A lightweight machine learning tool for Alzheimer's disease prediction

Abstract

INTRODUCTION

Alzheimer's disease (AD) is a progressive neurodegenerative disorder that needs better predictive tools. Using the National Alzheimer's Coordinating Center Uniform Data Set, this study developed machine learning (ML) models and a practical clinical tool for AD prediction.

METHODS

Data from 52,537 individuals (22,371 with AD) and more than 200 variables were processed with MissForest imputation and genetic algorithm‐based selection. Multiple ML models were trained, and interpretability was performed using SHAP and permutation importance. A LightGBM model was refined through iterative backward feature elimination (IBFE) followed by manual refinement.

RESULTS

LightGBM performed best (receiver operating characteristic‐area under the curve [ROC‐AUC] 0.91, accuracy 82.0%). Key predictors included arthritis, age, body mass index, and heart rate. A 19‐feature model retained accuracy (81.2%) and ROC‐AUC (0.90).

DISCUSSION

This lightweight tool predicts AD using mostly routine variables. Limitations include its cross‐sectional nature, and would need external validation. An interactive web app and GitHub resource are available.

Highlights

• Developed a lightweight ML based tool using 19 routinely available features.

• The lightweight model achieved an ROC-AUC of 0.90 for Alzheimer's disease prediction on NACC multicenter data.

• Genetic algorithm, IBFE, and manual refinement enabled optimal feature selection.

• Tool hosted on an open‐access platform for clinical and research use.

• SHAP analysis provided model interpretability and feature‐level insights.

1. INTRODUCTION

Alzheimer's disease (AD) is a chronic neurodegenerative disorder  that is among the largest contributors to disability‐adjusted life years (DALYs) among neurological disorders and has transitioned from a clinical to a biologically defined entity. ¹ , ² Over 55 million individuals live with dementia, and a majority in low‐ and middle‐income countries (LMICs), with the numbers doubling every two decades and projected to reach 139 million by 2050, with marked regional and age‐specific variations, alongside a global financial burden for AD and related dementias approaching $16.9 trillion. ³ , ⁴ , ⁵ , ⁶ , ⁷ While abnormal biomarkers are sufficient to diagnose AD under the revised National Institute on Aging‐Alzheimer's Association (NIA‐AA) criteria, clinical evaluation and cognitive testing such as the Alzheimer's Disease Assessment Scale–Cognitive Subscale (ADAS‐Cog) remain valuable, and has shown predictive validity as a screening tool in both mild cognitive impairment and AD. ⁸ , ⁹ However, the absence of a curative treatment and reliance on post‐diagnostic management emphasize the need for strategies enabling earlier risk identification. ¹⁰

Traditional prediction models have incorporated demographics, lifestyle, cardiovascular, and genetic risk factors, analyzed through regression‐based approaches. ¹¹ , ¹² , ¹³ These methods are constrained by the limited ability to capture the heterogeneous interactions driving AD. In contrast, machine learning (ML) models can process high‐dimensional, multifactorial data and have shown promise in dementia research. ¹⁴ , ¹⁵ Common ML models include artificial neural networks (ANNs), capable of modeling non‐linear relationships ¹⁶ ; decision tree classifiers (DTCs) ¹⁷ ; and ensemble methods such as gradient boosting machines (GBMs), which have been used for AD. ¹⁸ , ¹⁹ , ²⁰ Notably, Light Gradient Boosting Machine (LightGBM) is optimized for speed, offering rapid training on large datasets. ²¹

Prior ML studies on AD prediction or classification, including those based on genetic data, have used relatively small or homogeneous cohorts. ²² , ²³ , ²⁴ Large multicenter analyses with externally relevant, easily screenable features remain limited. We therefore applied multiple ML models to the National Alzheimer's Coordinating Center (NACC) Uniform Data Set (UDS) to identify risk factors and develop a lightweight, interpretable clinical tool for AD prediction.

2. METHODOLOGY

2.1. Data source

We utilized the NACC's UDS, including all data entries available up to September 2024. Only the most recent visit data for each patient were considered to avoid redundancy and ensure consistency in cross‐sectional prediction modeling. The etiological diagnosis of AD followed standardized criteria: initially, the National Institute of Neurological and Communicative Disorders and Stroke (NINCDS) and the Alzheimer's Disease and Related Disorders Association (ADRDA, now the Alzheimer's Association [AA]) criteria and, in later versions, the National Institute on Aging (NIA) and the AA guidelines for AD dementia. Our primary objective was to identify and predict features contributing to the development of AD using ML techniques on this multicenter dataset. Figure 1 depicts the entire workflow, detailing the techniques used for handling the dataset and the modeling process.

FIGURE 1.

Workflow of data preprocessing, feature selection, and machine learning modeling for Alzheimer's disease prediction. Created in BioRender. Suresh, V. (2025) https://BioRender.com/siish3x.

2.2. Data preprocessing

2.2.1. Deduplication and initial inclusion

We retained only the latest visit for each individual to avoid repeated measurements. Following this, deduplication was performed to ensure each entry represented a unique individual.

2.2.2. Variable selection and cleaning

We included all variables from the following relevant categories in the NACC dataset: A1 Subject Demographics, A2 Co‐participant Demographics, A3 Subject Family History, A4 Subject Medications, A5 Subject Health History, B1 Physical Exam, B2 Health/History Information Section (HIS) and Cardiovascular History, D1 Clinician Diagnosis, and D2 Clinician‐Assessed Medical Conditions for model development.

This resulted in an initial feature space of more than 200 variables. During subsequent data cleaning, specific codes indicating missing, unknown, or irrelevant entries were uniformly converted to “Unknown” to streamline downstream handling. Variables with more than 50% missing data were excluded from further analysis. Additionally, variables that were unrelated or did not satisfy the logical cause‐and‐effect relationship with AD were removed (e.g., post‐diagnostic variables or those clearly downstream of diagnosis).

2.2.3. Encoding and imputation

Categorical variables were transformed using one‐hot encoding. To handle missing values, we implemented the MissForest imputation algorithm, a random forest (RF)‐based iterative imputation technique. This method, as demonstrated by Vinutha et al., has been shown to outperform methods such as Multiple Imputation by Chained Equations (MICE) when applied to the NACC dataset. ²⁵ Additionally, to address the class imbalance, we used the Synthetic Minority Oversampling Technique (SMOTE) to generate synthetic samples for the minority class.

2.2.4. Feature selection

To identify the most informative predictors of AD, we utilized a genetic algorithm‐based feature selection framework, inspired by the protocol suggested by Vinutha et al. ²⁵ The parameters used were as follows:

1. Population size: 50

2. Number of generations: 20

3. Crossover probability: 0.9

4. Mutation probability: 0.1

This evolutionary optimization approach helped identify a subset of variables that maximized classification performance across models.

2.2.5. ML modeling

We then evaluated six ML models on the processed dataset, including RF, ANN, Extreme Gradient Boosting (XGB), DTC, LightGBM, and LR.

RESEARCH IN CONTEXT

1. Systematic review: This study did not involve a systematic review or meta‐analysis. Instead, we developed and validated a novel ML‐based clinical tool through secondary analysis of the NACC dataset.

2. Interpretation: A lightweight LightGBM model using 19 features predicted Alzheimer's disease with an ROC‐AUC = 0.90.

3. Future directions: Prospective validation in primary care and memory clinic settings, integration into electronic health record systems, and assessment across international and underserved populations are warranted. Incorporating longitudinal data would help in temporal prediction and better clinical utility.

A fully connected feedforward ANN was implemented using TensorFlow's Keras Application Programming Interface (API) for binary classification tasks. The architecture comprised an input layer, followed by four hidden layers with 128, 64, 32, and 16 neurons, respectively. Each hidden layer employed ReLU activation, batch normalization, on the first three layers and progressively decreasing dropout for regularization to prevent overfitting. The output layer consisted of a single neuron with a sigmoid activation function. The Adam optimizer with a learning rate of 0.001 was used for training, alongside a binary cross‐entropy loss function. Additionally, class weights were computed to address the residual class imbalance in the dataset. An XGB model was trained using its inbuilt cross‐validation functionality. An RF model with 100 decision trees and a Decision Tree Classifier were both trained using the Gini impurity criterion for node splitting. Additionally, LightGBM, a fast and efficient gradient boosting framework, was optimized using Optuna for hyperparameter tuning across 30 trials to maximize performance. The dataset was split into training and test sets at an 80:20 ratio, with 30% of the training data further used as a validation subset during model fitting for the ANN.

To interpret model predictions, SHAP (SHapley Additive exPlanations) values or permutation importance were computed, depending on model type. SHAP summary and bar plots were generated to identify the most influential features. Each model was evaluated on the test set using accuracy, precision, recall, harmonic mean of precision and recall (F1‐Score), and receiver operating characteristic‐area under the curve (ROC‐AUC) score.

2.2.6. Clinical tool development

To facilitate real‐world application, we aimed to develop a lightweight, interpretable clinical tool. Using iterative backward feature elimination (IBFE) followed by manual refinement with a LightGBM model, we selected the smallest subset of features that maintained a model performance threshold of ROC‐AUC ≥ 0.90.

The final LightGBM model, trained on this optimized feature set, was exported using pickle and deployed via a Flask‐based web application.

3. RESULTS

3.1. Participant characteristics

Our study utilized data from the NACC's UDS from the September 2024 data freeze, collected across more than 40 centers in the United States. The data included a wide range of assessments, such as questionnaires covering lifestyle factors, health conditions, demographic information, family history, physical measures, biological sample assays, and imaging. After receiving approval and obtaining data from NACC, we carefully cleaned the data and included records of individuals with and without AD. The data included 22,371 diagnosed with Alzheimer's and 30,166 without. The demographic characteristics and other measures of the included Alzheimer's patients are detailed in Table S1. After applying feature selection using the parameters specified for the genetic algorithm, a total of 58 high‐performing features were identified and subsequently tested across various ML models.

3.2. Model performance

The comparative performance of the evaluated machine learning models is summarized in Table 1. The LightGBM model demonstrated the strongest performance, with an ROC‐AUC of 0.91, an accuracy of 82%, and balanced sensitivity of 77.95% and specificity of 85.02%. Its precision for the Alzheimer's class was 79.54%, and the F1‐score was 0.79, indicating discriminatory power and balanced prediction ability for both classes. Calibration analysis showed close agreement between predicted and observed probabilities (Brier score: 0.113), indicating well‐calibrated estimates.

TABLE 1.

Comparative performance of various machine learning models.

Metric	Artificial neural network	Extreme gradient boosting	Random forest	Decision tree classifier	Naive bayes	Logistic regression	LightGBM
Accuracy	67.00%	82.29%	79.37%	72.65%	54.90%	64.83%	82.00%
Sensitivity (Recall)	53.00%	77.64%	73.25%	68.72%	76.92%	64.02%	77.95%
Specificity	77.14%	85.77%	83.94%	75.59%	38.50%	65.43%	85.02%
Precision (PPV)	63.40%	80.30%	77.31%	67.71%	48.22%	58.05%	79.54%
Negative Predictive Value (NPV)	72.52%	84.39%	81.11%	76.44%	69.14%	71.12%	83.62%
F1‐Score	0.66	0.79	0.75	0.68	0.59	0.61	0.79
ROC‐AUC Score	0.7187	0.9096	0.8807	0.7215	0.6443	0.7062	0.9100

Abbreviations: LightGBM, Light Gradient Boosting Machine; NPV, negative predictive value; PPV, positive predictive value; ROC‐AUC, receiver operating characteristic‐area under the curve.

The XGB model closely followed, achieving an accuracy of 82.29%, with a precision of 80.30% and recall of 77.64%, and an F1‐score of 0.79 for the Alzheimer's class. It achieved an ROC‐AUC of 0.9096 and a specificity of 85.77%.

The RF classifier achieved an accuracy of 79.37% and an ROC‐AUC score of 0.8807. It showed a precision of 77.31%, a recall of 73.25%, a specificity of 83.94%, and an F1‐score of 0.75 for the Alzheimer's class.

The DTC achieved an accuracy of 72.65%, with a precision of 67.71%, a recall of 68.72%, a specificity of 75.59%, an ROC‐AUC of 0.7215, and an F1‐score of 0.68 for the Alzheimer's class. While it performed better than ANN and Naïve Bayes Traditional (NBT), it was outperformed by ensemble models like XGB and LightGBM.

ANN performed less robustly, with an accuracy of 67%. It showed a lower recall of 53.00%, a precision of 63.40% for the Alzheimer's class, and a ROC‐AUC score of 0.7187, indicating limited sensitivity despite a specificity of 77.14%. The model's weighted F1‐score was 0.66.

The LR model had an accuracy of 64.83%, a precision of 58.05%, a recall of 64.02%, and an F1‐score of 0.61 for the Alzheimer's class. It achieved an ROC‐AUC of 0.7062 and a specificity of 65.43%.

The NBT model performed the worst, with an accuracy of only 54.90%. While it had a recall of 76.92%, its specificity was very low, 38.50%. It had a precision of 48.22%, an F1‐score of 0.59, and a ROC‐AUC score of 0.6443.

Figures 2A–H depict all the confusion matrices of the models.

FIGURE 2.

Confusion matrices. (A) Confusion matrix–artificial neural network, (B) confusion matrix–extreme gradient boosting, (C) confusion matrix–random forest, (D) confusion matrix–decision tree classifier, (E) confusion matrix–Naive Bayes, (F) confusion matrix–logistic regression, (G) confusion matrix–light gradient boosting, (H) confusion matrix–lightweight clinical tool.

Figures 3A–H depict all the ROC curves.

FIGURE 3.

Receiver operating characteristic (ROC) curves. (A) ROC curve–artificial neural network, (B) ROC curve–extreme gradient boosting, (C) ROC curve–random forest, (D) ROC curve–decision tree classifier, (E) ROC curve–Naive Bayes, (F) ROC curve–logistic regression, (G) ROC curve–light gradient boosting, (H) ROC curve–lightweight clinical tool.

Feature importance for each model can be visualized in Figures 4A–H.

FIGURE 4.

Feature Importances. (A) Feature (permutation) importances–artificial neural network, (B) feature importances–extreme gradient boosting, (C) feature importances–random forest, (D) feature importances–decision tree classifier, (E) feature importances–Naive Bayes, (F) feature importances–logistic regression, (G) feature importances–light gradient boosting, (H) feature importances–lightweight clinical tool.

3.3. Lightweight clinical tool

Further feature selection through IBFE and manual refinement ultimately yielded a set of 19 features (after an initial minimized set of 58 features). The retained features and their corresponding feature importance scores are presented in Table S2.

The final LightGBM model after hyperparameter tuning using Optuna achieved an accuracy of 0.8120 and an AUC of 0.90. The model demonstrated a specificity of 84.0%, a sensitivity of 77.4%, positive predictive value (PPV) of 77.9%, and negative predictive value (NPV) of 83.6%, with balanced performance across classes as reflected by precision, recall, and F1‐scores (macro‐average: 0.81). Calibration analysis demonstrated close alignment between predicted probabilities and observed outcomes, with a Brier score of 0.126 for the uncalibrated model and 0.125 after isotonic calibration, indicating well‐calibrated estimates. The magnitude of feature importance and individualized contributions to model predictions, as visualized through SHAP values, are shown in Figures S1 and S2, respectively. The model was hosted on a GitHub repository and deployed on Streamlit, enabling use in clinical settings. ²⁶ , ²⁷

4. DISCUSSION

Using data from over 52,000 individuals across more than 40 centers, gradient boosting approaches outperformed other models, with LightGBM achieving the best balance of accuracy (82.0%), ROC‐AUC (0.91), sensitivity, and specificity. By contrast, ANNs performed modestly, consistent with prior evidence that deep learning is less suited to structured tabular data, whereas boosting algorithms efficiently capture non‐linear interactions, accommodate mixed variable types, and train rapidly. ²⁸ , ²⁹ , ³⁰ , ³¹

Feature importance analyses identified spinal arthritis, age, body mass index (BMI), heart rate, weight, height, depression, and stroke as key predictors. Arthritis warrants attention: Over 90% of individuals with arthritis in our cohort had osteoarthritis (OA), and studies have identified associations between OA and dementia, wherein chronic inflammation may contribute to shared pathogenic pathways. ³² , ³³ , ³⁴ Reverse causality is also plausible, as lack of physical activity has been associated with dementia, ³² which could predispose to arthritis. Moreover, limited mobility may restrict adherence to healthcare and access to services, complicating the interpretation of the association. A Mendelian randomization study suggests heterogeneity: Hip OA has been associated with reduced dementia risk, Kneehip OA inversely associated with unspecified dementia, and ankylosing spondylitis identified as an independent risk factor for dementia. ³⁵ Large cohort studies similarly reported associations between spondylosis and AD risk, ³⁶ while rheumatoid arthritis appears protective, possibly due to anti‐inflammatory treatments. ³⁷ , ³⁸ These mixed findings emphasize the need for future work to disentangle causal effects from confounding by treatment or physical activity.

Other predictors aligned with established epidemiology. Older age and low BMI positively contributed to AD diagnosis predictions, a relationship consistent with findings from large cohort studies. ³⁹ , ⁴⁰ Some studies vary, and some suggest that a higher BMI or obesity leads to an increased risk of dementia, suggesting a complex relationship. ⁴¹ , ⁴² , ⁴³ Active depression in the last 2 years was strongly associated with AD diagnosis in our model, suggesting it may represent both a prodromal manifestation and a long‐term risk factor. Heart rate showed weaker effects: While our model mostly associated lower resting rates with AD, other studies have suggested no associations or causal links. ⁴⁴ , ⁴⁵ Stroke and seizures also contributed positively, reflecting overlap with vascular comorbidity. This is consistent with the cohort study by Honig et al, which reported an association between stroke and AD among older individuals. ⁴⁶

A critical innovation of our model is the deliberate exclusion of post‐diagnostic features such as cognitive or memory scores. While prior NACC‐based ML models have relied heavily on clinician‐administered assessments to achieve high classification accuracy, ⁴⁷ , ⁴⁸ our approach uses fewer and more routinely collected variables, improving feasibility for early screening, especially in resource‐limited settings. Our model instead relies on 19 demographic, clinical, and history variables, most of which are easily obtainable at first visits, enhancing practicality and translational value. SHAP analysis further supports interpretability, allowing clinicians to understand the contribution of individual features to diagnostic predictions.

The strengths of our work include the use of a large multicenter dataset, comparison across multiple ML algorithms, feature selection using genetic optimization and recursive elimination, assessment of calibration, and development of a lightweight tool deployed via Streamlit and GitHub for real‐world use. All selected features are routinely available in clinical practice and history taking, ensuring feasibility. Reproducibility of ML pipelines can be influenced by several sources of variability. In our case: data imputation, feature selection using the genetic algorithm, class balancing through SMOTE oversampling, and feature importance and direction as per SHAP analysis could potentially contribute to such variations. To assess the potential impact of imputation‐related variability, we re‐ran the entire pipeline five times, performing data imputation independently in each run for the LightGBM model. Additionally, we excluded OTHCOND ‐ representing other medical conditions or procedures within the past 12 months, as its broad and heterogeneous nature could introduce ambiguity, and confirmed that its removal did not materially alter the model's performance. Across these experiments, model performance remained within a reasonably stable range, with LightGBM accuracy varying from 0.777 to 0.824 (AUC: 0.86‐0.91), and the lightweight tool ranging from 0.737 to 0.804 (AUC: 0.81‐0.88). Furthermore, some variables were initially coded such that 1 = recent/active and 2 = remote/inactive, and were numerically switched to maintain a consistent directional order for modeling. This conversion was applied solely for coding consistency and did not necessarily imply severity. A sensitivity analysis was subsequently performed by recoding these variables into binary presence/absence states, which resulted in a further minor reduction in performance (LightGBM AUC: 0.83; lightweight tool AUC: 0.80) indicating that the original ordinal coding provided slightly greater discriminative stability. Limitations include residual class imbalance despite SMOTE, variations in the output of feature selection techniques like the genetic algorithm, potential bias from missing values even after MissForest imputation, possible variations with each iteration, SHAP analysis, and feature importances. Minor variations in performance metrics and confusion matrices can be noted across repeated runs. Measurements such as blood pressure or heart rate were recorded at a single time point, which may not fully reflect chronic exposures. As the NACC is clinic‐based, selection bias may limit generalizability to community or LMIC populations.

Consistent with NACC guidance, the dataset's heterogeneity in recruitment strategies and incomplete exposure histories may constrain its use for evaluating risk factors for dementia. Finally, as the models were trained on cross‐sectional data, they predict AD diagnosis and should not be interpreted as predictors of future risk trajectories.

5. CONCLUSIONS AND FUTURE DIRECTIONS

We developed a lightweight LightGBM‐based model that predicts AD using 19 routinely available, mostly clinical, history, and demographic variables. The model achieved high accuracy, was well calibrated, and offers practical interpretability through SHAP analysis. Unlike prior approaches that relied on post‐diagnostic cognitive measures, our tool is designed for feasibility at the first clinical encounter. Nonetheless, its cross‐sectional nature, clinic‐based dataset, and reliance on imputation warrant cautious interpretation, and results may not directly generalize to community or LMIC settings. Future work should focus on external and prospective validation, incorporation of longitudinal data, and careful consideration of ethical and regulatory requirements before clinical deployment.

AUTHOR CONTRIBUTIONS

Vinay Suresh: Conceptualization; methodology; investigation; data curation; formal analysis; visualization; project administration; writing—original draft, writing—review and editing. Tulika Nahar: Writing—original draft; writing—review and editing. Arkansh Sharma: Writing—original draft; writing—review and editing. Suhrud Panchawagh: Formal analysis; software; validation; writing—review and editing. Omer Mohammed: Investigation; writing—review and editing. Muneeb Ahmad Muneer: Writing—original draft; writing—review and editing. Devansh Mishra: Writing—original draft; writing—review and editing. Amogh Verma: Supervision; validation; data curation; investigation; writing—original draft; writing—review and editing. Vivek Sanker: Supervision; writing—review and editing. Ayush Mishra: Formal analysis; software; validation; writing—review and editing. Hardeep Singh Malhotra: Supervision; conceptualization; writing—review and editing. Ravindra Kumar Garg: Supervision; conceptualization; writing—review and editing.

CONFLICT OF INTEREST STATEMENT

The authors have no relevant conflicts of interest to disclose. The author disclosures are available in Supporting Information.

GENERATIVE AI USE STATEMENT

Paperpal and ChatGPT, were used for language refinement, technical editing and cleaning the code for analysis. These tools did not contribute to study design, or substantive content development. All scientific decisions and analyses were performed and validated by the authors, who take full responsibility for the content.

CLINICAL TRIAL REGISTRATION DETAILS/NUMBER

Not applicable, as this study does not report a clinical trial.

RESEARCH REGISTRY NUMBER

Not applicable.

HUMAN ETHICS APPROVAL DECLARATION:

As our research is secondary in nature and based on de‐identified data from the NACC Uniform Data Set, the requirement for informed consent and institutional ethics board approval was not applicable.

Supporting information

Table S1: In the first row, switch the values so that:

30,166 = without AD

22,371 = with AD

Supporting Information

Suresh V, Nahar T, Sharma A, et al. A lightweight machine learning tool for Alzheimer's disease prediction. Alzheimer's Dement. 2025;17:e70187. 10.1002/dad2.70187

DATA AVAILABILITY STATEMENT

Data supporting this study were obtained with formal permission from the NACC. The code used for the analysis is available from the corresponding author upon reasonable request.