Background

Alzheimer’s disease (AD) is a neurodegenerative disorder that primarily affects older individuals and is characterized by progressive cognitive decline, memory loss, and functional impairment [1]. The pathogenesis of AD is multifactorial, involving complex interactions between genetic, environmental, and lifestyle factors [2]. Currently, there is no cure for AD, and available treatments primarily aim to alleviate symptoms and slow disease progression [3]. Mild cognitive impairment (MCI), a critical precursor of AD, progresses to dementia at a rate of approximately 10–15% annually [4,5,6]. Therefore, identifying MCI patients at elevated risk of conversion and intervening early are critical clinical priorities [7].

Genetic factors play a central role in AD susceptibility. Genome-wide association studies (GWAS) have identified numerous single nucleotide polymorphisms (SNPs) associated with AD risk [8,9,10]. While many AD risk loci are shared across populations, the effect sizes of specific alleles differ by ancestry. For example, APOE ε4 exerts a weaker effect in East Asians compared to Europeans [11]. Large multi-ancestry GWAS have demonstrated convergence on common pathways, including lipid metabolism and synaptic integrity [12]. Recent reviews further highlight that ancestry-specific allele frequencies and linkage disequilibrium (LD) structures critically influence the portability and interpretability of polygenic and machine learning (ML)-based risk models [13, 14]. East Asian populations, in particular, tend to have longer LD blocks, which may enhance tagging efficiency of risk alleles but limit fine-mapping resolution [15]. These factors underscore the need for ancestry-specific prediction models, particularly in underrepresented populations such as Koreans.

Machine learning has shown growing promise in genetic risk prediction for complex traits and neurological disorders. Systematic reviews have demonstrated its application to psychiatric conditions [16, 17], neurological diseases [18], and AD [19], where ML-based extensions of polygenic risk scoring have improved prediction beyond conventional frameworks [20]. Best-practice recommendations further emphasize that threshold-independent metrics (AUC, PR-AUC) and calibration measures should accompany accuracy to ensure robust interpretability [21].

Prior studies in European cohorts have demonstrated that polygenic approaches can inform prediction of MCI progression. Desikan et al. [22] developed and validated a polygenic hazard score (PHS) that integrated APOE and additional SNPs, showing strong associations with age-specific AD risk and MCI-to-AD conversion. Tan et al. [23] demonstrated that PHS was enriched for AD-related amyloid and tau deposition and correlated with accelerated cognitive decline. More recently, Rowe et al. [24] systematically reviewed ML-based genetic prediction approaches for AD, concluding that APOE remains the strongest predictor, while polygenic or PRS-based models provide additional but modest improvements. However, most of these studies have been conducted in European or North American cohorts. Importantly, Lewis & Vassos [25] highlighted that PRS derived from European datasets often underperform in non-European populations due to allele frequency and LD differences, underscoring the urgent need for ancestry-diverse research.

In this context, we aimed to develop and evaluate predictive models for dementia risk in Korean patients with MCI using SNP chip data and ML algorithms. Using data from the Biobank Innovations for Chronic Cerebrovascular Disease with Alzheimer’s Disease Study (BICWALZS), we conducted GWAS to identify dementia-associated SNPs and constructed three predictive models: Model 1 (AD + Vascular dementia [VD] vs. subjective cognitive decline [SCD]), Model 2 (AD vs. SCD), and Model 3 (union of both sets). By systematically comparing multiple ML classifiers, applying strict cross-validation and calibration procedures, and validating predictions against longitudinal follow-up outcomes, our study extends existing literature by addressing an underrepresented ancestry group and provides proof-of-concept evidence for the feasibility of SNP-based risk stratification in Korean MCI populations.

Methods

Study population

This retrospective cohort study used data from the Biobank Innovations for Chronic Cerebrovascular Disease with Alzheimer’s Disease Study (BICWALZS), a component of the national biobanking program (Korea Biobank Project) led by the Korea Disease Control and Prevention Agency [26]. BICWALZS, launched in October 2016, involved five major institutions: Ajou University Hospital, Samsung Medical Center, Inha University Hospital, Pusan National University Hospital, and Chonnam National University Hospital. A centralized coordination system ensured standardized specimen handling, adherence to best practices, and compliance with ethical and legal regulations across sites [26].

Between October 2016 and December 2020, 1,013 participants underwent comprehensive clinical evaluations including blood tests and standardized neuropsychological assessments. Within this cohort, subsets additionally received brain magnetic resonance imaging (MRI) (n = 817), brain amyloid positron emission tomography (PET) (n = 713), SNP microarray analysis (n = 949), actigraphy measurements (n = 200), and patient-derived dermal fibroblast sampling (n = 175). Annual follow-up was conducted for all participants; extended biannual evaluations, including neuropsychological testing, actigraphy, and brain MRI, were performed for individuals with cortical amyloid accumulation, subcortical vascular pathology, the APOE ε4 allele, or significant cognitive decline [26].

All 1,013 participants completed assessments for cognitive function and daily living abilities, including the Seoul Neuropsychological Screening Battery, 2nd Edition (SNSB-II), the Korean Mini-Mental State Examination, the Korean version of the Short Form of the Geriatric Depression Scale, the Montgomery-Åsberg Depression Rating Scale, the Barthel Activity of Daily Living Score, and the Seoul Instrumental Activity of Daily Living Score. The SNSB-II assesses multiple cognitive domains, including language, visuospatial ability, memory, and frontal/executive functions [27]. Diagnostic classifications followed standard criteria: subjective cognitive decline (SCD) as self- or informant-reported cognitive decline without objective cognitive impairment [28]; MCI per the expanded Mayo Clinic criteria [6, 29]; AD per the National Institute on Aging and Alzheimer’s Association (NIA-AA) criteria [30]; and VD per the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) [31, 32].

For the neuroimaging-based categorization, a subset was classified as cognitively normal (CN), Alzheimer’s disease-related cognitive impairment (ADCI), or vascular cognitive impairment (VCI) based on Mini-Mental State Examination (MMSE) scores and two key neuroimaging biomarkers: cortical amyloid burden and subcortical vascular burden. Cognitive decline was identified based on age-, sex-, and education-adjusted z-scores (≤ −1). Individuals with amyloid pathology on PET scans or white-matter hyperintensities due to radiation injury, multiple sclerosis, vasculitis, or leukodystrophy were excluded from the CN and VCI groups [26].

For association analyses, cognitive status was available for 692 participants with confirmed Korean Biobank Array (KoreanChip) genotypes. Participants were classified as having SCD (n = 81), MCI (n = 389), AD (n = 144), or VD (n = 60). Eighteen participants with incompatible diagnostic labels were excluded, yielding a final analytic cohort of 674. Among those with MCI, 80 had 2-year follow-up; of these, 61 had both baseline genotypes and 2-year outcomes and were used for temporal validation.

Statistics

Analyses of variance (ANOVA) and chi-square tests compared demographic, clinical, cognitive, and genotypic characteristics across cognitive group (CN, MCI, and dementia). For variables with significant overall differences, Tukey’s test and Fisher’s exact test were used for post-hoc comparisons. Analyses were performed in R (v4.4.3) [33].

Genome-wide association study

SNP genotyping used the Korea Biobank Array (Axiom KORV1.1–96; Thermo Fisher Scientific) at DNA Link Inc. (Seoul, Republic of Korea) [34]. Marker-level QC employed SNPolisher (poly high resolution, mono high resolution, and absence of minor homozygosity). In addition, SNPs with a call rate < 95% or a Hardy-Weinberg equilibrium test p-value < 10⁻⁶ were excluded. Sample-level QC included call rate ≥ 97%, heterozygosity within ± 2 SD, multidimensional scaling within ± 0.02 of reference, singleton count ≤ 15, identity-by-state checks, and sex mismatch detection. After QC, 769,894 SNPs from 949 participants remained [26].

Genotypes were imputed using the Korean Imputation Service (KIS, https://coda.nih.go.kr/frt/index.do) with the Phase 1 Korean reference panel (n = 4,799; GRCh37/hg19) [35]. Phasing used SHAPEIT4 (v4.2.2) [36], and imputation used Minimac4 (v1.0.2) [37, 38]. Post-imputation QC in PLINK v1.9 required per-sample call rate ≥ 98%, per-variant missingness ≤ 2%, minor allele frequency (MAF) ≥ 0.01, and Hardy–Weinberg equilibrium p ≥ 1 × 10⁻⁶. Sites with allele mismatches, duplicates, or monomorphism were removed, and strand harmonization (allele switches/strand flips) was confirmed during imputation QC. The merged dataset comprised 9,681,273 SNPs across 692 individuals with a genotyping rate of 1.00. Chromosome-wise imputation QC indicated 88.98% mean reference overlap, with 49.76% of input sites retained after excluding mismatched/low-quality variants.

SNP functional annotation

Following GWAS analysis, comprehensive functional annotation of identified SNPs was performed using SNPnexus (https://www.snp-nexus.org/v4/) based on the hg19 genome build. The annotation pipeline included multiple analytical components: (1) functional region mapping and predicted functional effects based on genomic location, (2) SIFT (Sorting Intolerant From Tolerant) analysis for missense variant impact prediction, (3) eQTL (expression quantitative trait loci) data integration to assess regulatory effects on gene expression, (4) cross-referencing with GWAS Catalog for previously reported disease associations, and (5) validation against the Genetic Association Database (GAD) for comprehensive disease-associated SNP classification. This multi-layered approach enabled systematic categorization of SNPs across various disease classes and identification of potential pleiotropic effects.

Subtyping MCI using AD-associated SNP loci

We constructed predictive matrices from dementia-associated SNPs identified by GWAS. Logistic regression association tests in PLINK (v1.9) [39] adjusted for age, sex, and the top 10 principal components (PC1–PC10). We reported odds ratios (ORs) with 95% confidence intervals (CIs), Wald test p-values, FDR-adjusted q-values, and case/control minor allele frequencies (MAFs). FDR (Benjamini–Hochberg) was applied genome-wide within each comparison (AD vs. SCD; AD + VD vs. SCD); tabled q-values therefore reflect genome-wide multiplicity control. A suggestive threshold (p < 1 × 10⁻⁵) was used only for ML feature selection, not for formal discovery. SNPs meeting this threshold were defined once per case–control subset, and then fixed as predictor panels (54, 60, and 76 SNPs).

Within cross-validation, all model training, probability calibration (Platt scaling), and threshold selection (Youden’s J, F1-max) were confined to training folds and applied unchanged to held-out folds to avoid leakage. Given the modest sample size, fold-level GWAS within cross-validation (CV) folds was not pursued due to instability; instead, we combined fixed panels with nested CV and temporal validation to balance rigor and stability.

Based on p < 1 × 10⁻⁵, we constructed three models: Model 1 (54 SNPs; AD + VD vs. SCD), Model 2 (60 SNPs; AD vs. SCD), and Model 3 (76 SNPs; union) (Table 2). No SNPs were retained for SCD vs. VD; this subset was not modeled. Final matrices were 285 × 54 (Model 1; cases = 204, controls = 81), 225 × 60 (Model 2; cases = 144, controls = 81), and 225 × 76 (Model 3), with a sensitivity analysis using 285 × 76. Genotypes were hard calls (0/1/2) with no missingness. LD clumping (r²<0.1 within 250 kb) estimated approximately 11, 4, and 14 loci for Models 1–3, respectively (Supplementary Table S2); most signals mapped to the APOE–PVRL2–TOMM40 block, with additional non-APOE tagging variants on chromosomes 3, 6, 9, 10, 14, and 20.

Machine learning algorithms

SNPs meeting the suggestive threshold were used as features for six supervised learning algorithms: RF, KNN, ANN, SVM, and XGBoost, and LightGBM. These algorithms were selected because of their complementary strengths in handling high-dimensional genetic data and their proven effectiveness in disease risk prediction [12]. Given moderate class imbalance, all CV was stratified. We used nested CV: inner 5-fold for hyperparameter tuning (AUC-maximizing), outer 10 × 10 repeated stratified CV for unbiased estimation. Grids were:

  • RF: n_estimators = 200; max_depth ∈ {0,15}; min_samples_leaf = 1.

  • KNN: k ∈ {3,5,7}; distance = Euclidean; weighting = 1/distance.

  • SVM: Linear kernel (C ∈ {1,10}); RBF kernel (γ ∈ {0.005,0.01}, C = 10).

  • ANN: Hidden layers = i or i/2; learning_rate = 0.1; momentum = 0.3; epochs = 300.

  • XGBoost: n_estimators ∈ {200,300,500}; max_depth ∈ {3,4,5}; learning_rate ∈ {0.03,0.05,0.10}; subsample ∈ {0.8,1.0}; colsample_bytree ∈ {0.8,1.0}; reg_lambda ∈ {0.5,1.0}.

  • LightGBM: n_estimators ∈ {200,300,500}; num_leaves ∈ {15,31,63}; learning_rate ∈ {0.03,0.05,0.10}; subsample ∈ {0.8,1.0}; colsample_bytree ∈ {0.8,1.0}; reg_lambda ∈ {0.0,1.0}.

We summarized performance primarily with AUC and PR-AUC, alongside balanced accuracy, sensitivity, and specificity. Probabilities were calibrated with Platt scaling; calibration quality was quantified with Brier scores and visualized with calibration plots (Supplementary Table S3; Supplementary Fig. S2–S5). Operating thresholds (Youden’s J, F1-max) were selected within training folds and then applied unchanged to held-out folds and to the temporal validation cohort.

Risk assessment of predicted groups using two-year follow-up data

For a prospective check of generalization, we applied pre-specified thresholds to the 2-year follow-up subset (n = 61 with genotypes; 14 converters). For each algorithm and SNP panel, we report sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) with 95% CIs (Supplementary Table S4), and event enrichment across risk strata (Supplementary Table S5). Overall discrimination was modest (AUROC approximately 0.45–0.55), consistent with the limited number of events. Youden’s J yielded more balanced sensitivity/specificity, whereas F1-max increased sensitivity to converters (approximately 0.86–0.93) at the expense of specificity (approximately 0.19–0.30). Despite wide CIs, event rates were generally higher in predicted high-risk groups, supporting a genuine—albeit modest—predictive signal.

Results

Baseline characteristics of study participants

The study included 674 individuals categorized into three groups: cognitively normal (CN, n = 81), MCI (n = 389), and dementia (n = 204) (Table 1). The average age of the participants was 72.88 years (standard deviation = 7.33), with patients with dementia being the oldest (74.19 years) followed by those with MCI (72.67 years), and the CN (70.59 years). Females were predominant (65.73%) across all groups, although the proportion was lower in the dementia group (59.8%) compared to the MCI (67.61%) and CN (71.6%) groups. In terms of vascular health indicators, systolic and diastolic blood pressures were higher in the CN group than in the MCI and dementia groups, which may reflect differences in vascular health potentially associated with cognitive impairment progression. Body mass index (BMI) was highest in the CN group (24.76 kg/m²) and lowest in the dementia group (23.3 kg/m²), with statistically significant differences observed between the CN and dementia groups, as well as between the MCI and dementia groups (both p < 0.05). This finding is consistent with weight loss commonly observed in dementia, possibly due to underlying metabolic or nutritional factors [40]. Regarding lipid and metabolic profiles, total cholesterol and low-density lipoprotein levels were highest in the dementia group, whereas high-density lipoprotein levels were the lowest, suggesting a possible association between lipid dysregulation and dementia risk. Altered lipid metabolism has been implicated in amyloid deposition and neuroinflammation, which contributes to the pathogenesis of AD [41]. Notably, glucose levels were also highest in patients with dementia (111.03 mg/dL), which is consistent with existing evidence linking impaired glucose metabolism and cognitive decline [42]. Insulin resistance and hyperglycemia are associated with an increased risk of AD, possibly through mechanisms involving oxidative stress, neuroinflammation, and tau hyperphosphorylation [43, 44]. Cognitive function assessments revealed expected patterns: MMSE scores were highest in the CN group (26.96), followed by those with MCI (24.42), and lowest in patients with dementia (18.91), with all group differences being statistically significant (p < 0.05). The Clinical Dementia Rating and Geriatric Depression Scale scores also showed significant trends corresponding to cognitive status, with patients with dementia exhibiting the most severe impairment. These findings reinforce the progressive nature of cognitive decline and its impact on functional independence and mental health [45, 46]. Genotypic analysis of APOE revealed a significantly higher prevalence of ε4 carriers (E3/E4 or E4/E4) in patients with dementia (43.63%) than in those with MCI (26.48%) and CN (13.58%) (p < 0.001). Conversely, ε4 non-carriers were most common in the CN group (86.42%) and least common in patients with dementia (56.37%), consistent with the well-established role of APOE ε4 as a genetic risk factor for AD. The presence of APOE ε4 increases amyloid deposition, disrupts lipid homeostasis, and contributes to synaptic dysfunction, thereby accelerating the pathophysiological processes underlying AD [47, 48]. These findings highlight key demographic, metabolic, and genetic differences among cognitive groups, which may contribute to cognitive decline and inform early dementia risk assessment. Understanding these baseline characteristics is essential for interpreting predictive modeling outcomes. The interplay between vascular health, lipid metabolism, glucose regulation, and genetic susceptibility underscores the multifactorial nature of dementia pathogenesis and supports the integration of these factors into predictive models for more accurate and personalized risk stratification [49].

Table 1 Baseline characteristics of the study participants
Full size table

Identification of SNP loci associated with AD

Figure 1 illustrates the pipeline used to predict AD using an artificial intelligence model based on SNP chip data. First, SNP chip data from the BICWALZS dataset, which comprised 674 individuals, were analyzed, and dementia-associated SNP loci were selected based on the analysis results. The selected SNPs were used to construct a predictive matrix for 285 individuals, including those with SCD, AD, and VD. This SNP matrix was used to develop and train the predictive models. The accuracy of the models was evaluated using 10-fold cross-validation. To validate the models’ predictions, 2-year follow-up data were used to assess whether the cases that progressed from MCI to AD or VD aligned with the predictions. This pipeline provides a structured approach to leverage genetic data to predict dementia risk and validate these predictions over time. The GWAS-based SNP selection identified 54 SNPs for AD + VD vs. SCD (Model 1), 60 SNPs for AD vs. SCD (Model 2), and 76 SNPs for the combined union set (Model 3). These SNPs were selected based on their statistical significance in association with dementia, as determined by p-values below a predefined threshold (P < 10−5), and were subsequently used to construct the genetic input feature for dementia prediction models (Table 2). As expected under conservative genome-wide BH correction and a modest sample size, q-values were uniformly high. LD clumping revealed that these corresponded to 4, 11, and 14 approximately independent loci, respectively. The APOE locus accounted for the strongest and most consistent associations, but additional non-APOE loci were also present (e.g., chromosomes 3, 6, 9, 10, 14, and 20), reflecting broader polygenic contributions (Supplementary Table S2).

Fig. 1
Fig. 1
Full size image

Prediction pipeline for Alzheimer’s disease (AD) using artificial intelligence models based on SNP chip data. At baseline, SNP data from the BICWALZS dataset (N = 674: SCD = 81, MCI = 389, AD = 144, VD = 60) were analyzed to identify dementia-associated loci. Three predictive SNP panels (54, 60, and 76 SNPs) were constructed and used to generate genotype matrices for model training. Models were developed in case–control subsets and evaluated with repeated 10 × 10 cross-validation for performance stability, calibration, and threshold-specific metrics. Temporal validation subset: from baseline MCI (n = 389), n = 61 had genotype + 2-year outcomes (14 AD converters/47 non-converters) and were used to assess prospective performance

Table 2 Association results for SNPs with suggestive evidence (p < 1×10⁻⁵)
Full size table

Functional annotation of AD-associated SNPs reveals pleiotropic links to multiple disease categories

We elucidated the biological functions of the SNPs identified through the GWAS analysis of the BICWALZS dataset through a comprehensive annotation using SNPnexus (https://www.snp-nexus.org/v4/) [50]. The annotation provided crucial information, including dbSNP identifiers, nearby genes, predicted functional implications based on genomic location, and previously reported biological functions from the Genetic Association Databases and GWAS databases (Supplementary Table 1) [51, 52]. Figure 2 shows the disease category associations of the identified SNPs. The UpSet plot revealed that the largest number of SNPs were associated with metabolic disorders (n = 60). As shown at the bottom of the graph, two specific SNPs (rs769449 and rs429358) were present across multiple disease categories along with other SNPs. Rs429358 represents the well-characterized APOE ε4-associated variant (p.Cys130Arg), which is predicted as deleterious by SIFT analysis, while rs769449 is an intronic variant within the APOE gene that may contribute to gene regulation through enhancer elements or other regulatory mechanisms. Both variants demonstrate the expected pleiotropic effects consistent with APOE’s central role in multiple pathological processes (Supplementary Table S2). Notably, substantial intersections were observed among various combinations of disease categories, suggesting potential genetic pleiotropy of AD-associated SNPs. The analysis showed that the identified SNPs were associated with multiple disease categories, with particularly strong associations with metabolic, cardiovascular, and neurological conditions. These results indicate associations of genetic variants with multiple disease, and suggest a shared genetic architecture between AD and these disease groups.

Fig. 2
Fig. 2
Full size image

Functional annotation of AD-SNP loci reveals overlapping disease associations. The UpSet plot illustrates the intersection sizes (top) of different disease categories (left) associated with the annotated SNPs. The x-axis shows combinations of disease categories, while the y-axis indicates the number of SNPs in each intersection. Blue dots connected by lines denote the combinations of categories, and the height of each bar reflects the corresponding intersection size. The plot highlights significant overlaps, particularly among metabolic, cardiovascular, and neurological disorders

Classification results for the risk of progression from MCI to AD using machine learning algorithms

The classification of MCI-to-AD progression risk was assessed using six ML algorithms (RF, KNN, ANN, SVM, XGBoost, LightGBM) across three SNP sets (Table 3).

Table 3 Model performance (Best per algorithm per model, 10 × 10 cross-validation, mean ± SD)
Full size table

  • Model 1 (AD + VD vs. SCD; 54 SNPs; n = 285): Boosting models achieved the best overall performance, with XGBoost (AUC 0.866 ± 0.077; PR-AUC 0.939 ± 0.040) and LightGBM (AUC 0.859 ± 0.078; PR-AUC 0.936 ± 0.044) outperforming RF, KNN, and ANN (AUCs ~ 0.847–0.858). SVM achieved high accuracy but lower AUC (0.766 ± 0.088).

  • Model 2 (AD vs. SCD; 60 SNPs; n = 225): RF achieved the highest AUC (0.829 ± 0.085), closely followed by KNN and ANN (0.828 each). By contrast, XGBoost provided the highest PR-AUC (0.866 ± 0.085), suggesting stronger performance in precision–recall space despite slightly lower ROC-based AUC. LightGBM performed modestly across both metrics (AUC 0.809 ± 0.092; PR-AUC 0.861 ± 0.077).

  • Model 3 (Union panel; 76 SNPs; n = 225): XGBoost achieved the highest AUC (0.881 ± 0.074; PR-AUC 0.924 ± 0.055), with LightGBM performing similarly (AUC 0.868 ± 0.079; PR-AUC 0.921 ± 0.056). ANN also performed strongly (AUC 0.873 ± 0.083).

  • Model 3 sensitivity analysis (Union panel; n = 285): Boosting methods again yielded the strongest results (XGBoost AUC 0.864 ± 0.077; PR-AUC 0.938 ± 0.041; LightGBM AUC 0.858 ± 0.079; PR-AUC 0.935 ± 0.043).

Across all datasets, AUC was treated as the primary criterion, with PR-AUC considered complementary. The best-performing algorithms by AUC are highlighted in Table 3. Notably, boosting algorithms consistently provided the highest or near-highest performance across all models, with stable results across folds. Optimal configurations converged to shallow trees (XGBoost: depth 3–5, LR 0.03; LightGBM: 15 leaves), indicating well-regularized solutions without overfitting. Supplementary Table S3 further shows that threshold-specific sensitivity and specificity remained stable across folds, and calibration plots confirmed good agreement between predicted and observed risks (Supplementary Fig. S2–S5), with Brier scores in the 0.116–0.183 range.

Observed clinical outcomes of predicted risk groups using two-year follow-up data

Given that no SNP loci specifically associated with VD were detected in the GWAS analysis, we focused on exploring whether SNP-based risk groupings could reflect actual clinical progression to AD. Using the 2-year follow-up data of 80 individuals initially diagnosed with MCI, we examined whether those classified into high-risk groups at baseline were more likely to show cognitive decline over time. Of the 80 participants, 61 had both baseline genotype data and 2-year follow-up clinical outcomes, and 14 progressed from MCI to AD during the follow-up period (Table 4).

Table 4 Actual and predicted changes for cognitive function status (MCI to AD)
Full size table

At the F1-max threshold, model-level sensitivities (Sensitivityᵃ: true positive rate) ranged from 0.571 to 1.000 across classifiers, with distinct patterns across models. In Model 1, sensitivities were generally high (0.857–1.000), with KNN achieving 1.000 and ANN/SVM/XGB all at 0.929, while RF and LGBM each reached 0.857. In Model 2, sensitivities were slightly lower (0.714–0.857); KNN and ANN both achieved 0.857, RF/SVM/XGB each reached 0.786, and LGBM was lowest at 0.714. Model 3 showed the widest variation (0.571–1.000): KNN again achieved 1.000, while ANN dropped to 0.643 and LGBM had the lowest sensitivity at 0.571.

In addition, subject-level prediction agreement (Sensitivityᵇ: the proportion of algorithms that correctly predicted each converter) revealed heterogeneity across individuals. Several subjects (e.g., 1, 2, 9, 10, 12, 13) were consistently identified as converters by all algorithms (Sensitivityᵇ = 1.0). In contrast, subjects 11 (Sensitivityᵇ = 0.30) and subjects 6 (Sensitivityᵇ = 0.45) were misclassified by most algorithms, underscoring clinical heterogeneity in conversion trajectories.

These findings indicate that Model 1 provided the most stable and accurate predictions overall, with high sensitivity and agreement across classifiers. Model 2 achieved moderate performance, with ANN and KNN showing relatively stronger alignment. Model 3 displayed substantial variability across algorithms, ranging from perfect prediction (KNN) to poor discrimination (LGBM), suggesting greater inconsistency in its SNP panel. While overall predictive alignment with true outcomes was encouraging in some classifiers, the presence of misclassified converters highlights the limitations of SNP-based stratification alone. Integration of additional clinical, metabolic, and vascular markers will likely be necessary to improve risk stratification and reduce misclassification.

As a temporal validation, we applied pre-specified CV-derived thresholds to an independent 2-year follow-up cohort (n = 61; 14 converters). Sensitivity, specificity, PPV, and NPV with 95% CIs are reported in Supplementary Table S4, and risk group validation in Supplementary Table S5. Although discrimination was modest (AUROC approximately 0.45–0.55), calibration remained acceptable in the follow-up cohort, with no systematic over- or under-estimation of conversion risk (Supplementary Fig. S2-S5). This analysis demonstrates the models’ prospective performance without retraining. With 14 conversion events, precision of discrimination estimates is limited (wide CIs, nonsignificant p-values), which likely underestimates longer-term performance. Despite this, the direction of enrichment (higher event rates in high-risk groups across several algorithms) and acceptable calibration support the presence of a genuine, albeit modest, predictive signal.

To further contextualize these results, we conducted a detailed clinical analysis of subject 6 and subject 11, the two individuals who progressed to AD but showed the lowest subject-level sensitivity (Sensitivityᵇ = 0.45 and 0.30, respectively). Compared with other converters correctly predicted by most models, these two cases exhibited distinct deviations across vascular, metabolic, and hematologic markers. Subject 6 presented with moderately elevated systolic blood pressure (Z = + 1.02) but reduced diastolic blood pressure (Z = − 1.04), suggesting hemodynamic imbalance. Multiple hematologic indices, including hematocrit (Z = − 0.97), hemoglobin (Z = − 0.89), and MCV/MCH (approximately − 0.7), were lower than the group mean, consistent with mild anemia or nutritional insufficiency. Albumin (Z = − 0.35) was also reduced. This constellation of metabolic and hematologic alterations may have influenced disease progression through pathways not captured by SNP-based genetic models [53]. Subject 11 showed a markedly different profile, with diastolic blood pressure (Z = + 1.79) and pulse (Z = + 1.53) both elevated, consistent with vascular stress. BMI (Z = + 1.82) indicated overweight status, while total cholesterol (Z = − 0.74) and HDL cholesterol (Z = − 1.10) were below average. Albumin (Z = − 0.17) and hemoglobin/hematocrit (Z approximately − 0.6) were also modestly reduced. This combination of vascular burden, metabolic dysregulation, and borderline hematologic insufficiency again suggests risk pathways not accounted for in genetic classifiers.

Taken together, these cases illustrate how clinical heterogeneity among MCI patients—ranging from anemia-like profiles to vascular–metabolic stress—may contribute to discrepancies between genetic risk predictions and actual outcomes. Such findings underscore the importance of integrating metabolic, vascular, and hematologic biomarkers with SNP data to enhance the sensitivity and specificity of dementia risk stratification models.

This individual-level divergence supports the notion that clinical trajectories may diverge despite similar genetic profiles. Cases misclassified as low risk often displayed vascular/metabolic abnormalities (e.g., dyslipidemia, obesity, high diastolic BP), pointing to clinical modifiers outside the SNP panels. Conversely, high-risk predictions among non-converters may reflect preclinical processes or progression beyond the two-year observation period. Overall, these results do not demonstrate definitive predictive power but reflect an exploratory attempt to link SNP-based stratification with short-term outcomes. They highlight both potential utility and limitations: (i) genetic classifiers can capture subsets of converters, but (ii) misclassification arises where metabolic/vascular profiles dominate or where follow-up is too short to capture latent progression. Future work should integrate additional markers (imaging, CSF, clinical/biochemical) and extend follow-up horizons to improve dementia risk prediction.

Discussion

The primary objective of this study was to develop predictive models to classify patients with MCI into high- and low-risk groups for dementia using SNP chip data and ML algorithms. Through GWAS analysis, we identified several SNP loci that were significantly associated with dementia, and these SNPs were subsequently used to train and evaluate the ML models. We identified 54 SNPs (AD + VD vs. SCD), 60 SNPs (AD vs. SCD), and 76 SNPs (Union panel), mapping to 4–14 approximately independent loci after LD clumping. As expected, most signals arose from the APOE–PVRL2–TOMM40 block, but additional tagging variants were identified at other loci, underscoring both the central role of APOE and the polygenic nature of dementia risk. Consistent with the conservative nature of genome-wide FDR correction and the limited sample size, q-values were uniformly high. This supports our use of the p < 1 × 10⁻⁵ threshold not as evidence of genome-wide significance, but as a pragmatic criterion for SNP feature selection in machine learning models. Notably, adjusting for PC1–PC10 increased the number of suggestive SNPs (AD + VD: 38→54; AD: 44→60) without evidence of inflation (λGC = 1.021–1.029; Supplementary Fig. S1). Thus, the increase likely reflects improved capture of tagging variants in the APOE region rather than technical inflation.

Given the modest sample size and high-dimensional feature space, overfitting was a key concern. We mitigated this risk through strict separation of feature selection and prediction, repeated stratified cross-validation with nested tuning, and calibration confined to training folds. The convergence of boosting models to shallow, regularized configurations further suggests stable, non-overfit solutions. A second challenge in genetic prediction is class imbalance, as unstable predictions can arise when converters are relatively few compared to non-converters. Our analysis explicitly accounted for imbalance by stratifying all cross-validation folds, reporting balanced accuracy in addition to AUC and PR-AUC, calibrating probabilistic outputs, and applying threshold selection strategies within training folds. These safeguards ensured that evaluation metrics were not biased by unequal class sizes and that generalization performance was robust. Boosting methods consistently achieved the strongest or comparable performance, particularly in the union SNP panel (Model 3), where XGBoost reached an AUC of 0.881 and PR-AUC of 0.924. LightGBM showed similar performance, underscoring the robustness of gradient boosting in high-dimensional SNP classification tasks. Importantly, classical methods such as RF and KNN also remained competitive in certain models (e.g., RF in Model 2, AUC 0.829), highlighting that multiple algorithmic strategies can capture predictive signal. SVM, while achieving high accuracy, tended to underperform in AUC, suggesting lower discrimination ability despite correct classifications in balanced datasets. Taken together, these findings indicate that although class imbalance was present, it was effectively mitigated in our modeling framework through stratified validation, balanced accuracy reporting, and calibration. Boosting approaches (XGBoost, LightGBM) consistently provided well-regularized and high-performing strategies for SNP-based dementia risk prediction, while corroborating the stability of results across diverse classifiers. Furthermore, the convergence of boosting hyperparameters to shallow depth and low learning rates supports the interpretation that the predictive signal is genuine rather than an artifact of overfitting.

To further examine real-world applicability, we prospectively evaluated pre-specified thresholds in a temporally separated follow-up cohort (n = 61, 14 converters). As expected given the limited number of events, discrimination was modest (AUROC approximately 0.45–0.55), with relatively high false-positive rates. Nonetheless, event rates were generally higher in predicted high-risk groups, and odds ratios were often > 1, suggesting directional consistency. Importantly, thresholds determined by Youden’s J and F1-max showed systematic differences: Youden’s J yielded more balanced sensitivity and specificity, while F1-max prioritized converter detection, achieving higher true positive rates at the expense of specificity. From a clinical perspective, this trade-off is meaningful—minimizing missed converters may outweigh preserving specificity in early intervention or preventive contexts. Accordingly, we report F1-max–based operating points as the primary analysis while also providing Youden’s J for balance-oriented use cases.

These results should also be interpreted in light of the small sample size (n = 61) and short follow-up duration (2 years, 14 conversion events). Such constraints inherently cap the achievable AUROC, since many genetically high-risk individuals may not yet have converted but could do so with longer observation. Thus, the modest discrimination observed here likely represents a conservative lower-bound estimate of predictive ability. The enrichment of conversion events in high-risk groups nonetheless suggests that the models captured biologically relevant signal despite statistical underpowering.

Case-level discrepancies (e.g., subjects 6 and 11) suggest that vascular, metabolic, and hematologic factors can drive progression outside genetic risk captured by SNP panels, whereas uniformly detected converters likely reflect stronger genetic load. These observations support positioning SNP-based models as screening/stratification aids to triage individuals for confirmatory modalities (neuroimaging, CSF, longitudinal testing), rather than stand-alone diagnostics.

To explore the biological significance of the selected SNPs, we conducted functional annotation using SNPnexus. The strongest signals mapped to the APOE–PVRL2–TOMM40 block, including rs429358 (APOE ε4, p.Cys130Arg) and rs769449 (APOE promoter/intronic variant), both of which are well-established AD risk alleles. These findings reaffirm APOE’s central role in dementia risk. Several additional loci (e.g., on chromosomes 9, 14, and 20) were also detected, but they most likely represent LD-tagging rather than novel independent associations. The UpSet plot highlighted substantial pleiotropy: many SNPs overlapped across metabolic, cardiovascular, and neurological disease categories. This multicategory association suggests that these variants may contribute to a shared genetic architecture across diverse complex disorders, consistent with the pleiotropic nature of AD pathogenesis [54,55,56]. For instance, APOE-linked variants influence cholesterol transport and lipid metabolism, pathways that are directly relevant to both AD and vascular disease. Other variants overlapped with loci previously implicated in vascular traits and other neurodegenerative disorders. This convergence suggests that shared biological pathways—spanning lipid homeostasis, vascular integrity, and synaptic maintenance—contribute to AD pathogenesis.

This pleiotropic profile may help explain the observed clinical heterogeneity. Some individuals genetically classified as high risk did not progress within the 2-year follow-up, while some low-risk individuals converted. This discrepancy likely reflects the interplay between genetic predisposition and additional non-genetic or systemic factors, including vascular comorbidity and metabolic health. Thus, our findings support the view that AD is not a unitary disease entity but a complex syndrome with multiple pathophysiological pathways that potentially converge on similar clinical phenotypes [49, 57]. Integrating SNPs associated with AD and mixed dementia (AD + VD) in Model 3 yielded the most robust cross-validation performance, suggesting that accounting for overlapping neurodegenerative and cerebrovascular mechanisms may improve predictive accuracy. However, temporal validation revealed heterogeneous subject-level responses across algorithms, underscoring variability when models are applied prospectively. This approach acknowledges the growing evidence of mixed pathologies in late-life dementia and provides a more comprehensive genetic risk assessment [58, 59].

This study has several limitations. The cohort was relatively small (n = 674) and restricted to the Korean population, which may limit generalizability. The follow-up period (2 years) was short, with only 14 conversion events, limiting statistical power and precluding robust survival analyses. Modest sensitivity and high false-positive rates indicate that genetic markers alone provide incomplete stratification. Furthermore, no external SNP cohorts with dementia outcomes were available for replication. Importantly, the transferability of polygenic or ML-based risk models across ancestries is not guaranteed, as allele frequencies and LD structures differ substantially between populations. East Asian cohorts, for example, exhibit longer LD blocks, which may improve tagging but reduce fine-mapping resolution. Prior reviews emphasize that such ancestry-specific genetic architectures affect the portability of PRS and ML models [13, 14]. In addition, clinical characteristics such as years of education, BMI, and comorbidities were not incorporated into the predictive models, even though they are known to influence dementia risk. Future studies should assess whether integrating such factors with SNP-based predictors enhances predictive accuracy and clinical utility. Thus, while our models demonstrate proof-of-concept utility in a Korean cohort, future efforts must include multi-ethnic validation to assess generalizability and refine prediction across diverse populations.

Another methodological consideration concerns the interface between GWAS-based feature selection and cross-validation. In principle, performing GWAS separately within each training fold would provide the strictest safeguard against leakage. However, given the modest case–control sample sizes, fold-level GWAS would likely have produced unstable and inconsistent SNP panels. We therefore defined parsimonious panels once per subset (p < 1 × 10⁻⁵, adjusted for age, sex, and PC1–PC10) and carried them forward unchanged for model development. This strategy, combined with the use of a composite endpoint (any dementia; AD + VD vs. SCD) in Model 1, enhanced stability while capturing a broader spectrum of associated variants. To mitigate residual optimism, we combined this fixed-panel strategy with repeated nested cross-validation (with calibration and thresholding confined to training folds) and a temporally separated validation cohort. The latter, despite modest discrimination (AUROC approximately 0.45–0.55), provided a conservative lower-bound estimate of real-world performance, reinforcing that our design prioritized stability and minimized bias under the sample size constraints.

These limitations underscore the need for larger, multi-center, and multi-ethnic cohorts with extended follow-up. Future research should expand prediction accuracy and applicability by incorporating multimodal data (neuroimaging, CSF biomarkers, cognitive assessments) and multi-omics strategies (transcriptomics, proteomics, epigenomics). Longer follow-up will also clarify whether current false positives eventually convert, refining clinical interpretation. In addition, feature selection strategies informed by biological pathway enrichment or network analysis may help identify functionally relevant SNP combinations.

Finally, the ethical implications of genetic risk prediction must be considered. Early identification of high-risk individuals may facilitate timely interventions and care planning but also raises concerns regarding psychological impact, privacy, discrimination, and access to therapies. As predictive models advance, careful integration with genetic counseling frameworks and policy development will be essential.

Conclusions

We examined the potential of using SNP chip data and ML algorithms to classify individuals with MCI into high- and low-risk groups for progression to AD. We trained and evaluated multiple ML models using GWAS-identified SNP loci to assess how these genetic risk classifications aligned with 2-year clinical outcomes. Our findings revealed that certain models, particularly those using 54 SNPs from the AD + VD vs. SCD comparison (Model 1), demonstrated the most consistent alignment with observed clinical conversions. Of the 14 participants who progressed to AD, 12–14 were identified as high-risk by several classifiers. Model 2 (60 SNPs from AD vs. SCD) captured 11–12 converters depending on the algorithm, while Model 3 (76 SNPs, union panel) showed the widest variability, ranging from 8 to 14 correctly identified converters across classifiers. These results indicate that genetic classifiers can enrich for converters at a group level, but individual-level prediction remains constrained by short follow-up and non-genetic modifiers. Accordingly, SNP-based ML should complement, not replace, clinical and biomarker assessments. Future studies should aim to enhance model robustness by incorporating additional biomarkers such as neuroimaging, cerebrospinal fluid data, and longitudinal cognitive assessments. Expanding to larger and more diverse cohorts will also improve generalizability. Validation across ethnically diverse populations is essential to assess transportability and reduce performance bias. Multi-omics integration may also improve the biological interpretability and clinical applicability of such models. As understanding of dementia pathophysiology deepens, combining genetic insights with other clinical and environmental factors is essential to develop more reliable tools for early risk identification and personalized intervention planning. Beyond their immediate application, these models may contribute to the long-term development of predictive frameworks that support timely interventions and resource allocation in clinical settings. In summary, SNP-based ML offers a feasible framework for risk stratification in Korean MCI populations, with clear paths for improvement via longer follow-up and multimodal integration.