15]3.43 [2.72–4.22]0.7193.55 [3.07–4.14]3.43 [2.72–4.22]0.732Imaging biomarkersICV (mm3)1,474,191.40±147,609.391,501,773.82±116,920.280.7151,474,937.00±136,993.091,505,103.53±117,069.350.6251,438,749.25±181,576.351,508,117.50±106,870.190.312Right HC (mm3)1,162.05±118.511,217.69±112.870.1341,169.42±103.341,226.41±118.610.0541,061.31±103.431,250.76±118.600.004Left HC (mm3)1,163.76±125.151,198.02±107.110.3651,157.52±97.581,209.91±115.690.0641,069.50±97.631,234.92±115.710.007Total HC (mm3)2,325.81±240.342,415.71±212.740.4962,326.94±196.672,436.32±226.470.0062,130.85±196.712,485.64±226.500.004Total HC z-score-0.32±1.100.09±0.970.222-0.31±0.900.19±1.040.052-1.22±1.100.31±0.820.004WMH grade (G1/2/3)3/6/5 (21.4/42.9/35.7)23/18/10 (45.1/35.3/19.6)0.0325/9/10 (20.8/37.5/41.7)21/15/5 (51.2/36.6/12.2)0.01111/2/2 (68.8/12.5/12.5)32/14/0 (65.3/28.6/0.0)0.025Lacunes1 (7.1)3 (5.9)>0.9992 (8.3)2 (4.9)0.5563 (18.8)2 (4.1)0.170Microbleed1 (7.1)3 (5.9)>0.9991 (4.2)2 (4.9)>0.9990 (0.0)3 (6.1)0.743Data are presented as median [IQR] for plasma biomarkers and as mean±SD or n (%) for imaging biomarkers. Continuous variables with approximately normal distributions (intracranial volume, hippocampal volumes, hippocampal volume z-score) are shown as mean±SD; skewed variables (plasma biomarkers pTau181, GFAP, NfL, pTau181/Aβ42 on the log scale) are shown as median [IQR]. Group comparisons for normally distributed variables used an independent samples t-test; comparisons for skewed variables used the Mann–Whitney U test. Categorical variables (WMH grade, lacunes, microbleeds) were compared using χ2 or Fisher’s exact test as appropriate. Progressors (≥1 pt) include individuals with any decrease in the K-MMSE score at four-year follow-up, while progressors (≥2 pt) represent a subgroup with a more stringent clinical decline of 2 points or more, analyzed to assess the robustness of findings against measurement variability. Total hippocampal volume z-scores were calculated based on an East Asian normative dataset to adjust for age, sex, and intracranial volume.Aβ42, amyloid-beta 42; G, grade; GFAP, glial fibrillary acidic protein; HC, hippocampal volume; ICV, intracranial volume; IQR, interquartile range; NfL, neurofilament light chain; pTau181, phosphorylated tau at threonine-181; SD, standard deviation; WMH, white matter hyperintensities. Baseline biomarker values between the groups Data are presented as median [IQR] for plasma biomarkers and as mean±SD or n (%) for imaging biomarkers. Continuous variables with approximately normal distributions (intracranial volume, hippocampal volumes, hippocampal volume z-score) are shown as mean±SD; skewed variables (plasma biomarkers pTau181, GFAP, NfL, pTau181/Aβ42 on the log scale) are shown as median [IQR]. Group comparisons for normally distributed variables used an independent samples t-test; comparisons for skewed variables used the Mann–Whitney U test. Categorical variables (WMH grade, lacunes, microbleeds) were compared using χ2 or Fisher’s exact test as appropriate. Progressors (≥1 pt) include individuals with any decrease in the K-MMSE score at four-year follow-up, while progressors (≥2 pt) represent a subgroup with a more stringent clinical decline of 2 points or more, analyzed to assess the robustness of findings against measurement variability. Total hippocampal volume z-scores were calculated based on an East Asian normative dataset to adjust for age, sex, and intracranial volume. Aβ42, amyloid-beta 42; G, grade; GFAP, glial fibrillary acidic protein; HC, hippocampal volume; ICV, intracranial volume; IQR, interquartile range; NfL, neurofilament light chain; pTau181, phosphorylated tau at threonine-181; SD, standard deviation; WMH, white matter hyperintensities. WMH grades differed significantly between converters and non-converters (χ2=6.86; p=0.032), with converters showing higher moderate-to-severe WMHs (G2 42.9%, G3 35.7%) than non-converters (G2 35.3%, G3 19.6%). The Prog-1 group also showed higher WMH grades than the NP-1 group (χ2=8.26; p=0.011), with 79.2% showing moderate-to-severe WMHs (G2 37.5%, G3 41.7%) versus 48.8% in the NP-1 group (G2 36.6%, G3 12.2%). Predictive performance of biomarkers In multivariable logistic regression models adjusted for age and sex, individual plasma biomarkers (pTau181, GFAP, NfL, and pTau181/Aβ42 ratio) did not significantly predict MCI conversion or cognitive progression at either the ≥1-point (Prog-1) or ≥2-point (Prog-2) thresholds (all p>0.05 for ORs) (Table 3). Among plasma biomarkers, NfL showed the highest discriminative estimates; however, it demonstrated statistically significant discrimination for MCI conversion based on AUC analysis, while its overall performance remained modest (MCI conversion: AUC 0.686, p=0.008; Prog-1: AUC 0.624, p=0.075) (Table 3), and these findings were not consistently observed across outcomes or in regression analyses. Table 3Predictive performance of baseline biomarkers for cognitive outcomes in A-SCDBiomarkerOutcomeOR (95% CI)p (OR)AUC (95% CI)p (AUC)CutoffSens/Spec (%)PPV/NPV (%)Plasma biomarkerspTau181MCI conversion1.12 (0.65–1.94)0.6840.591 (0.426–0.756)0.280>22.464.3/41.223.1/80.8Progressor (≥1 pt)1.31 (0.78–2.21)0.3120.604 (0.467–0.741)0.138>26.854.2/65.948.1/71.1Progressor (≥2 pt)1.35 (0.77–2.36)0.2840.640 (0.483–0.798)0.081>26.862.5/67.338.5/84.6GFAPMCI conversion1.05 (0.61–1.82)0.8420.512 (0.345–0.679)0.887>96.457.1/49.023.5/80.6Progressor (≥1 pt)1.14 (0.68–1.92)0.5310.548 (0.395–0.701)0.510>105.550.0/58.541.4/66.7Progressor (≥2 pt)1.18 (0.65–2.12)0.4980.582 (0.412–0.752)0.320>112.456.3/61.232.1/81.1NfLMCI conversion1.48 (0.85–2.58)0.1650.686 (0.548–0.824)0.008>20.864.3/66.734.6/87.2Progressor (≥1 pt)1.25 (0.74–2.10)0.4020.624 (0.487–0.761)0.075>18.266.7/56.147.1/74.2Progressor (≥2 pt)1.62 (0.92–2.85)0.0920.662 (0.523–0.801)0.022>20.168.8/63.337.9/86.1pTau181/Aβ42MCI conversion1.22 (0.71–2.09)0.4720.618 (0.455–0.781)0.160>0.2457.1/66.732.0/85.0Progressor (≥1 pt)1.19 (0.70–2.02)0.5180.584 (0.432–0.736)0.281>0.2258.3/58.545.2/70.6Progressor (≥2 pt)1.28 (0.72–2.28)0.3940.621 (0.458–0.784)0.145>0.2562.5/65.337.0/84.2Imaging biomarkersTotal HCMCI conversion1.62 (0.95–2.75)0.0740.765 (0.548–0.981)0.016≤2,24878.6/72.544.0/92.5Progressor (≥1 pt)4.21 (2.15–8.24)<0.0010.933 (0.797–1.000)<0.001≤2,27487.5/87.880.8/92.3Progressor (≥2 pt)5.84 (2.65–12.80)<0.0010.952 (0.845–1.000)<0.001≤2,24893.8/85.768.2/97.7WMHsMCI conversion2.10 (1.18–3.75)0.0120.835 (0.651–1.000)<0.001≥G264.3/82.450.0/89.4Progressor (≥1 pt)1.85 (1.05–3.28)0.0330.857 (0.685–1.000)<0.001≥G262.5/87.875.0/80.0Progressor (≥2 pt)2.05 (1.12–3.74)0.0190.871 (0.712–1.000)<0.001≥G268.8/83.757.9/89.1Data are presented for three outcomes: “MCI conversion” (diagnostic change from SCD to MCI), “Progressor (≥1 pt)” (any decrease in K-MMSE score at four-year follow-up), and “Progressor (≥2 pt)” (a clinical decline of 2 points or more, analyzed as a sensitivity analysis). To improve clinical interpretability, ORs for continuous biomarkers (pTau181, NfL, and total HC) are reported per one standard deviation increase (for plasma biomarkers) or decrease (for HC volume) to represent clinically meaningful effect sizes, rather than unit-based increments (e.g., 1 mm3 or 1 pg/mL). Logistic regression models were adjusted for age and sex (for plasma biomarkers) or age, sex, and intracranial volume (for imaging biomarkers). All AUC values, 95% CIs, and p-values were calculated using the DeLong method, with p-values representing the comparison against a null hypothesis of AUC=0.5. Optimal cutoff values were identified using the Youden index (Sensitivity + Specificity - 1). For WMHs, Grade 2 or higher (≥G2) was used as the categorical cutoff. p-values indicate statistical significance (p<0.05).Aβ42, amyloid-beta 42; A-SCD, amyloid-negative SCD; AUC, area under the receiver operating characteristic curve; CI, confidence interval; G, grade; GFAP, glial fibrillary acidic protein; HC, hippocampal volume; K-MMSE, Korean Mini-Mental State Examination; MCI, mild cognitive impairment; NfL, neurofilament light chain; NPV, negative predictive value; OR, odds ratio; pTau181, phosphorylated tau at threonine-181; PPV, positive predictive value; SCD, subjective cognitive decline; Sens, sensitivity; Spec, specificity; WMHs, white matter hyperintensities. Predictive performance of baseline biomarkers for cognitive outcomes in A-SCD Data are presented for three outcomes: “MCI conversion” (diagnostic change from SCD to MCI), “Progressor (≥1 pt)” (any decrease in K-MMSE score at four-year follow-up), and “Progressor (≥2 pt)” (a clinical decline of 2 points or more, analyzed as a sensitivity analysis). To improve clinical interpretability, ORs for continuous biomarkers (pTau181, NfL, and total HC) are reported per one standard deviation increase (for plasma biomarkers) or decrease (for HC volume) to represent clinically meaningful effect sizes, rather than unit-based increments (e.g., 1 mm3 or 1 pg/mL). Logistic regression models were adjusted for age and sex (for plasma biomarkers) or age, sex, and intracranial volume (for imaging biomarkers). All AUC values, 95% CIs, and p-values were calculated using the DeLong method, with p-values representing the comparison against a null hypothesis of AUC=0.5. Optimal cutoff values were identified using the Youden index (Sensitivity + Specificity - 1). For WMHs, Grade 2 or higher (≥G2) was used as the categorical cutoff. p-values indicate statistical significance (p<0.05). Aβ42, amyloid-beta 42; A-SCD, amyloid-negative SCD; AUC, area under the receiver operating characteristic curve; CI, confidence interval; G, grade; GFAP, glial fibrillary acidic protein; HC, hippocampal volume; K-MMSE, Korean Mini-Mental State Examination; MCI, mild cognitive impairment; NfL, neurofilament light chain; NPV, negative predictive value; OR, odds ratio; pTau181, phosphorylated tau at threonine-181; PPV, positive predictive value; SCD, subjective cognitive decline; Sens, sensitivity; Spec, specificity; WMHs, white matter hyperintensities. In contrast, neuroimaging markers demonstrated more consistent predictive performance. For hippocampal volume (per SD decrease), the association with MCI conversion did not reach conventional statistical significance in the adjusted logistic regression model (OR 1.62, 95% CI 0.95–2.75, p=0.074). However, its discriminative performance was statistically significant (AUC 0.765, 95% CI 0.548–0.981, p=0.016) (Fig. 1A). For cognitive progression defined as ≥1-point decline (Prog-1), total hippocampal volume was associated with progression (OR 4.21 per SD decrease, 95% CI 2.15–8.24, p<0.001) and demonstrated high discriminative estimates (AUC 0.933, 95% CI 0.797–1.000, p<0.001) (Fig. 1B). When applying the more stringent threshold for cognitive decline (Prog-2, ≥2-point decrease), the estimated effect size of hippocampal volume was larger (OR 5.84, 95% CI 2.65–12.80, p<0.001) and demonstrated the highest AUC estimate among all tested biomarkers (AUC 0.952, 95% CI 0.845–1.000, p<0.001) (Table 3), although this should be interpreted cautiously given the limited sample size. ROC curves for imaging biomarkers for predicting cognitive outcomes. ROC curves illustrate the discriminative performance of total hippocampal volume (red line) and WMH grade (blue line), adjusted for age, sex, and ICV. The shaded areas represent the 95% CIs derived from 1,000 bootstrap iterations. A: Prediction of four-year conversion from SCD to MCI. ROC: converter vs. nonconverter (adjusted). B: Prediction of four-year cognitive decline (≥1-point decrease in K-MMSE score). ROC: progressor vs. nonprogressor (adjusted). AUC values with corresponding 95% CIs are indicated within each panel. AUC, area under the curve; CI, confidence interval; ICV, intracranial volume; K-MMSE, Korean Mini-Mental State Examination, 2nd edition; MCI, mild cognitive impairment; ROC, receiver operating characteristic; SCD, subjective cognitive decline; WMHs, white matter hyperintensities. WMH grade (≥G2) also was significantly associated with MCI conversion (OR 2.10, 95% CI 1.18–3.75, p=0.012) and cognitive progression at both thresholds (Prog-1: OR 1.85, p=0.033; Prog-2: OR 2.05, p=0.019). While WMH grade demonstrated significant discriminative performance (AUC 0.835–0.871, all p<0.001), its effect sizes and AUC estimates were relatively lower than those of hippocampal volume across all cognitive outcomes (Table 3). The hippocampal volume models showed good calibration, with non-significant Hosmer–Lemeshow tests (p=0.74 for MCI conversion; p=0.61 for Prog-1) and favorable Brier scores (0.15 and 0.11, respectively). Direct comparisons of AUCs using bootstrap resampling (1,000 iterations) were performed to compare the relative prognostic performance of imaging markers and plasma biomarkers using a resampling-based approach (Fig. 2). This analysis showed that neuroimaging markers generally demonstrated higher point estimates for AUC compared to all plasma biomarkers. Hippocampal volume exhibited the most pronounced ΔAUC relative to individual plasma biomarkers. For cognitive progression (Prog-1), hippocampal volume showed numerically higher AUC estimates than all individual plasma biomarkers (all p<0.05), with ΔAUC values ranging from 0.177 to 0.248 (Fig. 2A). Similarly, WMH grade showed higher discriminative estimates than GFAP and NfL for Prog-1 (Fig. 2B), although its superiority over pTau181 and pTau181/Aβ42 reached only a non-significant trend. For MCI conversion, hippocampal volume exhibited higher AUCs than pTau181, NfL, and the pTau181/Aβ42 ratio for predicting MCI conversion. It should be noted that Table 3 presents the absolute AUC values for each biomarker, whereas Fig. 2 illustrates the bootstrap-derived ΔAUC and its 95% CI to provide a more resampling-based comparison of the differences in performance. These comparisons are based on internal bootstrap estimates within a limited sample. Bootstrap-based comparisons of prognostic performance between imaging and plasma biomarkers using standardized variables. ΔAUC were estimated using bootstrap resampling (1,000 iterations) at the participant level, with re-estimation of the logistic regression model and recalculation of the AUC in each resample. Prior to analysis, all continuous imaging and plasma biomarkers were standardized (z-scored) to ensure methodological harmonization and comparability of effect sizes. For each outcome, ΔAUC was calculated as (AUC of imaging biomarker model) – (AUC of plasma biomarker model), and error bars indicate 95% CIs derived from the bootstrap distribution. A positive value indicates higher AUC estimates for the imaging marker compared with the plasma biomarker. Red dots and lines represent MCI conversion while blue dots and lines represent cognitive progression (Prog-1). A: Comparisons between hippocampal volume and plasma biomarkers. B: Comparisons between WMH grade and plasma biomarkers. A ΔAUC was considered statistically significant if the 95% CIs did not include zero. This bootstrap-based approach was employed to assess the robustness of comparative AUC estimates, but does not mitigate limitations related to the small sample size and limited number of events. While Table 3 reports the absolute adjusted AUC values using DeLong’s method for each biomarker model, whereas Fig. 2 presents pairwise ΔAUCs derived from the bootstrap distribution; these values represent distinct statistical estimations and should not be directly compared numerically. All models were adjusted for age and sex, with neuroimaging models additionally adjusted for intracranial volume. *p<0.05. Aβ, amyloid-beta; ΔAUC, difference in area under the curve; AUC, area under the curve; CI, confidence interval; GFAP, glial fibrillary acidic protein; MCI, mild cognitive impairment; NfL, neurofilament light; pTau181, phosphorylated tau at threonine-181; SCD, subjective cognitive decline; WMHs, white matter hyperintensities. In sensitivity analyses with sequential adjustment for education, baseline K-MMSE, and vascular risk burden, the overall patterns of association remained consistent across both plasma and imaging biomarkers, with no material changes in effect estimates or statistical significance (Supplementary Table 2 in the online-only Data Supplement). DISCUSSION In this four-year longitudinal study of individuals with amyloid PET-negative SCD, hippocampal volume showed consistent prognostic performance across outcomes, with higher discriminative estimates, than plasma-based biomarkers in terms of predictive efficacy. These comparative findings based on bootstrap-derived ΔAUC should be interpreted cautiously given the limited sample size and exploratory nature of the analysis. The results suggest that reduced baseline hippocampal volume was associated with global cognitive decline and the transition from SCD to MCI, suggesting its potential relevance as a risk marker across the amyloid-tau-neurodegeneration (ATN) spectrum.11 Smaller hippocampal volume at baseline was associated with higher odds of cognitive decline (OR 4.21 per SD decrease for Prog-1 and OR 5.84 for Prog-2) (Table 3). This finding supports the role of neurodegeneration markers as prognostic markers of subclinical decline in amyloid-negative cohorts. Consistent with the ATN framework, neurodegeneration without amyloid pathology (A–T+N+) has been linked to faster clinical progression compared to A–T–N–profiles.6,7,12 This suggests that neurodegenerative processes leading to hippocampal atrophy, independent of Aβ accumulation, contribute to cognitive decline in patients with A-SCD. While we could not histologically confirm specific etiologies, in the context of hippocampal atrophy, various non-AD tauopathies, such as PART and LATE, have been known to target CA1 pyramidal neurons and the subiculum, causing synaptic and dendritic spine loss before cell death.13 The observed hippocampal atrophy in our study may reflect heterogeneous non-AD-type neurofibrillary changes or TAR DNA-binding protein 43 (TDP-43)-related hippocampal pathology, rather than a specific underlying pathology. Following the LATE consensus criteria,22 such TDP-43-related hippocampal pathology may underlie atrophy independent of AD-type neurofibrillary changes. Mild PART pathology correlates with hippocampal shrinkage on MRI,13,35 while microglial activation and complement-mediated synaptic pruning further accelerate volume loss with aging.7 Quantitative MRI volumetry therefore may capture a downstream manifestation of heterogeneous non-AD neurodegenerative processes, rather than indicating a specific underlying etiology, which may not be fully reflected by blood-based amyloid markers.12 Our findings are consistent with recent large-scale plasma biomarker comparisons showing limited discriminative utility of AD-specific plasma markers in amyloid-negative or mixed-etiology populations.14,15 These results collectively suggest that while ultrasensitive blood-based assays provide accessible biomarker quantification, their prognostic performance remains suboptimal in specific non-AD neurodegenerative trajectories where amyloid pathology is absent. Consequently, quantitative MRI measurements of hippocampal volume may serve as a useful indicator of disease progression in A-SCD, with potential clinical utility for risk stratification. Regarding vascular contributions, WMH grade (≥G2) also demonstrated significant associations for both MCI conversion and cognitive decline across all models (p<0.05) (Table 3). While its discriminative performance (AUC 0.835–0.871) was statistically significant and showed higher discriminative performance than several plasma biomarkers, the magnitude of the effect was relatively more modest compared to that of hippocampal volume. Although WMHs indicate cerebral small vessel disease and cognitive impairment,12,13 their limited prognostic value relative to hippocampal atrophy may be related to short follow-up periods or visual WMHs grading underestimating subtle leukoaraiosis. Recent research suggests that automated WMHs quantification correlates better with executive dysfunction than visual grading.36 The effect of WMHs on cognition can be complex, interacting with other neurodegenerative processes or modified by vascular risk factors.34 Baseline plasma pTau181, GFAP, NfL, and pTau181/Aβ42 ratios did not show consistent prognostic associations for four-year clinical conversion to MCI or K-MMSE decline within this A-SCD population. This finding is consistent with a recent multicohort longitudinal study by Abbas et al. (2025),37 which reported that plasma GFAP did not increase significantly over time and showed no association with cognitive decline in amyloid-negative cognitively unimpaired individuals. In that study, based on longitudinal data from multiple well-characterized cohorts, plasma GFAP remained stable across follow-up and was unrelated to cognition in the amyloid-negative group. Together with our results, these convergent findings indicate that astrocytic activation reflected by plasma GFAP may occur independently of AD-specific amyloid pathology and may not serve as a reliable prognostic biomarker in amyloid-negative trajectories. This aligns with recent neuropathological and plasma biomarker studies, indicating that current AD-specific plasma biomarkers, particularly pTau181, are highly specific for amyloid-β-driven tau phosphorylation3,15,18 and have limited utility in non-AD etiologies.16,17,38,39 These results may reflect the biological heterogeneity among A- individuals and the limited sensitivity of current plasma assays for non-AD neurodegeneration. While subgroups of SCD with elevated GFAP and NfL levels may be at higher risk,40,41 these AD-focused plasma biomarkers lack the sensitivity to detect subtle molecular changes in etiologically heterogeneous A-SCD in single time-point measurements.2,5,7,42,43,44 Our visual reading approach for amyloid negativity was validated using quantitative centiloid analysis. Only one visually A- participant had a CL exceeding 25 (26.6) and remained cognitively stable without conversion. When amyloid negativity was redefined as CL <25, the findings remained consistent, suggesting that visual classification largely agrees with the quantitative assessment. This aligns with prior work showing that standardized visual interpretation corresponds to quantitative thresholds45 and validation studies proposing centiloid cutoff ranges in the 25–35 CL interval.31 Our study had several strengths. This longitudinal analysis provides novel prognostic data from a well-characterized amyloid PET-negative SCD cohort over four years, addressing a critical gap in the literature, as most biomarker studies have predominantly focused on amyloid-positive populations. We assessed cognitive trajectories using both categorical (MCI conversion) and continuous measures of global cognition, enabling the detection of subtle declines during the medium-term follow-up. However, this study had some limitations. First, the limited sample size (n=65) and small number of conversion events (n=14) may have constrained the statistical power. This low event-per-variable ratio may have contributed to model instability and wider CIs, which necessitates caution in generalizing these results. Although we employed bootstrap resampling to assess the robustness of the estimates, this approach does not overcome the limitations related to the small sample size or potential model instability, and the results regarding MCI conversion should be interpreted as exploratory, and further validation in larger cohorts is required. Second, we relied on single-time-point baseline measurements for plasma and imaging biomarkers, potentially missing dynamic changes that could enhance prognostic accuracy. Third, the use of K-MMSE as the sole measure for cognitive progression and the dichotomization of score changes may be subject to a ceiling effect and regression to the mean, as evidenced by the significantly higher baseline scores in the progressor groups. To address the potential measurement variability of a 1-point decline, we performed a sensitivity analysis using a ≥2-point threshold (Prog-2), which yielded consistent results and partially mitigated this limitation. Additionally, allowing verbal memory performance down to the 7th percentile at baseline may have enriched the cohort with individuals at relatively higher risk of progression. Fourth, we did not subclassify non-AD pathologies, such as LATE or PART, which may have contributed to hippocampal atrophy without amyloidosis. Therefore, the underlying etiologies of neurodegeneration could not be determined in this study, and our interpretation regarding these specific etiologies remains speculative without pathological confirmation. Fifth, our analyses were limited to plasma pTau181 levels measured using the Simoa HD-X platform. Evidence suggests that pTau217, particularly when quantified by mass spectrometry or Lumipulse platforms, shows greater sensitivity in cognitively unimpaired individuals.46,47 Finally, tau pathology was not directly measured using tau-PET or plasma MTBR-tau243. Although plasma pTau181 was included, it mainly reflects AD-specific, Aβ-driven tau phosphorylation and may underrepresent non-AD tauopathies.16,17 ATN-based studies have shown that tau pathology in A- individuals significantly influences clinical progression compared to that in A-T- individuals.48 This should be explored in future studies using larger, multi-ethnic cohorts with multi-modal tau assessments. In conclusion, in this specific cohort of amyloid PET-negative individuals with SCD, hippocampal atrophy demonstrated consistent predictive performance among the tested markers, showing higher and more consistent discriminative performance than AD-focused plasma biomarkers (pTau181, GFAP, NfL, and pTau181/Aβ42 ratio) in forecasting four-year cognitive outcomes, although these findings should be interpreted cautiously. These findings suggest that MRI-based neurodegeneration markers may have a role in risk stratification in A- trajectories. Future studies with larger, multi-center cohorts and long-term follow-up are warranted to validate these findings and to explore the integration of multi-modal biomarkers to further refine prognostic accuracy in cognitive decline. Acknowledgements The authors thank all participants and staff of the CoSCo study for their contributions to this research. Footnotes Author Contributions: Conceptualization: Bora Yoon, Dong Won Yang.Data curation: all authors.Formal analysis: Bora Yoon.Funding acquisition: Dong Won Yang.Investigation: Yun Jeong Hong, Kee Hyung Park, SangYun Kim, Min Jeong Wang, Seong Hye Choi, Jee Hyang Jeong, Dong Won Yang.Methodology: Bora Yoon, Dong Won Yang.Project administration: Dong Won Yang.Resources: Yun Jeong Hong, Kee Hyung Park, SangYun Kim, Min Jeong Wang, Seong Hye Choi, Sungmin Kang, Jee Hyang Jeong, Dong Won Yang.Supervision: Jee Hyang Jeong, Dong Won Yang.Visualization: Bora Yoon.Writing—original draft: Bora Yoon.Writing—review & editing: Bora Yoon, Hyuk-je Lee, Yun Jeong Hong, Kee Hyung Park, SangYun Kim, Min Jeong Wang, Seong Hye Choi, Jee Hyang Jeong, Dong Won Yang. Conceptualization: Bora Yoon, Dong Won Yang. Conceptualization: Bora Yoon, Dong Won Yang. Data curation: all authors. Data curation: all authors. Formal analysis: Bora Yoon. Formal analysis: Bora Yoon. Funding acquisition: Dong Won Yang. Funding acquisition: Dong Won Yang. Investigation: Yun Jeong Hong, Kee Hyung Park, SangYun Kim, Min Jeong Wang, Seong Hye Choi, Jee Hyang Jeong, Dong Won Yang. Investigation: Yun Jeong Hong, Kee Hyung Park, SangYun Kim, Min Jeong Wang, Seong Hye Choi, Jee Hyang Jeong, Dong Won Yang. Methodology: Bora Yoon, Dong Won Yang. Methodology: Bora Yoon, Dong Won Yang. Project administration: Dong Won Yang. Project administration: Dong Won Yang. Resources: Yun Jeong Hong, Kee Hyung Park, SangYun Kim, Min Jeong Wang, Seong Hye Choi, Sungmin Kang, Jee Hyang Jeong, Dong Won Yang. Resources: Yun Jeong Hong, Kee Hyung Park, SangYun Kim, Min Jeong Wang, Seong Hye Choi, Sungmin Kang, Jee Hyang Jeong, Dong Won Yang. Supervision: Jee Hyang Jeong, Dong Won Yang. Supervision: Jee Hyang Jeong, Dong Won Yang. Visualization: Bora Yoon. Visualization: Bora Yoon. Writing—original draft: Bora Yoon. Writing—original draft: Bora Yoon. Writing—review & editing: Bora Yoon, Hyuk-je Lee, Yun Jeong Hong, Kee Hyung Park, SangYun Kim, Min Jeong Wang, Seong Hye Choi, Jee Hyang Jeong, Dong Won Yang. Writing—review & editing: Bora Yoon, Hyuk-je Lee, Yun Jeong Hong, Kee Hyung Park, SangYun Kim, Min Jeong Wang, Seong Hye Choi, Jee Hyang Jeong, Dong Won Yang. Conflicts of Interest: The authors have no potential conflicts of interest to disclose. Funding Statement: This study was supported by the Korea Institute for Advancement of Technology (KIAT) grant funded by the Ministry of Trade, Industry, and Energy, Republic of Korea (P0018061, Industry Innovation Foundation Building Program) and the Ministry of Health and Welfare, Republic of Korea (HI18C0530). The funders had no role in the study design, data collection, data analysis, interpretation, manuscript writing, or decision to submit the article for publication. Availability of Data and Material Supporting data for this study are available upon reasonable request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions. Supplementary Materials The online-only Data Supplement is available with this article at https://doi.org/10.3988/jcn.2025.0659. Supplementary Table 1 Comparison of completers and dropouts Supplementary Table 2 Sensitivity logistic regression analyses with additional covariate adjustment References Associated Data Supplementary Materials Supplementary Table 1 Comparison of completers and dropouts Supplementary Table 2 Sensitivity logistic regression analyses with additional covariate adjustment Data Availability Statement Supporting data for this study are available upon reasonable request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.