AD-istics: A Machine Learning Framework for Optimized Detection of Dementia and Alzheimer's Disease

Transcription:

Dementia and Alzheimer's disease are leading conditions affecting millions of seniors today. We often struggle to receive an effective diagnosis for various reasons such as difficulties in accessibility, time, cost, and accuracy, and they pose many strains upon our healthcare system.

I want to solve these problems, and I want to introduce AD-istics. AD-istics is a dual-parallel machine learning framework that takes in two types of data: clinical data and MRI image data. AD-istics consists of two machine learning models, an XGBoost model and a Vision Transformer, all combined under one software interface.

With AD-istics, I hope to help seniors who have essentially raised our generation today and take a lot of the stress off of our doctors' shoulders. I hope to help doctors. I hope to create a tool for doctors.

I am excited to meet everyone at CWSF. Thank you.

Personal Reflection:

I have been studying dementia and Alzheimer's disease since I was 15. The conditions are devastating, often consuming other people and causing them to lose themselves. I have always feared my mother forgetting who I am or what my name is, fueling my motivation to explore these topics as deeply as possible.

Problem Statement:

In many healthcare systems today, people living with dementia/Alzheimer's disease (AD) often face barriers in receiving an effective diagnosis due to long wait times, high costs, and limited accuracy of detections.

According to research, significant proportions of AD patients go misdiagnosed solely based on traditional diagnostic methods such as clinical assessments (Dokholyan et al., 2022). Although current diagnostic techniques are advancing significantly, accessibility poses immense concern. 75% of all dementia cases go undiagnosed across the globe, up to 90 percent in low- and middle-income countries (Alzheimer’s Disease International, n.d.). This problem is often attributed to limited accessibility of healthcare professionals (Figure 1; Bassetti et al., 2024). If an individual can receive a diagnosis, “it may take several appointments and tests over many months before a diagnosis of Alzheimer's disease can be confirmed” (National Health Service [NHS], n.d.), figure 2 (Author et al., 2008). If a diagnosis can be made after meeting these three conditions, the overall healthcare burden and costs only add to the problem (Figure 3 (Nandi et al., 2024)).

I set out to solve these problems and create an effective solution to dementia detection.

Parallel Machine Learning Framework:

To create an effective solution, a two-way AI framework mimicking real dementia detection was developed.

Workflow: There is a clinical side accepting clinically administered data and an image side accepting MRI scan data (brain scans). A dual parallel workflow is used for applicability, credibility, and further accessibility, giving users options when being tested.

Clinical: An AI model was trained on a wide range of clinical data for the detection of Alzheimer's disease. This model is called XGBoost (Figure 2) (Mustafa & Ahmad, 2024). The goal is to aid those with limited access to neurologists but with access to healthcare professionals trained in administering necessary tests for ease of diagnosis.

Image: An additional AI model was trained on a dataset of brain MRI scans. This model is called a visual image transformer (ViT) (Figure 3) (Shah, 2022). This model allows for a dementia diagnosis with simply one brain MRI scan.

Data Collection:

Clinical data was taken from a dataset on Kaggle, collected from 2,149 patients with various features

I extracted and used important features for ease of AD-istics.

MRI data was retrieved from an open OASIS dataset on Kaggle, consisting of over 200 slices of 416 patients with various levels of dementia.

Careful steps were taken to prevent data leakage. Figure 4

Testing the Framework:

Clinical: An 80%/20% train/test split and fivefold cross-validation were used to test the XGBoost model (figure 5 (scikit-learn developers, n.d.)) -

Image: A 60%/20%/20% train/validation/test split and fivefold cross-validation was used to test the ViT

K-fold cross validation allows for the more accurate representations of a model's metrics

Framework Optimization:

The image model was the main focus in this area; various methods of model optimization were taken to increase generalizability, mitigate limitations, and obtain ideal metrics.

The Prototype/Application:

The machine learning framework was combined under a software interface for demonstration to depict how the design would be used in the real world. This was done utilizing the Python library Streamlit (Figure 1).

The app receives an input from a healthcare professional (either clinical or image-dependent on the tab) and outputs a prediction of whether the patient is demented or non-demented or has AD or non-AD with probability scores.

Confidence stratification was implemented when the model had little confidence in specific patients (when probabilities are close to 50/50).

I recognize that a limitation often employed on XGBoost models is difficulty in extrapolating outside of the training data range. Thus, a range analysis (Figure 2) was conducted and showed that the data values of patients within the used dataset account for a majority of patients in the real world. However, as an extra precaution, a verification was implemented, indicating further evaluation is needed if data is out of range.

Visual Image Transformer Results (Matrix Reports):

Results of test set in the 60/20/20 dataset split:

Non-Demented: Percision: 96% | Recall: 81% | f1-score: 88%

Demented: Precision: 76% | Recall: 94% | f1-score: 84%

Accuracy: 86%

AUC: 0.878 (Figure 3)

Results of 5-fold cross-validation:

Non-Demented: Precision: 92% ± 2.3% | Recall: 80% ± 4.6% | f1 Score: 85.4% ± 3.2%

Demented: Precision: 73% ± 4.3% | Recall: 89.2% ± 2.7% | f1 score: 80% ± 3.2%

Accuracy: 83.2% ± 3.3%

XGBoost Matrix Results (Matrix Reports):

Results of test set in the 80/20 split:

Non-AD: Precision: 95% | Recall: 97% | f1 Score: 96%

AD Positive: Precision: 94% | Recall: 90% | f1 Score: 92%

Accuracy: 95%

AUC: 0.937

Results of 5-fold cross-validation:

Non-AD: Precision: 95.0% ± 1.4% | Recall: 96.6% ± 0.89% | F1 Score: 96.2% ± 1.3%

AD Positive: Precision: 93.8% ± 1.6% | Recall: 90.8% ± 3.1% | F1 Score: 92.4% ± 2.3%

Accuracy: 95.0% ± 0.96%

Discussion:

General:

Overall, despite dealing with skewed datasets (i.e., Figure 4) and limited data, both models had strong performance.

Interestingly, both models complement one another, demonstrating potential for future ensemble approaches. The clinical model demonstrates a higher AD precision yet lacks relative recall, and the image model lacks dementia precision relative to dementia recall.

Image-based model

Considering the sole input was MRI images, the image-based model performed strongly.

65%-95% of MRI scans are correctly classified with AD diagnoses. The MRI-based ViT is deemed to be effective in both saving time with relatively similar accuracies in these diagnoses (Klöppel et al., 2008).

During the developing phase, I had to decide whether I should focus on non-demented or demented recall due to the tradeoff. After careful consideration, I decided a greater false-positive rate needing further evaluation may be prioritized over a greater false-negative rate.

Clinical model:

Performed well on given data with minimal deviations

In reference to the statistic in the "why" section regarding difficulties in developing a clinical diagnosis solely based on clinical data, the XGboost model deems beneficial

Further Discussion/Result Analysis:

Overall, the prototype of AD-istics solves the initial problem I had in research in creating an effective solution for the detection of dementia. ADistics takes seconds to formulate predictions and costs minimal and is widely accessible, being a software interface, and overall is deemed to be user-friendly to help neurologists and doctors diagnose dementia and AD with ease.

The system works effectively in creating an effective solution.

Clinical (Tabular) Model:

Relevant to specific results of the clinical model, the clinical model performs with strong metrics and minimal deviations on the 5-fold cross-validation, representing strong and stable performance with optimistic results in the real world (considering the wide range of values in the dataset).

Image Model:

The image-based model-specific results needs further reflection

The deviations overall in 5-fold cross-validations are relatively higher compared to the clinical model (less stability of results across folds). This is likely a result of the limited dataset size (416 patients), resulting in immense variability (± patient differences of results).

The model struggles with predicting dementia correctly (low dementia precision) relative to all of its predictions. and struggles to outline all non-demented cases in the test dataset (low non-demented recall). The tradeoff with non-demented percision and demented recall was considered.

The model generalizes well, performing well on 5-fold validation and have stable convergence and decreasing values between train loss and validation loss throughout epochs trained

Taken together, AD-istics has strong potential in aiding doctors in diagnosing dementia/AD and helping many seniors.

Future Works on AD-istics:

In the future, I hope to utilize the compatibility of the two models to create an ensemble model that could additionally run in parallel, solely working based on restricted data it is provided.

Improve image model's metrics and robustness for further medical application

Incorporate clinical explainability using Grad-CAM (heat maps of MRI) (Figure 1) (Sertkaya et al., 2024) and SHAP analysis.

Addition of other domains (such as DARWIN writing dataset) for further application (figure 2 is annotated) (Gallagher, 2020)

Applying AD-istics:

Incorporating real, clinical testing of AD-istics to understand generalizability and applicability

-I would like to sincerely thank my mentor, Gaumit Kauts, for his insights and teachings on choosing machine learning models and help with tips on model optimization, understanding aspects of machine learning and analysis, and his overall support in being an absolutely amazing mentor. I truly appreciate you.

-I would like to deeply thank the Edmonton Science Fair Committee and judges for their feedback to improve my project. Without them, this opportunity to present at CWSF would not be possible.

-I would additionally like to thank my family for their endless support and resources they have provided me and for always being by my side when things were difficult and stressful. It means a lot to me.

-Finally, I would like to thank my high school, Lillian Osborne High School. My immense fascination for STEM could have never been developed without my amazing teachers and support from school staff.

(N.d.). AI Neural Network Visualization in Human Brain. Retrieved April 25, 2026, from https://www.presentationgo.com/presentation/ai-neural-network-brain/.

2024 alzheimer’s disease facts and figures - 2024 - alzheimer’s & dementia - wiley online library. (n.d.-a). https://alz-journals.onlinelibrary.wiley.com/doi/10.1002/alz.13809

3.1. Cross-validation: evaluating estimator performance. (n.d.). Retrieved April 29, 2026, from https://scikit-learn.org/stable/modules/cross_validation.html.

Accorroni, A., Arnesen, A., & Vuletic, V. (2024). General neurology: Current challenges and future implications. Retrieved April 29, 2026,.

Armstrong, J. (2022). Navigating the Path Forward for Dementia in Canada: The Landmark Study Report #1. Retrieved April 29, 2026, from https://alzheimer.ca/en/research/reports-dementia/navigating-path-forward-landmark-report-1.

Boesch, G. (2023). Vision Transformers (ViT) in Image Recognition. https://viso.ai/deep-learning/vision-transformer-vit/. Retrieved April 29, 2026,.

Crouch, M. (2022). Concerned About Alzheimer’s Disease? How to Get an Accurate Diagnosis. Retrieved April 29, 2026, from https://www.aarp.org/health/conditions-treatments/getting-accurate-alzheimers-diagnosis/.

Dokholyan, N. V., Mohs, R. C., & Bateman, R. J. (2022). Challenges and progress in research, Diagnostics, and therapeutics in alzheimer’s disease and related dementias. Alzheimer’s & Dementia: Translational Research & Clinical Interventions, 8(1). https://doi.org/10.1002/trc2.12330

Day, G. (2022). Alzheimer Disease Biomarkers in Clinical Practice: A Blood-Based Diagnostic Revolution. Retrieved April 29, 2026, from https://www.researchgate.net/figure/Diagnostic-challenges-in-Alzheimer-disease-The-accurate-diagnosis-of-Alzheimer-disease_fig1_366097172.

Early-onset alzheimer’s disease | johns hopkins medicine. (n.d.-b). https://www.hopkinsmedicine.org/health/conditions-and-diseases/alzheimers-disease/earlyonset-alzheimer-disease

freeCodeCamp.org (n.d.-c). YouTube. https://www.youtube.com/watch?v=JwSS70SZdyM

Gallagher, W. (2020). Future Apple Pencil may feature full touch-sensitive controls. Retrieved April 21, 2026, from https://appleinsider.com/articles/20/01/09/future-apple-pencils-may-feature-full-touch-sensitive-controls.

Klöppel S;Stonnington CM;Barnes J;Chen F;Chu C;Good CD;Mader I;Mitchell LA;Patel AC;Roberts CC;Fox NC;Jack CR;Ashburner J;Frackowiak RS; (n.d.). Accuracy of dementia diagnosis: A direct comparison between Radiologists and a computerized method. Brain : a journal of neurology. https://pubmed.ncbi.nlm.nih.gov/18835868/#:~:text=Abstract,specificity:%2080/85

Liu, S., Masurkar, A. V., Rusinek, H., Chen, J., Zhang, B., Zhu, W., Fernandez-Granda, C., & Razavian, N. (2022). Generalizable deep learning model for early alzheimer’s disease detection from structural mris. Scientific Reports, 12(1). https://doi.org/10.1038/s41598-022-20674-x

Lianne and Justin (n.d.-b). YouTube. https://www.youtube.com/watch?v=aLOQD66Sj0g

Mayo Foundation for Medical Education and Research. (n.d.). Diagnosing alzheimer’s: How alzheimer’s is diagnosed. Mayo Clinic. https://www.mayoclinic.org/diseases-conditions/alzheimers-disease/in-depth/alzheimers/art-20048075

Mustafa, R., & Ahmad, T. (2024). Network architecture of XGBoost. Retrieved April 20, 2026,.

Nandi, A., Counts, N., Bröker, J., Malik, S., Chen, S., Han, R., Klusty, J., Bloom, D., Vigo, D., Tortorice, D., & Seligman, B. (2024). Projected cost of ADRDs formal and informal care in the United States, 2020 to 2060. Retrieved April 29, 2026, from https://pubmed.ncbi.nlm.nih.gov/38331952/.

Prevalence and incidence. Dementia Statistics Hub. (2024, November 5). https://dementiastatistics.org/about-dementia/prevalence-and-incidence/.

Sertkaya, M., Ergen, B., türkoğlu, M., & Tonkal, Ö. (2024). Accurate diagnosis of dementia and Alzheimer’s with deep network approach based on multi‐channel feature extraction and selection. Retrieved April 29, 2026,.

Single Brain Scan could diagnose alzheimer’s disease | technology networks. (n.d.-c). https://www.technologynetworks.com/neuroscience/news/single-brain-scan-could-diagnose-alzheimers-disease-362899/

Speechly, C., Bridges-Webb, C., & Passmore, E. (2008). The pathway to dementia diagnosis. Retrieved April 24, 2026, from https://www.mja.com.au/journal/2008/189/9/pathway-dementia-diagnosis.

Uygar Kurt (n.d.-a). YouTube. https://www.youtube.com/watch?v=Vonyoz6Yt9c.

Waiting your turn: Wait times for health care in Canada, 2024 report | Fraser Institute. (n.d.-d). https://www.fraserinstitute.org/studies/waiting-your-turn-wait-times-for-health-care-in-canada-2024

XGBoost explained: A beginner’s guide | by Jamie Crossman-Smith | low code for data science | medium. (n.d.-e). https://medium.com/low-code-for-advanced-data-science/xgboost-explained-a-beginners-guide-095464ad418f