Resource-stratified machine learning framework for cognitive status classification and mild cognitive impairment to dementia progression prediction.

Resource-stratified machine learning framework for cognitive status classification and mild cognitive impairment to dementia progression prediction - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Alzheimers Res Ther . 2026 Mar 7;18:86. doi: 10.1186/s13195-026-02006-7 Search in PMC Search in PubMed View in NLM Catalog Add to search Resource-stratified machine learning framework for cognitive status classification and mild cognitive impairment to dementia progression prediction Jingmei Yang Jingmei Yang 1 Department of Electrical & Computer Engineering, Division of Systems Engineering, Department of Biomedical Engineering, Boston University, 8 St. Mary’s St, Boston, MA 02215 USA Find articles by Jingmei Yang 1 , Huitong Ding Huitong Ding 2 Department of Anatomy & Neurobiology and Framingham Heart Study, Chobanian & Avedisian School of Medicine, Boston University, 72 E Concord St, Boston, MA 02118 USA Find articles by Huitong Ding 2 , Samad Amini Samad Amini 1 Department of Electrical & Computer Engineering, Division of Systems Engineering, Department of Biomedical Engineering, Boston University, 8 St. Mary’s St, Boston, MA 02215 USA Find articles by Samad Amini 1 , Boran Hao Boran Hao 1 Department of Electrical & Computer Engineering, Division of Systems Engineering, Department of Biomedical Engineering, Boston University, 8 St. Mary’s St, Boston, MA 02215 USA Find articles by Boran Hao 1 , Cody Karjadi Cody Karjadi 2 Department of Anatomy & Neurobiology and Framingham Heart Study, Chobanian & Avedisian School of Medicine, Boston University, 72 E Concord St, Boston, MA 02118 USA Find articles by Cody Karjadi 2 , Rhoda Au Rhoda Au 2 Department of Anatomy & Neurobiology and Framingham Heart Study, Chobanian & Avedisian School of Medicine, Boston University, 72 E Concord St, Boston, MA 02118 USA 3 Department of Neurology and Medicine, Chobanian & Avedisian School of Medicine, Boston University, 72 E Concord St, Boston, MA 02118 USA 4 Department of Epidemiology, School of Public Health, Boston University, 72 E Concord St, Boston, MA 02118 USA Find articles by Rhoda Au 2, 3, 4 , Ioannis Ch Paschalidis Ioannis Ch Paschalidis 1 Department of Electrical & Computer Engineering, Division of Systems Engineering, Department of Biomedical Engineering, Boston University, 8 St. Mary’s St, Boston, MA 02215 USA 5 Faculty of Computing & Data Sciences, Boston University, 665 Commonwealth Ave, Boston, MA 02215 USA Find articles by Ioannis Ch Paschalidis 1, 5, ✉ Author information Article notes Copyright and License information 1 Department of Electrical & Computer Engineering, Division of Systems Engineering, Department of Biomedical Engineering, Boston University, 8 St. Mary’s St, Boston, MA 02215 USA 2 Department of Anatomy & Neurobiology and Framingham Heart Study, Chobanian & Avedisian School of Medicine, Boston University, 72 E Concord St, Boston, MA 02118 USA 3 Department of Neurology and Medicine, Chobanian & Avedisian School of Medicine, Boston University, 72 E Concord St, Boston, MA 02118 USA 4 Department of Epidemiology, School of Public Health, Boston University, 72 E Concord St, Boston, MA 02118 USA 5 Faculty of Computing & Data Sciences, Boston University, 665 Commonwealth Ave, Boston, MA 02215 USA ✉ Corresponding author. Received 2025 Oct 26; Accepted 2026 Feb 28; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13081641  PMID: 41794745 Abstract Background Widespread access to diagnosis of cognitive decline remains inadequate. Assessment tools rely on neuroimaging, biofluid markers, or lengthy neuropsychological batteries. This reliance limits their use in primary care and low-resource settings and contributes to healthcare disparities. We developed and validated a three-level, resource-stratified machine learning framework to provide scalable dementia screening and prognostic risk stratification across diverse healthcare settings. Methods We used data from 31,081 participants in the National Alzheimer’s Coordinating Center. We trained Gradient Boosted Tree models for multi-class detection (Cognitively Intact/Mild Cognitive Impairment [MCI]/Dementia) and 10-year risk stratification (MCI-to-dementia progression). The framework tiered inputs by resource intensity. The minimal-resource Level 1 included demographic and basic functional data. The moderate-resource Level 2 added standard cognitive tests. The high-resource Level 3 added comprehensive neuropsychological batteries. Results The Level 3 model achieved high classification performance (AUC: 93.98%), and the Level 1 model achieved comparable performance (AUC: 91.53%). For MCI-to-dementia progression, the framework showed strong prognostic performance. The Level 3 and Level 1 models achieved AUCs of 85.90% and 81.90% for predicting progression over a 2-year window, respectively. Key predictors remained consistent across all resource levels, such as difficulty managing finances and self-reported cognitive decline, and sociodemographic factors, including education level and Black race. Conclusions The tiered system offers a scalable and accessible approach for dementia detection and prognostic stratification. It helps non-specialists conduct initial evaluations and identify high-risk individuals with minimal data. High-risk patients may then be triaged for specialized assessment. This resource-stratified framework offers a strategy to expand diagnostic capacity and reduce care inequities, especially in low-resource settings. Supplementary Information The online version contains supplementary material available at 10.1186/s13195-026-02006-7. Keywords: National Alzheimer’s Coordinating Center Data, Healthcare Disparities, Resource-Stratified Screening, Cognitive Assessment, MCI-to-Dementia Progression, Machine Learning Background Dementia affects over 57 million people and is the seventh leading cause of death globally [ 1 , 2 ]. The associated annual costs are projected to reach $2.8 trillion by 2030 [ 3 ]. This burden presents a societal challenge. Despite decades of research, therapeutic development has been limited, and most clinical trials have failed [ 4 – 10 ]. The lack of effective therapies underscores the importance of accessible identification. Diagnosis enables intervention with lifestyle modifications [ 11 – 17 ]. However, a major gap exists in diagnostic accessibility. This problem disproportionately affects low- and middle-income countries, where more than two-thirds of people with dementia reside [ 18 , 19 ]. In these settings, advanced diagnostic infrastructure is often severely limited or absent. Diagnostic services outside capital cities are available in only one in four low-income countries compared to two in three high-income countries 20 . Even within high-income nations, many underserved populations face significant hurdles to accessing care. Geographic remoteness and prohibitive costs contribute to underdiagnosis and delays in obtaining a specialist consultation [ 20 , 21 ]. The primary care setting is the first point of contact for most patients with cognitive concerns, yet it is ill-equipped for comprehensive dementia evaluation [ 22 – 27 ]. While recent guidelines encourage expanded cognitive assessment in primary care, providers face resource constraints as they lack access to advanced diagnostic modalities like neuroimaging and specialized laboratory facilities for cerebrospinal fluid or genetic analysis [ 28 ]. Furthermore, the time-intensive nature of traditional cognitive assessment is incompatible with the primary care workflow. For example, a comprehensive neuropsychological evaluation can require over an hour to administer, a timeframe that is not feasible in a regular visit. These limitations create a disconnect between where patients first seek help and where diagnostic tools are available. While recent machine learning models report high diagnostic accuracy for dementia, their real-world utility is limited by their design choice [ 29 , 30 ]. They predominantly rely on resource-intensive data. These data modalities, including neuroimaging, genetic markers, and cerebrospinal fluid biomarkers, are absent in the clinical settings where most patients need help [ 31 – 38 ]. Such settings include primary care clinics and other low-resource contexts, like rural areas and low- and middle-income countries, which lack the required specialized equipment and personnel. This creates a paradox where the most powerful analytical tools are developed for the best-resourced clinics rather than for the contexts that face the greatest diagnostic burden. Consequently, the potential of machine learning to expand diagnostic reach remains largely unrealized. We propose a resource-stratified framework for cognitive assessment designed to accommodate diverse clinical settings. The framework uses a cumulative three-tier structure: Level 1 incorporates demographic data and basic functional measures. Level 2 builds upon Level 1 by adding standardized cognitive screening tools. Level 3 further extends Level 2 with comprehensive neuropsychological assessments, typically available only in specialized centers. Using data from the National Alzheimer’s Coordinating Center (NACC) [ 39 ], we developed and validated machine learning models for two tasks: multi-class classification of cognitive status (cognitively intact [CI], Mild Cognitive Impairment [MCI], dementia) and prognostic risk stratification of MCI-to-dementia progression. This tiered framework may allow healthcare providers to select models appropriate for their available resources. This approach may help improve access to cognitive assessment tools in settings where advanced diagnostics are limited, particularly for underserved populations. Methods Data We utilized the Uniform Data Set (UDS) from the NACC, which harmonizes longitudinal data from 35 Alzheimer’s Disease Research Centers (ADRCs) across the United States. The UDS provides standardized annual clinical, cognitive, and behavioral assessments for participants across the cognitive continuum. We included participants aged ≥ 50 years with at least one UDS visit and clinical diagnosis (CI, MCI, or dementia). All participants provided informed consent through their respective centers. Cognitive status diagnoses in the NACC UDS are assigned through clinical evaluation at each participating ADRC. Depending on the center’s practice, diagnoses are made by either a consensus panel or a single physician (typically the clinician conducting the neurological examination). The cognitive status classification (cognitively intact, MCI, or dementia) is determined using the NACC-derived variable NACCUDSD, which integrates comprehensive clinical assessment including cognitive testing, functional evaluation, medical history, and neurological examination findings. Participants who are cognitively intact with cognitive intact have NACCUDSD = 1, those with MCI have NACCUDSD = 3, and those with dementia have NACCUDSD = 4. Resource-stratified assessment framework As shown in Fig. 1 , we developed a hierarchical framework that stratifies dementia assessment from readily available patient information to comprehensive clinical evaluations, designed for healthcare settings with varying time, expertise, and resource constraints. Our framework incorporates measures including demographic information (DEMO), neuropsychiatric and functional assessments (Other), cognitive screening tools (Mini-Mental State Examination [MMSE] and Montreal Cognitive Assessment [MoCA]), and comprehensive domain-specific neuropsychological testing (NP). Starting with basic measures available in primary care (Level 1: DEMO + Other), the framework progressively adds cognitive screening (Level 2: Level 1 + MMSE + MoCA), and detailed neuropsychological testing (Level 3: Level 2 + NP). Fig. 1. Open in a new tab Overview of our resource-stratified framework and modeling pipeline. We utilized longitudinal data from the NACC UDS, selecting participants aged ≥ 50 years with at least one visit. Level 1 includes basic demographics and questionnaire-based assessments (DEMO + Other); Level 2 adds cognitive screening tools (MMSE and MoCA); and Level 3 incorporates comprehensive neuropsychological testing (NP). Models were trained at each level for two key tasks: classification of cognitive status (CI, MCI, dementia) and prediction of progression from MCI to dementia Level 1 (DEMO + Other) comprises fundamental participant information that are low-burden measures that can be readily implemented in primary care settings. The demographic features (age, sex, race, education) combined with basic neuropsychiatric and functional assessments – specifically the Neuropsychiatric Inventory (NPI), Geriatric Depression Scale (GDS), and Functional Activities Questionnaire (FAQ) – provide an efficient starting point for cognitive decline risk assessment. These questionnaire-based measures can be collected during routine primary care visits by general practitioners or trained staff, requiring minimal specialized resources while capturing both modifiable and non-modifiable risk factors. Level 2 (DEMO + Other + MMSE + MoCA) augments Level 1 with two widely used cognitive screening instruments: MMSE and MoCA. These standardized tools, while requiring basic clinical training for administration, can be feasibly implemented in outpatient clinics and community settings within 5–10 min. Though they provide objective measures of global cognitive function and serve as standard first-line assessments, they may have limited sensitivity to subtle cognitive changes. This intermediate level represents a balance between clinical utility and resource demands, offering more direct cognitive assessment than Level 1 while remaining accessible to most healthcare settings. Level 3 (DEMO + Other + MMSE + MoCA + NP) integrates a comprehensive neuropsychological (NP) test battery that assesses multiple cognitive domains: memory, attention, working memory, processing speed, language, visuospatial skills, and executive function. These detailed assessments provide greater diagnostic specificity and can detect subtle cognitive changes that may be missed by Level 1 and Level 2 screening tools. While offering gold-standard cognitive profiling and better differentiation between cognitive statuses, these tests require substantial clinical expertise and time to administer, typically limiting their availability to specialized memory clinics or research centers. This level represents a significant increase in resource requirements, usually necessitating referral to specialists. Resource level selection is guided by healthcare setting capabilities and available clinical infrastructure. Level 1 assessments can be implemented in primary care and community health centers where demographic information and questionnaire-based measures are accessible. Level 2 is applicable to outpatient clinics equipped with cognitive screening tools. Level 3 is designed for specialty memory clinics with access to comprehensive neuropsychological testing. Patients may enter at any level based on clinical presentation and setting context. Direct Level 3 evaluation may be warranted for presentations requiring specialist assessment, including rapid cognitive decline, early-onset symptoms, or atypical presentations with predominant language, behavioral, or visuospatial deficits. Data preprocessing and feature selection We used participant-level stratified data splitting to ensure all records from the same participant were assigned exclusively to one set to prevent data leakage from having the same participant’s data across different sets. As shown in Fig. 2 , the preprocessing pipeline comprised missing values handling, statistical testing, data transformation, and feature selection. All preprocessing parameters were learned exclusively from training data and then applied to the test set. Fig. 2. Open in a new tab Preprocessing and feature selection pipeline Features with more than 50% missing values in the training set were excluded. For remaining features, we implemented an age-based imputation strategy accounting for the age-dependent nature of cognitive measures (see Figure S1). At baseline visits, missing values were imputed within age groups (50–59, 60–69, 70–79, 80–89, 90 + years) using age-group-specific statistics: means for numerical features, medians for ordinal features, and modes for categorical features. For follow-up visits, we used the last observation carried forward to maintain temporal integrity and prevent leakage from future instances. Feature selection followed a two-step approach. First, statistical significance testing was performed using Analysis of Variance (ANOVA) for numerical features and chi-square tests for ordinal and categorical features, with a threshold of 0.05. Only features passing this criterion were retained. Significant features then underwent standard transformations: one-hot encoding for categorical and ordinal features, and Min-Max normalization for numerical features to scale values to a 0–1 range. Second, recursive feature elimination (RFE) was applied to select the optimal feature subset. The complete lists of selected features for cognitive status classification and MCI-to-dementia progression prediction are provided in Supplementary Tables S1 and S2, respectively. Subsequently, hyperparameter optimization was performed using grid search on the training data and validated on the validation set. All transformation parameters (encoding mappings, minimum and maximum values) and optimal hyperparameters learned from training data were applied to the test set. Multi-class classification for cognitive status We developed ML models to classify participants’ cognitive status (CI, MCI, dementia) at each resource level. This modeling approach allowed us to analyze how classification performance evolved from basic demographic information to comprehensive neuropsychological evaluations. By examining the predictive capabilities across resource levels, we could quantify the relative contributions of different clinical measures to diagnostic accuracy. We implemented seven classification algorithms: Decision Tree (DT), Oblique Decision Tree (ODT), Random Forest (RF), Oblique Random Forest (ORF), Gradient Boosting Tree (GBT), Support Vector Machine (SVM), and Logistic Regression (LR). We selected these seven algorithms to compare three complementary modeling approaches. Tree-based ensemble methods are widely adopted in healthcare domain for risk prediction and disease detection due to their ability to handle mixed data types, capture non-linear relationships, and provide interpretable importance rankings [ 34 , 40 – 42 ]. As for oblique variants, these extensions relax the axis-aligned split constraint of traditional trees, enabling them to capture complex multivariate patterns and potentially higher-order interactions between cognitive measures [ 43 ]. As for linear and kernel-based models, these approaches provide comparisons based on different modeling assumptions. Logistic regression implements linear decision boundaries while SVM employs kernel-based non-linear transformations. This combination allowed us to evaluate three key contrasts: (a) linear versus non-linear approaches, (b) single-learner versus ensemble methods, and (c) axis-aligned versus oblique decision boundaries. Model performance was evaluated using multiple metrics. The Area Under the Receiver Operating Characteristic curve (AUC) was calculated using a one-vs-rest approach, where each cognitive status was evaluated as the positive class against all remaining categories combined as the negative class. With three cognitive status categories, this resulted in three individual AUCs corresponding to CI vs. others, MCI vs. others, and dementia vs. others. These three AUCs were then macro-averaged to produce a single aggregate metric. Overall classification accuracy measured the proportion of correctly classified cases across all cognitive status categories. The class-specific recall rates for CI, MCI, and dementia categories quantified each model’s ability to identify true positive cases within each cognitive status. Given the natural class imbalance, we employed weighted F1-scores to ensure fair evaluation across all cognitive status categories. MCI progression to dementia prediction To identify MCI participants at elevated risk for dementia conversion, we conducted a longitudinal analysis over 10 years. Participants diagnosed with MCI at baseline were followed through subsequent clinical visits to monitor conversion to dementia. Only participants who completed at least one follow-up visit after their baseline MCI diagnosis were included in this analysis. We developed binary classification models to predict whether MCI patients would convert to dementia within 10 years. The models can generate predictions for variable timeframes (e.g., 3 years, 5 years, 10 years) and provide individualized probability scores (0–1 scale) for dementia conversion within the specified interval. We applied the same resource-stratified framework and selected the GBT algorithm based on its superior performance in cognitive status classification. Models were trained at each resource level to predict MCI-to-dementia conversion, with each level incorporating progressively comprehensive clinical features. Performance was evaluated using AUC, accuracy, recall, and weighted F1 scores. Software All analyses were performed in Python. Data processing used NumPy and Pandas. Machine learning models were implemented using Scikit-learn (LR, SVM, RF, GBT), and Treeple (ODT, ORF). Feature selection employed Scikit-learn’s RFECV, hyperparameter optimization used GridSearchCV, and cross-validation used StratifiedKFold. Data preprocessing applied Limescale, and performance metrics were computed using Scikit-learn’s metrics module. Results Classification of cognitive impairment status: distinguishing CI, MCI, and dementia A total of 31,081 participants were included, comprising 14,109 (45.39%) with CI, 8,400 (27.03%) with MCI, and 8,572 (27.58%) with dementia (see Table 1 ). Significant demographic differences were observed across cognitive status groups. Cognitive impairment was associated with advanced age (80–89 years: 16.92% CI, 20.11% MCI, 24.28% dementia), male sex (34.83% CI vs. 50.04% MCI vs. 50.22% dementia), lower educational attainment (CI: 15.95 ± 2.88 vs. MCI: 15.23 ± 3.21 vs. dementia: 14.77 ± 3.45 years), and higher White representation (81.99% CI vs. 81.06% MCI vs. 84.67% dementia). Table 1. Baseline demographic characteristics of study participants for cognitive status classification. Cognitive status categories include CI, MCI, and dementia Feature Level CI MCI Dementia P -value N = 14,109 (45.39%) N = 8400 (27.03%) N = 8572 (27.58%) Age 50–59 1369 (9.71%) 579 (6.89%) 1039 (12.72%) < 0.05 60–69 4687 (33.27%) 2368 (28.57%) 2202 (26.96%) 70–79 5362 (38.03%) 3598 (43.37%) 3126 (38.29%) 80–89 2384 (16.92%) 1667 (20.11%) 1983 (24.28%) 90+ 307 (2.18%) 188 (2.27%) 222 (2.72%) Sex Male 4914 (34.83%) 4204 (50.04%) 4305 (50.22%) < 0.05 Female 9195 (65.17%) 4196 (49.96%) 4267 (49.78%) Education Years 15.95 ± 2.88 15.48 ± 3.21 14.77 ± 3.45 < 0.05 Race White 11,402 (81.99%) 6814 (81.06%) 7482 (84.67%) < 0.05 Black 2227 (16.03%) 1299 (15.45%) 854 (9.65%) Asian 366 (2.63%) 237 (2.82%) 153 (1.73%) Other 114 (0.82%) 50 (0.59%) 83 (0.94%) Open in a new tab Values for categorical variables (age group, sex, and race) are presented as count (percentage within group). Education is presented as mean ± standard deviation in years. For Race, the Other group includes participants who identified as American Indian/Alaska Native, Native Hawaiian/Pacific Islander, or additional underrepresented racial groups in the study cohort. P-values were derived from chi-square tests for categorical variables (age group, sex, and race) and one-way ANOVA for years of education. GBT consistently outperformed all other models across resource levels (Fig. 3 ). GBT performance increased progressively: Level 1 (AUC 91.53 ± 0.17%, accuracy 81.05 ± 0.27%), Level 2 (AUC 92.53 ± 0.16%, accuracy 82.08 ± 0.27%), and Level 3 (AUC 93.98 ± 0.13%, accuracy 83.38 ± 0.27%). DeLong tests showed that all performance improvements between resource levels were statistically significant ( p < 0.001 for all pairwise comparisons across CI, MCI, and dementia classifications). This resource-stratified improvement was most pronounced for MCI detection, with recall advancing from 54.02% (L1) to 58.95% (L3), while CI (from 90.96% to 91.91%) and dementia (from 84.90% to 87.72%) maintained consistently high performance across all levels. Figure 4 shows the top 10 most important features from the GBT model. Fig. 3. Open in a new tab Performance comparison for classification models across resource levels. Models include Decision Tree (DT), Oblique Decision Tree (ODT), Random Forest (RF), Oblique Random Forest (ORF), Gradient Boosting Tree (GBT), Support Vector Machine (SVM), and Logistic Regression (LR)Figure 3: Performance comparison for classification models across resource levels. Models include Decision Tree (DT), Oblique Decision Tree (ODT), Random Forest (RF), Oblique Random Forest (ORF), Gradient Boosting Tree (GBT), Support Vector Machine (SVM), and Logistic Regression (LR). Three resource levels are represented: L1 (blue), L2 (green), and L3 (purple), with error bars indicating variability across runs. ( a ) One-vs-rest AUC scores for each model across resource levels. ( b ) Overall classification accuracy across resource levels. ( c ) Weighted F1-scores across resource levels. ( d ) Class-specific recall rates for the Dementia class. ( e ) Class-specific recall rates for the MCI class. ( f ) Class-specific recall rates for the Cognitively Intact (CI) class Fig. 4. Open in a new tab Feature importance rankings across resource levels in cognitive status classification models. Top 10 features ranked by importance scores for each resource level, displayed in descending order. Resource levels represent cumulative combinations: Level 1 (demographic and basic clinical measures), Level 2 (adding cognitive screening), and Level 3 (incorporating neuropsychological testing). Key features include: DECIN_1.0 (informant-reported memory decline), SHOPPING (shopping alone), REMDATES (remembering dates/appointments), TRAVEL (traveling out of neighborhood), DECSUB_1 (subject-reported memory decline), BILLS (managing bills/checks), TAXES (managing tax records), EVENTS (keeping track of current events), PAYATTN (paying attention), NACCAGE (age at visit), NACCMMSE (total MMSE score), MEMUNITS (total number of story units recalled), TRAILB (trail making test B – total number of seconds to complete), ANIMALS (total number of animals named in 60 s) All trained models demonstrated consistent resource-dependent performance improvements, with AUC increasing from 90% (L1) to 94% (L3) and accuracy from 80% (L1) to 83% (L3) (Table 2 ). MCI detection presented the primary diagnostic challenge, with recall rates of 50–55% (L1) improving to 59% (L3), substantially lower than CI (85–90% across levels) and dementia (80–90% across levels). Table 2. Classification performance of GBT models across resource levels for cognitive status prediction (CI, MCI, and dementia) Resource AUC Accuracy F1 Recall CI Recall M Recall D Level 1 91.53 ± 0.17 81.05 ± 0.27 80.61 ± 0.29 90.96 ± 0.29 54.02 ± 0.80 84.90 ± 0.56 Level 2 92.53 ± 0.16 82.08 ± 0.27 81.72 ± 0.28 91.08 ± 0.30 56.32 ± 0.77 86.66 ± 0.51 Level 3 93.98 ± 0.13 83.38 ± 0.27 83.05 ± 0.28 91.91 ± 0.28 58.95 ± 0.77 87.72 ± 0.52 Open in a new tab Values represent mean ± standard deviation across 100 runs. AUC: area under the receiver operating characteristic curve (averaged across one-vs-rest for each class); F1: weighted F1-score; Recall CI/M/D: sensitivity for cognitively intact/MCI/dementia respectively. Resource levels are cumulative, where Level 1 includes demographic and basic clinical measures (DEMO + Other); Level 2 adds cognitive screening tests (MMSE, MoCA); and Level 3 incorporates comprehensive neuropsychological testing (NP) We conducted analyses to evaluate alternative classification tasks (Tables S5 and S6). Binary classification of CI versus any cognitive impairment (combining MCI and dementia) achieved sensitivity of 84.14–86.32% across resource levels. When dementia cases were excluded and only CI versus MCI was examined, the models achieved MCI sensitivity of 66.84–71.11%. Identifying MCI participants at risk for dementia progression To identify MCI participants at highest risk for dementia conversion, we analyzed 10,460 participants with baseline MCI diagnosis followed for up to 10 years. Of these, 3,762 (35.97%) progressed to dementia while 6,698 (64.03%) remained stable (Table 3 ). Participants who converted to dementia were typically older, more often male, and more likely to be White. Table 3. Demographic characteristics of MCI participants. Participants were divided based on whether they developed dementia during 10-year follow-up Feature Level Non-Progression Progression to Dementia P -value 6698 (64.03%) 3762 (35.97%) Age 50–59 362 (5.41%) 170 (4.52%) < 0.05 60–69 1630 (24.34%) 707 (18.80%) 70–79 2646 (39.50%) 1611 (42.82%) 80–89 1714 (25.58%) 1068 (28.40%) 90+ 346 (5.16%) 206 (5.47%) Sex Male 3088 (46.10%) 1898 (50.45%) < 0.05 Female 3610 (53.90%) 1864 (49.55%) Education Years 15.48 ± 3.20 15.62 ± 3.14 < 0.05 Race White 5275 (78.73%) 3292 (87.48%) < 0.05 Black 1191 (17.78%) 354 (9.41%) Asian 184 (2.75%) 102 (2.71%) Other 48 (0.72%) 14 (0.37%) Open in a new tab Values for categorical variables (age, sex, and race) are presented as count (percentage within group). Education is presented as mean ± standard deviation in years. For Race, the Other group includes participants who identified as American Indian/Alaska Native, Native Hawaiian/Pacific Islander, or additional underrepresented racial groups in the study cohort. P-values were derived from chi-square tests for categorical variables (age group, sex, and race) and one-way ANOVA for years of education. Analysis of MCI to dementia progression over 10 years revealed a clear temporal pattern (Figure S2). Most conversions occurred within the first few years after diagnosis, peaking at year 2 (1,366 participants, 13.06%) and declining steadily thereafter. The cumulative conversion rate plateaued after year 6, with one-third (35.97%) of participants progressing to dementia over the entire follow-up period. Table 4 ; Fig. 5 show that prediction accuracy declined as the follow-up window extended but stabilized after year 5. Models that incorporated in-depth cognitive assessments – neuropsychological testing – consistently outperformed those relying only on demographic, neuropsychiatric, and brief cognitive screening measures (L1 and L2). Short-term predictions were highly accurate across all models (AUC: 93–95% at 0–1 year window), but performance dropped by approximately 10% points by the 0–2 year window and plateaued thereafter, with 10-year AUCs ranging from 74% (L1) to 80% (L3). Level 3 demonstrated superior performance compared to Levels 1 and 2. The largest improvement occurred between Levels 2 and 3, underscoring the critical value of incorporating detailed neuropsychological profiling for predicting MCI-to-dementia progression. Comprehensive metrics across all prediction windows are provided in Supplement Table S3 and Table S4. Table 4. Performance metrics of resource-stratified models in predicting MCI-to-dementia progression across short-term (0–1, 0–2, 0–3 years) and 0–10 years windows Window Level AUC Accuracy Recall F1 0–10 1 74.78 ± 1.04 63.27 ± 1.87 83.19 ± 4.14 63.50 ± 2.17 0–10 2 77.22 ± 0.89 66.22 ± 2.15 81.38 ± 4.40 66.66 ± 2.35 0–10 3 79.55 ± 0.86 69.37 ± 1.98 80.42 ± 4.29 69.89 ± 2.02 0–1 1 93.99 ± 1.07 86.36 ± 1.92 87.78 ± 5.17 87.38 ± 1.63 0–1 2 95.19 ± 0.97 88.46 ± 1.68 87.62 ± 5.11 89.17 ± 1.46 0–1 3 95.48 ± 0.93 89.16 ± 1.58 87.44 ± 5.05 89.77 ± 1.41 0–2 1 81.90 ± 1.14 70.15 ± 1.73 89.42 ± 2.91 70.61 ± 1.78 0–2 2 84.03 ± 1.14 72.59 ± 1.87 87.70 ± 3.51 73.12 ± 1.87 0–2 3 85.90 ± 1.08 75.37 ± 1.80 87.43 ± 3.56 75.89 ± 1.76 0–3 1 78.90 ± 1.05 66.65 ± 1.74 88.86 ± 2.88 66.56 ± 1.98 0–3 2 81.32 ± 0.96 69.33 ± 1.99 86.97 ± 3.53 69.50 ± 2.17 0–3 3 83.59 ± 0.94 72.78 ± 1.83 86.68 ± 3.35 73.08 ± 1.89 Open in a new tab Values represent mean ± standard deviation across 100 runs. Resource levels are cumulative, where Level 1 includes demographic and basic clinical measures (DEMO + Other); Level 2 adds cognitive screening tests (MMSE, MoCA); and Level 3 incorporates comprehensive neuropsychological testing (NP). Fig. 5. Open in a new tab Performance comparison of MCI-to-dementia progression prediction across different metrics and prediction windows (0–1 to 0–10 years). Each plot shows mean performance values (points) with standard deviations (error bars). Level 1 (demographic and basic clinical measures), Level 2 (adding cognitive screening), and Level 3 (incorporating neuropsychological testing). The x-axis represents prediction windows from 0–1 to 0–10 years, while the y-axis indicates the corresponding performance metric values (AUC, accuracy, F1-score, and recall) for each resource level Accuracy and F1 score trends closely mirrored the AUC findings, reinforcing the consistent performance hierarchy (L3 > L2 > L1) and the pattern of declining performance over longer prediction windows with eventual stabilization after year 5. In the short term (0–1 year), all models performed well, with Level 3 achieving 89–90% accuracy and F1 scores, and Levels 1 and 2 reaching 85–87%. By year 10, performance declined across all levels, while Level 3 retained higher accuracy (69–71%) and F1 scores (70–72%) compared to Levels 1 and 2 (63–65%). The relative importance of predictive features shifted across resource levels (Fig. 6 ). At Level 1, which includes only demographics and basic functional/neuropsychiatric measures, age (NACCAGE) was the strongest predictor, followed by functional assessments of daily living activities: remembering dates and appointments (REMDATES), managing finances (BILLS, TAXES), and traveling independently outside the neighborhood (TRAVEL). Years of education (EDUC) and smoking history (SMOKYRS) contributed additional predictive value. Fig. 6. Open in a new tab Feature importance rankings across resource levels in MCI-to-dementia progression models. Top 10 features ranked by importance scores for each resource level, displayed in descending order. Resource levels represent cumulative combinations: Level 1 (demographic and basic clinical measures), Level 2 (adding cognitive screening), and Level 3 (incorporating neuropsychological testing). Key features include: NACCAGE (age at visit), REMDATES (remembering dates/appointments), BILLS (managing bills/checks), DECIN_1.0 (reported memory decline), TAXES (managing tax records), HRATE (resting heart rate), BPDIAS (diastolic blood pressure), TRAVEL (traveling out of neighborhood), SMOKYRS (total years smoked cigarettes), EDUC (years of education), NACCMMSE (total MMSE score), MMSEORDA (orientation subscale score – Time), MEMUNITS (total number of story units recalled), TRAILA and TRAILB (trail making test A and B – total number of seconds to complete), VEG (total number of vegetables named in 60 s), and ANIMALS (total number of animals named in 60 s) Level 2 added cognitive screening tools. Total MMSE score (NACCMMSE) became the dominant predictor, with the MMSE orientation to time subscale (MMSEORDA) also ranking among the top features. Demographic and functional measures remained in the top 10 but with reduced importance compared to Level 1. Level 3 incorporated comprehensive neuropsychological testing. Memory recall (MEMUNITS) became the single most important feature with substantially higher importance than other variables. Processing speed measures (Trail Making Test A and B completion times: TRAILA, TRAILB) and verbal fluency tasks (vegetables and animals named in 60 s: VEG, ANIMALS) dominated the remaining top positions. Demographic and functional assessments, while still present, contributed less to prediction when cognitive measures were available. These importance shifts follow a clear progression: demographic and functional assessments provide the strongest signals when comprehensive cognitive data are unavailable, while neuropsychological tests capture cognitive decline more directly with greater discriminative power when accessible. Informant-reported memory decline (DECIN_1.0) and difficulty managing bills (BILLS) maintained consistent importance across all three levels. The models incorporate the time interval from baseline MCI diagnosis to dementia diagnosis as an input feature alongside demographic, functional, and cognitive measures. This design may allow clinicians to specify the prediction window during model inference. When assessing a patient, the clinician could input the desired time horizon (e.g., 1 year, 3 years, or 10 years from the current MCI diagnosis), and the model would generate a corresponding conversion probability between 0 and 1. The prediction window is parameterized as a model input rather than fixed during training. This approach could potentially support flexible risk assessment tailored to individual patient care planning contexts. Discussion Our study establishes a flexible, resource-stratified framework for cognitive assessment and risk stratification across diverse clinical environments. This tiered approach offers a scalable solution to cognitive evaluation, as it has the potential to enable more accessible identification and stratification in settings with varying resource levels. The framework’s performance scaled with data intensity, and all tiers delivered strong diagnostic and prognostic accuracy. The Level 3 model, using comprehensive neuropsychological data, achieved a classification AUC of 93.98% and a 10-year prognostic AUC of 79.56%. Importantly, the Level 1 model demonstrated high utility with only basic inputs, reaching a classification AUC of 91.53%. These findings highlight the viability of a multi-level assessment strategy that expands diagnostic accessibility within real-world clinical constraints. As shown in Table 2 , The lower MCI sensitivity (54.02–58.95%) compared to CI (90.96–91.91%) and dementia (84.90–87.72%) is consistent with MCI’s status as the most diagnostically challenging category in cognitive assessment. This pattern is consistent across the literature and arises from three factors: (1) MCI’s clinical instability as a transitional state, (2) its etiological heterogeneity, and (3) the inherent limitations of clinical measures without biomarkers such as amyloid and tau pathology markers, and APOE4 genotyping for capturing this intermediate stage. MCI represents a transitional and unstable cognitive state. Unlike CI and dementia, which are relatively stable endpoints, MCI can follow multiple trajectories: progression to dementia, reversion to normal cognition, or fluctuation between states. MCI shows substantial instability, with reversion to normal cognition occurring in approximately 9% of clinical cohorts and up to 28% in population-based studies [ 44 ]. This instability makes MCI more difficult to capture in predictive models, as individuals with MCI exhibit more variable clinical trajectories and heterogeneous underlying pathologies. Our study includes all-cause MCI without differentiating by etiology, including cases driven by early Alzheimer’s pathology, vascular cognitive impairment, depression-related cognitive changes, medication effects, and other reversible causes. This heterogeneity reflects real-world clinical practice where the underlying cause of MCI is often unknown at initial presentation, but it also increases the challenge of prediction since different etiologies have fundamentally different progression patterns and respond to different predictive features. Our models use clinical, functional, and neuropsychological measures without neuroimaging, biofluid biomarkers, or genetic data. Even studies incorporating comprehensive multimodal data with neuroimaging and biomarkers report persistent challenges with MCI classification. For example, a study using structural MRI combined with CSF biomarkers, cognitive scores, and APOE4 status achieved only 58.8% accuracy for three-class classification (CI vs. MCI vs. dementia) [ 45 ]. Another study using multimodal data including hippocampal volume, CSF markers, and APOE4 genotyping reported an AUC of 0.68 for predicting conversion among APOE4-positive MCI participants, suggesting that even with genetic information, MCI progression remains difficult to predict [ 46 ]. Even advanced deep learning approaches using MRI show that MCI remains the most challenging category, with models achieving lower for MCI [ 47 ]. Given our intentionally accessible feature set designed for implementation across diverse healthcare settings with varying resources, the observed MCI sensitivity represents a reasonable performance. The 5% improvement from Level 1 to Level 3 (54.02–58.95%) demonstrates that progressively comprehensive assessment, from basic demographic and functional measures, to cognitive screening (MMSE), to detailed neuropsychological testing, improves MCI detection. Our framework offers a practical advantage over existing cognitive assessment models, which often rely on costly and inaccessible multimodal data. Many models integrate neuroimaging, cerebrospinal fluid biomarkers, and genetic data to achieve high accuracy, limiting their feasibility in resource-constrained settings [ 31 – 38 , 48 , 49 ]. For example, a model using hippocampal atrophy, amyloid PET, CSF biomarkers, and APOE genotyping achieved a 3-year MCI-to-dementia prognostic AUC of 80.50% [ 50 ]. Our Level 3 model demonstrated this performance, reaching an AUC of 83.70% over the same interval using only clinical and neuropsychological data. This suggests that competitive performance is feasible without reliance on resource-intensive inputs. Even among other resource-stratified frameworks, our approach is more accessible; the model by Ren et al. [ 51 ], for instance, required MRI for its highest tier, whereas our framework operates entirely independent of neuroimaging or biofluid markers. Finally, our model provides personalized prognostic risk scores by generating individualized conversion probabilities for specified assessment intervals. This granular risk stratification offers more specific clinical guidance than the clustering-based approaches in previous work, which stratified patients into risk groups but did not provide individualized probability estimates for specific time horizons. A challenge in dementia care is the trade-off between diagnostic precision and public health accessibility. Current paradigms often depend on resource-intensive specialist evaluations. This dependence creates barriers to care in underserved areas [ 23 , 52 ]. This constraint also postpones appropriate management and planning [ 53 ]. Our stratified framework addresses this challenge by enabling resource-adaptive risk stratification. Resource level selection depends on available clinical infrastructure rather than algorithmic cutoffs. This allows for flexible implementation across diverse clinical contexts. Within each tier, clinicians can integrate model-derived probabilities with clinical judgment to inform referral and management decisions. In resource-constrained settings where advanced assessments are unavailable, Level 1 provides an accessible screening tool using readily available data. This level identifies at-risk individuals within the general population. Conversely, when comprehensive neuropsychological testing is accessible, patients with complex presentations can proceed directly to Level 3 assessment. This flexibility extends screening capacity beyond specialist settings. It facilitates broader risk assessment while conserving specialist resources for patients presenting with complex clinical signs requiring comprehensive evaluation for a definitive diagnosis. Once high-risk individuals are identified, clinicians can implement closer follow-ups, recommend lifestyle modifications, or prioritize candidates for clinical trials [ 16 , 54 ]. Each tier has different time and resource requirements. Level 1 assessment (10–15 min) uses questionnaire-based measures. These can be collected by trained staff during routine primary care visits, particularly for patients aged ≥ 70 years with cognitive concerns. Level 2 (additional 10–15 min) adds MMSE cognitive screening to support specialist referral decisions in primary care settings. Level 3 (60–90 min) includes comprehensive neuropsychological testing administered by specialists in memory clinics, incorporating memory recall, verbal fluency, and processing speed. Implementation pathways vary by healthcare infrastructure. In high-resource settings, models support evidence-based triage from primary care (Level 1/2) to specialist evaluation (Level 3). In low-resource settings such as rural areas or developing countries, only Level 1 may be feasible. In these contexts, Level 1 provides risk assessment without advanced diagnostic infrastructure. For the MCI progression prediction task, Fig. 6 presents the most predictive features at each level, providing empirical guidance on assessment prioritization. Clinicians select resource levels based on their setting’s capabilities and the associated accuracy trade-offs, maximizing prognostic accuracy within their resource constraints. Our study uses functional assessments as predictors, and BILLS and TAXES rank among the top features (Fig. 6 ). These features derive from questionnaires. Collecting this data requires less time and expertise than comprehensive neuropsychological testing, which often demands hours of administration by trained specialists. We acknowledge a limitation regarding these features. They represent functional impairments that have already manifested. These variables signal clinical signs of decline rather than pre-symptomatic risk. Consequently, the reliance on these observable functional changes implies a correlation with disease status that likely drives the high performance metrics in our models. However, our framework aims to address a clinical gap. Rather than prioritizing new biomarkers or risk factors for pre-symptomatic detection, our approach offers utility for expanding diagnostic capacity in underserved regions. Many individuals with apparent functional decline remain undiagnosed due to a lack of specialists. We provide automated assessment tools designed to help non-specialists identify high-risk individuals in these contexts who need further evaluation. This capability supports broadened diagnostic coverage in underserved settings. Our framework is designed with the goal of expanding diagnostic reach and mitigating inequities in dementia care. Access to neurological specialists is a primary barrier to diagnosis in the rural regions of high-income countries and across many low- and middle-income countries. Our Level 1 model is intended to enable non-specialists to perform resource-stratified risk stratification. If validated, it could equip primary care providers, nurses, or community health workers with a scalable, evidence-based automated cognitive assessment tool. This capability is helpful for reducing diagnostic disparities. Our study has several limitations. First, the NACC data originates from specialized Alzheimer’s Disease Research Centers. NACC participants are typically recruited through academic medical centers and may exhibit selection bias toward higher educational attainment and superior healthcare access compared to general populations. This demographic profile differs from the underserved and low-resource settings. However, our selection of the NACC dataset was driven by methodological requirements: large sample size, comprehensive clinical assessments, and extended longitudinal follow-up. Alternative population-based datasets typically feature smaller samples, limited feature coverage, or shorter observation periods that would compromise model predictive power. Our primary objective was to establish a methodological blueprint demonstrating the feasibility of accurate cognitive assessment using accessible and clinical measures rather than resource-intensive features. External validation in general populations, particularly in low-resource settings, including rural areas and low- and middle-income countries, represents a critical next step. Second, the models demonstrated strong retrospective accuracy, yet their prospective clinical utility remains unknown. Prospective implementation trials are the next step to establish this utility. Such trials would evaluate the framework’s real-world feasibility, clinician acceptability, and impact on diagnostic timeliness. Our models also rely on cross-sectional data at the MCI diagnosis. This static approach may miss complex cognitive trajectories. Integrating longitudinal data from multiple visits is an important next step. Temporal modeling approaches, including recurrent neural networks or transformers, can enable dynamic risk updating and better characterize the conversion process. Incorporating accessible biomarkers represents a promising direction. Blood-based biomarkers offer several advantages. They are minimally invasive, cost-effective, and feasible to implement in primary care settings. This accessibility could enable earlier detection in community practices. Expanding screening and assessment capacity to primary care physicians would particularly benefit underserved and low-resource populations. Conclusions Overall, our resource-stratified framework represents a promising approach to enhancing assessment capacity across diverse clinical settings. This system could improve accessibility and equity by providing flexible assessments aligned with different resource levels. We anticipate this framework will become a valuable tool within real-world clinical settings for automated detection and large-scale assessments. Supplementary Information Supplementary Material 1. (584.8KB, docx) Acknowledgements We are grateful to the NACC for generously sharing the data. The NACC database is funded by National Institute on Aging (NIA)/National Institutes of Health (NIH) Grant U24 AG072122. NACC data are contributed by the NIA-funded ADRCs: P30 AG062429 (PI James Brewer, MD, PhD), P30 AG066468 (PI Oscar Lopez, MD), P30 AG062421 (PI Bradley Hyman, MD, PhD), P30 AG066509 (PI Thomas Grabowski, MD), P30 AG066514 (PI Mary Sano, PhD), P30 AG066530 (PI Helena Chui, MD), P30 AG066507 (PI Marilyn Albert, PhD), P30 AG066444 (PI John Morris, MD), P30 AG066518 (PI Jeffrey Kaye, MD), P30 AG066512 (PI Thomas Wisniewski, MD), P30 AG066462 (PI Scott Small, MD), P30 AG072979 (PI David Wolk, MD), P30 AG072972 (PI Charles DeCarli, MD), P30 AG072976 (PI Andrew Saykin, PsyD), P30 AG072975 (PI David Bennett, MD), P30 AG072978 (PI Neil Kowall, MD), P30 AG072977 (PI Robert Vassar, PhD), P30 AG066519 (PI Frank LaFerla, PhD), P30 AG062677 (PI Ronald Petersen, MD, PhD), P30 AG079280 (PI Eric Reiman, MD), P30 AG062422 (PI Gil Rabinovici, MD), P30 AG066511 (PI Allan Levey, MD, PhD), P30 AG072946 (PI Linda Van Eldik, PhD), P30 AG062715 (PI Sanjay Asthana, MD, FRCP), P30 AG072973 (PI Russell Swerdlow, MD), P30 AG066506 (PI Todd Golde, MD, PhD), P30 AG066508 (PI Stephen Strittmatter, MD, PhD), P30 AG066515 (PI Victor Henderson, MD, MS), P30 AG072947 (PI Suzanne Craft, PhD), P30 AG072931 (PI Henry Paulson, MD, PhD), P30 AG066546 (PI Sudha Seshadri, MD), P20 AG068024 (PI Erik Roberson, MD, PhD), P20 AG068053 (PI Justin Miller, PhD), P20 AG068077 (PI Gary Rosenberg, MD), P20 AG068082 (PI Angela Jefferson, PhD), P30 AG072958 (PI Heather Whitson, MD), and P30 AG072959 (PI James Leverenz, MD). Authors’ contributions JY conceived the study, conducted data analysis and experiments, and drafted the initial manuscript. IP designed the study, edited and revised the manuscript. HD edited the manuscript, conducted data analysis, and interpreted results. SA, BH, CK, and RA reviewed and revised the manuscript. All authors reviewed the manuscript and approved the final version. Funding The research was partially supported by the NSF under grants CCF-2200052, IIS-1914792, ECCS-2317079, and DEB-2433726, the NIH under grant UL54 TR00413, the Boston University Rajen Kilachand Fund for Integrated Life Science and Engineering, the National Institute on Aging under grants under grants AG062109, AG068753 , AG072654 , R03AG095992, and AG083735 . Data availability The datasets used in this study are publicly available through the National Alzheimer’s Coordinating Center (NACC) database. Data can be accessed through the NACC Data Request process. Declarations Ethics approval and consent to participate The data used in this study were obtained from the NACC database. All participants provided informed consent at their respective ADRCs prior to data collection and submission to NACC. As secondary users of de-identified data from the NACC database, consent was not necessary for the use of human subject data in this study. The use of NACC UDS data for this research was conducted in accordance with NACC data use agreements. Consent for publication All authors have reviewed the manuscript and consented to its publication. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. GBD 2019 Collaborators. Global mortality from dementia: Application of a new method and results from the Global Burden of Disease Study 2019. Alzheimers Dement Transl Res Clin Interv. 2021;7:e12200. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Arvanitakis Z, Shah RC, Bennett DA. Diagnosis and Management of Dementia. Rev JAMA. 2019;322:1589. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Alzheimer’s Association. 2019 Alzheimer’s disease facts and figures. Alzheimers Dement 2019;15: 321–387. 4. Yiannopoulou KG, Anastasiou AI, Zachariou V, et al. Reasons for Failed Trials of Disease-Modifying Treatments for Alzheimer Disease and Their Contribution in Recent Research. Biomedicines. 2019;7:97. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Kim CK, Lee YR, Ong L, et al. Alzheimer’s Disease: Key Insights from Two Decades of Clinical Trial Failures. J Alzheimers Dis. 2022;87:83–100. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Passeri E, Elkhoury K, Morsink M, et al. Alzheimer’s Disease: Treatment Strategies and Their Limitations. Int J Mol Sci. 2022;23:13954. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Reuben DB, Kremen S, Maust DT. Dementia Prevention and Treatment: A Narrative Review. JAMA Intern Med. 2024;184:563. [ DOI ] [ PubMed ] [ Google Scholar ] 8. Kim B-H, Kim S, Nam Y, et al. Second-generation anti-amyloid monoclonal antibodies for Alzheimer’s disease: current landscape and future perspectives. Transl Neurodegener. 2025;14:6. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Cummings J, Zhou Y, Lee G, et al. Alzheimer’s disease drug development pipeline: 2024. Alzheimers Dement Transl Res Clin Interv. 2024;10:e12465. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Asher S, Priefer R. Alzheimer’s disease failed clinical trials. Life Sci. 2022;306:120861. [ DOI ] [ PubMed ] [ Google Scholar ] 11. Livingston G, Huntley J, Liu KY, et al. Dementia prevention, intervention, and care: 2024 report of the Lancet standing Commission. Lancet. 2024;404:572–628. [ DOI ] [ PubMed ] [ Google Scholar ] 12. Kivipelto M, Solomon A, Ahtiluoto S, et al. The Finnish Geriatric Intervention Study to Prevent Cognitive Impairment and Disability (FINGER): Study design and progress. Alzheimers Dement. 2013;9:657–65. [ DOI ] [ PubMed ] [ Google Scholar ] 13. Rosenberg A, Ngandu T, Rusanen M, et al. Multidomain lifestyle intervention benefits a large elderly population at risk for cognitive decline and dementia regardless of baseline characteristics: The FINGER trial. Alzheimers Dement. 2018;14:263–70. [ DOI ] [ PubMed ] [ Google Scholar ] 14. Crivelli L, Calandri IL, Suemoto CK, et al. Latin American Initiative for Lifestyle Intervention to Prevent Cognitive Decline (LatAm-FINGERS): Study design and harmonization. Alzheimers Dement. 2023;19:4046–60. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Baker LD, Snyder HM, Espeland MA, et al. Study design and methods: U.S. study to protect brain health through lifestyle intervention to reduce risk (U.S. POINTER). Alzheimers Dement. 2024;20:769–82. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Ornish D, Madison C, Kivipelto M, et al. Effects of intensive lifestyle changes on the progression of mild cognitive impairment or early dementia due to Alzheimer’s disease: a randomized, controlled clinical trial. Alzheimers Res Ther. 2024;16:122. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Ngandu T, Lehtisalo J, Solomon A, et al. A 2 year multidomain intervention of diet, exercise, cognitive training, and vascular risk monitoring versus control to prevent cognitive decline in at-risk elderly people (FINGER): a randomised controlled trial. Lancet. 2015;385:2255–63. [ DOI ] [ PubMed ] [ Google Scholar ] 18. Kalaria R, Maestre G, Mahinrad S, et al. The 2022 symposium on dementia and brain aging in low- and middle‐income countries: Highlights on research, diagnosis, care, and impact. Alzheimers Dement. 2022;20:4290–314. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Mattap SM, Mohan D, McGrattan AM, et al. The economic burden of dementia in low- and middle-income countries (LMICs): a systematic review. BMJ Glob Health. 2022;7:e007409. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Seeher K, Cataldi R, Dua T, et al. Inequitable Access to Dementia Diagnosis and Care in Low-Resource Settings – A Global Perspective. Clin Gerontol. 2023;46:133–7. [ DOI ] [ PubMed ] [ Google Scholar ] 21. Gulline H, Carmody S, Yates M, et al. Equity of access in rural and metropolitan dementia diagnosis, management, and care experiences: an exploratory qualitative study. Int J Equity Health. 2025;24:74. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Alzheimer'sAssociation, 2024 Alzheimer’s disease facts and figures. Alzheimers Dement 2024;20:3708–3821. 10.1002/alz.13809. [ DOI ] [ PMC free article ] [ PubMed ] 23. Bradford A, Kunik ME, Schulz P, et al. Missed and Delayed Diagnosis of Dementia in Primary Care: Prevalence and Contributing Factors. Alzheimer Dis Assoc Disord. 2009;23:306–14. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. the EVIDEM-ED project, Koch T, Iliffe S. Rapid appraisal of barriers to the diagnosis and management of patients with dementia in primary care: a systematic review. BMC Fam Pract. 2010;11:52. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Boustani M, Schubert C, Sennour Y. The challenge of supporting care for dementia in primary care. Clin Interv Aging. 2007;2:631–6. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Fernandes B, Goodarzi Z, Holroyd-Leduc J. Optimizing the diagnosis and management of dementia within primary care: a systematic review of systematic reviews. BMC Fam Pract. 2021;22:166. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Hinton L, Franz CE, Reddy G, et al. Practice constraints, behavioral problems, and dementia care: primary care physicians’ perspectives. J Gen Intern Med. 2007;22:1487–92. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Atri A, Dickerson BC, Clevenger C, et al. Alzheimer’s Association clinical practice guideline for the Diagnostic Evaluation, Testing, Counseling, and Disclosure of Suspected Alzheimer’s Disease and Related Disorders (DETeCD-ADRD): Executive summary of recommendations for primary care. Alzheimers Dement. 2025;21:e14333. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Amini S, Hao B, Yang J, et al. Prediction of Alzheimer’s disease progression within 6 years using speech: A novel approach leveraging language models. Alzheimers Dement. 2024;20:5262–70. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Amini S, Zhang L, Hao B, et al. An artificial intelligence-assisted method for dementia detection using images from the clock drawing test. J Alzheimers Dis. 2021;83:581–9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Kung T-H, Chao T-C, Xie Y-R, et al. Neuroimage Biomarker Identification of the Conversion of Mild Cognitive Impairment to Alzheimer’s Disease. Front Neurosci. 2021;15:584641. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Xue C, Kowshik SS, Lteif D, et al. AI-based differential diagnosis of dementia etiologies on multimodal data. Nat Med. 2024;30:2977–89. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Suk H-I, Lee S-W, Shen D. Hierarchical feature representation and multimodal fusion with deep learning for AD/MCI diagnosis. NeuroImage. 2014;101:569–82. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Qiu S, Miller MI, Joshi PS, et al. Multimodal deep learning for Alzheimer’s disease dementia assessment. Nat Commun. 2022;13:3404. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Borchert RJ, Azevedo T, Badhwar A, et al. Artificial intelligence for diagnostic and prognostic neuroimaging in dementia: A systematic review. Alzheimers Dement. 2023;19:5885–904. [ DOI ] [ PubMed ] [ Google Scholar ] 36. Lee LY, Vaghari D, Burkhart MC, et al. Robust and interpretable AI-guided marker for early dementia prediction in real-world clinical settings. eClinicalMedicine. 2024;74:102725. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 37. Feng X, Provenzano FA, Small SA, et al. A deep learning MRI approach outperforms other biomarkers of prodromal Alzheimer’s disease. Alzheimers Res Ther. 2022;14:45. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 38. Javeed A, Dallora AL, Berglund JS, et al. Machine Learning for Dementia Prediction: A Systematic Review and Future Research Directions. J Med Syst. 2023;47:17. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 39. Besser L, Kukull W, Knopman DS, et al. Version 3 of the National Alzheimer’s Coordinating Center’s Uniform Data Set. Alzheimer Dis Assoc Disord. 2018;32:351–8. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 40. the Alzheimer’s Disease Neuroimaging Initiative, Chen D, Yi F, et al. A Stacking Framework for Multi-Classification of Alzheimer’s Disease Using Neuroimaging and Clinical Features. J Alzheimers Dis. 2022;87:1627–36. [ DOI ] [ PubMed ] [ Google Scholar ] 41. Martínez-Florez JF, Osorio JD, Cediel JC, et al. Short-Term Memory Binding Distinguishing Amnestic Mild Cognitive Impairment from Healthy Aging: A Machine Learning Study. J Alzheimers Dis. 2021;81:729–42. [ DOI ] [ PubMed ] [ Google Scholar ] 42. Reinke C, Doblhammer G, Schmid M, et al. Dementia risk predictions from German claims data using methods of machine learning. Alzheimers Dement. 2023;19:477–86. [ DOI ] [ PubMed ] [ Google Scholar ] 43. Liu X, Xiao Q, Gu Z, et al. Development and external validation of a machine learning-based model to predict postoperative recurrence in patients with duodenal adenocarcinoma: a multicenter, retrospective cohort study. BMC Med. 2025;23:98. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 44. Salemme S, Lombardo FL, Lacorte E, et al. The prognosis of mild cognitive impairment: A systematic review and meta-analysis. Alzheimers Dement Diagn Assess Dis Monit. 2025;17:e70074. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 45. for the Alzheimer’s Disease Neuroimaging Initiative, Gill S, Mouches P, et al. Using Machine Learning to Predict Dementia from Neuropsychiatric Symptom and Neuroimaging Data. J Alzheimers Dis. 2020;75:277–88. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 46. Apostolova LG, Hwang KS, Kohannim O, et al. ApoE4 effects on automated diagnostic classifiers for mild cognitive impairment and Alzheimer’s disease. NeuroImage Clin. 2014;4:461–72. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 47. Shah DR, Kapdi RA, Patel JS, et al. Three Class Classification of Alzheimer’s Disease Using Deep NeuralNetworks. Curr Med Imaging Former Curr Med Imaging Rev. 2023;19:e290922209274. [ DOI ] [ PubMed ] [ Google Scholar ] 48. Uddin KMM, Alam MJ, Jannat-E-Anawar, et al. A Novel Approach Utilizing Machine Learning for the Early Diagnosis of Alzheimer’s Disease. Biomed Mater Devices. 2023;1:882–98. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 49. Lim BY, Lai KW, Haiskin K, et al. Deep Learning Model for Prediction of Progressive Mild Cognitive Impairment to Alzheimer’s Disease Using Structural MRI. Front Aging Neurosci. 2022;14:876202. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 50. Pang Y, Kukull W, Sano M, Albin RL, Shen C, Zhou J, Dodge HH. Predicting progression from normal to MCI and from MCI to AD using clinical variables in the national Alzheimer's coordinating center uniform data set version 3: Application of machine learning models and a probability calculator. The Journal of Prevention of Alzheimer's Disease. 2023;10(2):301–13. 10.14283/jpad.2023.10. [ DOI ] [ PMC free article ] [ PubMed ] 51. Ren Y, Shahbaba B, Stark CEL. Improving clinical efficiency in screening for cognitive impairment due to Alzheimer’s. Alzheimers Dement Diagn Assess Dis Monit. 2023;15:e12494. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 52. Holm E, Jacobsen KK, De Lony TB, et al. Frequency of missed or delayed diagnosis in dementia is associated with neighborhood socioeconomic status. Alzheimers Dement Transl Res Clin Interv. 2022;8:e12271. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 53. Chen Y, Power MC, Grodstein F, et al. Correlates of missed or late versus timely diagnosis of dementia in healthcare settings. Alzheimers Dement. 2024;20:5551–60. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 54. Kivipelto M, Mangialasche F, Ngandu T. Lifestyle interventions to prevent cognitive impairment, dementia and Alzheimer disease. Nat Rev Neurol. 2018;14:653–66. [ DOI ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Material 1. (584.8KB, docx) Data Availability Statement The datasets used in this study are publicly available through the National Alzheimer’s Coordinating Center (NACC) database. Data can be accessed through the NACC Data Request process. Articles from Alzheimer's Research & Therapy are provided here courtesy of BMC ACTIONS View on publisher site PDF (2.6 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top