Canada

Dementia is largely underdiagnosed in early stages because cognitive and behavioural symptoms are often misattributed to aging. With global dementia cases on the rise, artificial intelligence offers a promising solution to support early detection. This study assesses whether easy-to-interpret white-box models can achieve diagnostic accuracy comparable to that of complex black-box models in dementia screening, striving to increase transparency without compromising predictive accuracy. With a sample of 358 adults from the OASIS-1 and OASIS-2 datasets, we trained white-box and black-box machine learning models on six clinical and demographic predictors of dementia: Mini Mental State Examination (MMSE), age, sex, education, socioeconomic score, and normalized whole-brain volume (nWBV). The Black-box models included Random Forest, XGBoost, and k-Nearest Neighbours, and the white-box model used LASSO-regularized logistic regression. AI Models were evaluated using accuracy, sensitivity, specificity, positive predictive values, and negative predictive values for both binary and multiclass classification tasks. In general, black-box models performed slightly better; however, the white-box model achieved a notably strong binary sensitivity. This suggests that white-box models were particularly effective in identifying individuals with cognitive impairment based on their clinical and demographic characteristics. While white-box models did not perform as strongly as black-box models, especially in terms of multiclass accuracy, they may still be helpful in healthcare, as their transparency enables physicians to interpret and verify predictions. The study&#39;s findings emphasize the trade-off between performance and interpretability.

Aging Powerfully: AI Blueprint for Chronic Disease Management and Social Determinants of Longevity

Comparing White-Box Versus Black-Box Predictive AI Models for Dementia Using OASIS Data

oasis

dementia diagnosis

ai interpretability

black-box

white-box

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. AAAI-26 will feature technical paper presentations, special tracks, invited speakers, workshops, tutorials, poster sessions, senior member presentations, competitions, and exhibit programs, and a range of other activities to be announced.<br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to AAAI 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://aaai.org/conference/aaai/aaai-26/registration/) first.

**Online Registration Form**: https://aaai.getregistered.net/conference-2026 

Registration Required

AAAI 2026

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

Dementia is largely underdiagnosed in early stages because cognitive and behavioural symptoms are often misattributed to aging. With global dementia cases on the rise, artificial intelligence offers a promising solution to support early detection. This study assesses whether easy-to-interpret white-box models can achieve diagnostic accuracy comparable to that of complex black-box models in dementia screening, striving to increase transparency without compromising predictive accuracy. With a sample of 358 adults from the OASIS-1 and OASIS-2 datasets, we trained white-box and black-box machine learning models on six clinical and demographic predictors of dementia: Mini Mental State Examination (MMSE), age, sex, education, socioeconomic score, and normalized whole-brain volume (nWBV). The Black-box models included Random Forest, XGBoost, and k-Nearest Neighbours, and the white-box model used LASSO-regularized logistic regression. AI Models were evaluated using accuracy, sensitivity, specificity, positive predictive values, and negative predictive values for both binary and multiclass classification tasks. In general, black-box models performed slightly better; however, the white-box model achieved a notably strong binary sensitivity. This suggests that white-box models were particularly effective in identifying individuals with cognitive impairment based on their clinical and demographic characteristics. While white-box models did not perform as strongly as black-box models, especially in terms of multiclass accuracy, they may still be helpful in healthcare, as their transparency enables physicians to interpret and verify predictions. The study's findings emphasize the trade-off between performance and interpretability.

technical paper

2025 National Inter-University Health Data & AI Inquiry Program - this program is a unique project-based inquiry and science communication program that brings together students from diverse academic backgrounds and research interests. Presented research projects explore how generative AI is being applied to medication management, mental health screening, and chronic disease monitoring to align with the future of pharmacy practice.

2025 National Inter-University Health Data & AI Inquiry Program is a unique project-based inquiry and science communication program that brings together students from diverse academic backgrounds and research interests. 

This research leverages deep learning techniques to investigate the connectivity between social determinants of health (SDoH) and chronic disease-related mortality and life expectancy globally. We integrated multiple datasets, including World Bank socioeconomic indicators, WHO global health estimates, and income inequality measures, spanning from 2000 to 2021. Using exploratory data analysis, we identified global trends and significant correlations between five chronic disease outcomes, Alzheimer’s, asthma, cardiovascular diseases, diabetes, and kidney diseases, and various socioeconomic factors. We implemented multiple predictive models, including linear regression, XGBoost, and recurrent neural networks, including Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM), validated using 3-fold cross- validation. The parameters are selected using hyperparameter tuning techniques. XGBoost consistently outperformed other methods across diseases, achieving superior accuracy as measured by mean squared error (MSE) and coefficient of determination (R²). To ensure interpretability, SHAP (SHapley Additive exPlanations) values were applied, highlighting influential socioeconomic factors like income inequality, diabetes prevalence, and life expectancy. Our analysis revealed distinct patterns based on income groups and geographic regions, emphasizing the necessity of specific public health strategies. This research uniquely contributes by integrating interpretable AI techniques, providing actionable insights into the socioeconomic drivers of chronic diseases, and guiding effective policy and intervention design aimed at promoting healthier aging populations worldwide.

Deep Learning Insights into Social Determinants of Chronic Disease and Longevity

Early, non-invasive detection of Alzheimer’s Disease (AD) significantly benefits public health planning, yet existing diagnostic tools are invasive, expensive, or clinic-dependent. This research develops a machine learning (ML) pipeline leveraging spontaneous speech, an accessible biomarker, using the balanced ADReSS 2020 dataset. We extracted 22 acoustic-prosodic features (including MFCCs, spectral centroid/bandwidth, RMS energy, fundamental frequency) combined with demographic metadata. After winsorizing outliers and applying standard scaling, four supervised ML algorithms (Gaussian Naïve Bayes, Logistic Regression, Support Vector Machine, and Gradient Boosting) were benchmarked. Optimal binning, novel to AD speech analysis, significantly enhanced Gradient Boosting performance, yielding 81.97% accuracy and ROC-AUC of 89.54%, outperforming other models by 7-30%. Feature analysis revealed pause- related spectral and pitch variability as key discriminators. Results indicate computationally efficient gradient boosting algorithms combined with optimal binning outperform complex deep learning models in interpretability and clinical feasibility, providing a strong foundation for practical, low-cost AD screening.

Detection of Alzheimer’s Dementia from Spontaneous Speech Using Gradient‑Boosted Classification

Introduction/Background: Chronic diseases and longevity are influenced by social determinants of health (SDOH) in urban environments where aging populations are more vulnerable to external stressors. This project examined how variations in SDOH impact disease risks and healthy life expectancy, and how these results could guide planning for healthier communities. Materials and Methods: We integrated records from the U.S. Census American Community Survey (ACS), Centers for Disease Control and Prevention (CDC), Environmental Protection Agency (EPA) AirNow, and United States Department of Agriculture (USDA) Food Access databases into a unified dataset covering 6,607 entries. Twelve features captured key SDOH factors relevant to aging. Machine learning models were trained to predict a community health index reflecting chronic disease prevalence and healthy aging. Model performance was validated using R², Mean Squared Error (MSE), and external test sets. Key Findings/Results: The ANN achieved excellent predictive accuracy (R² = 0.9998). Feature importance analyses identified housing cost burden, broadband access, and crowding as top predictors of community health outcomes. An interactive digital twin platform enables users to simulate policy interventions and observe changes in disease risk and longevity. Conclusion: This study introduces an innovative approach that integrates multi source data and machine learning to guide urban and health planning. By enabling testing of interventions, it offers a new way to design environments that help aging populations live longer, healthier lives. Future research could build on this by incorporating real time data and causal models to enhance impact.

Digital Twin Cities: Simulating Aging Outcomes By Modifying Social Determinants of Health

Neurodegenerative diseases such as Parkinson’s Disease (PD) and Alzheimer’s Disease (AD) are highly prevalent in aging populations, but current diagnostic processes remain invasive, costly, and inaccessible. This project proposes a novel, non-invasive, and interactive screening tool that takes handwriting data as input and implements a machine learning model to assess the risk of AD and PD. Handwriting data from the DARWIN dataset (n=174) was used to classify Alzheimer’s patients using logistic regression, feedforward neural networks (FNN), and random forest classifiers, with the random forest achieving the highest accuracy of 85.29%. For Parkinson’s detection, we employed spiral and wave drawings from multiple datasets and trained a convolutional neural network (CNN) using transfer learning with MobileNetV2, reaching a test accuracy of 81.25%. A graphical user interface (GUI) was developed to gamify the data collection process and deliver real-time predictions. The results are promising and suggest that drawing-based assessments can become a more accessible solution for early screening of neurodegenerative diseases in clinical and home environments.

Drawing To Diagnose: A Gamified Diagnostic Tool For Alzheimer’s And Parkinson’s Disease

Early detection of Alzheimer's disease and related neurological disorders is crucial for timely intervention and improved patient outcomes. Traditional diagnostic approaches often require extensive clinical assessments, creating barriers to widespread screening in aging populations. This study aims to develop and evaluate machine learning models for predicting Alzheimer's disease and neurological disorders using reduced clinical questionnaires enhanced through Large Language Model (LLM) feature expansion techniques. We utilized the National Alzheimer's Coordinating Center (NACC) dataset and developed a novel two-stage approach: (1) a baseline model using preprocessed NACC data, and (2) an LLM-enhanced model that expands a reduced 36-question questionnaire into 106 feature fields through medical knowledge augmentation. Three custom neural network architectures were trained and evaluated for diagnostic accuracy. The baseline model achieved 93% accuracy on the preprocessed NACC dataset. The LLM-enhanced approach demonstrated superior performance with 94% accuracy, effectively reducing the clinical burden while maintaining diagnostic precision. LLM-assisted feature expansion represents a promising approach for early neurological disorder detection, enabling comprehensive assessment through simplified questionnaires while achieving superior diagnostic accuracy compared to traditional methods.

Enhancing Early Alzheimer’s Detection in Aging Populations Using Novel LLM-assisted Feature Expansion

The global rise in the elderly population is putting an increasing strain on healthcare systems, particularly due to the increased chance of chronic illness, polypharmacy, and healthcare utilization. The traditional one-size-fits-all treatment approach often fails to address the unique and evolving needs of older adults, leading to adverse drug reactions, diminished quality of life, and inefficient use of medical resources. This study investigated the potential of personalized medicine — augmented by artificial intelligence (AI) — to improve elderly care through data-driven, individualized treatment plans. By integrating data from electronic health records, genomics, wearable devices, and lifestyle factors, we developed AI-enhanced models capable of tailoring treatments to each patient’s specific profile. These models allow for real-time treatment adjustments, predictive risk assessments, and optimized drug regimens. Our findings suggest that AI-driven personalized care can reduce adverse outcomes, enhance clinical effectiveness, and support proactive, scalable, and patient-centered healthcare for aging populations.

Improving Prediction of ADR in Elderly Patients Using Machine Learning Algorithms

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder that is a common type of dementia and is prevalent in aging populations. One hallmark of AD pathology is the hyperphosphorylation of tau, a microtubule-associated protein, which aggregates into neurofibrillary tangles that disrupt neuronal communication and trigger cell death. Microtubule Affinity Regulating Kinase 4 (MARK4), a serine/threonine kinase, initiates this pathological tau phosphorylation at Ser262, making it a promising therapeutic target. This study aimed to identify and optimize small-molecule inhibitors of MARK4 to mitigate tau aggregation in AD. Combining artificial intelligence (AI) tools and computational drug discovery techniques was employed, including structural modeling, molecular docking (via 1-Click Docking), and pharmacokinetic screening (using SwissADME and the BOILED-Egg model). The MARK4 catalytic domain (PDB ID: 5ES1) was used for docking simulations, centered at the ATP-binding site. Starting with known inhibitors, we designed and evaluated novel drug candidates (DCs) based on their docking scores, ligand efficiency, blood-brain barrier permeability, gastrointestinal absorption, CYP450 inhibition, and synthetic accessibility. Among the tested molecules, five DCs (5b, 6b, 7a, 8a, and 8c) demonstrated high binding affinity (≤ -8.7 kcal/mol), favorable pharmacokinetics, and CNS-targeting potential. These candidates surpassed baseline inhibitors and met key criteria for drug-likeness and synthetic feasibility. Our findings support the potential of AI to accelerate AD therapeutic development and propose five promising compounds for future in vitro and in vivo testing. This research highlights MARK4 inhibition as a viable strategy to halt tau-driven neurodegeneration in Alzheimer’s disease.

Inhibiting MARK4 to Prevent Tau Accumulation in Alzheimer's: A Docking Study 

National Inter-University Health Data and AI Inquiry Program Abstract: Strokes are a leading cause of mortality in Canada and long-term disability worldwide, occurring either through blockage of blood or rupture of blood vessels in the brain. The impact of strokes is expected to increase as global populations age. Although clinical risk factors are prioritized in stroke prediction, lifestyle behaviours and social determinants of health (SDOH) also play a crucial role. Prevention of stroke not only affects populations on a personal and societal level but may also reduce its impact on healthcare systems. This project aims to aid individuals in predicting stroke risk by utilizing machine learning and providing preventative measures as well as promoting healthier aging based on public Canadian health data. The model uses Random Forest Regression to replace empty values in the dataset with an approximation. It tests the accuracy by predicting existing crude rate data and comparing the approximations to the actual value. The model demonstrated accurate predictions for lower crude rate values, but struggled with extreme values due to limited data availability. Results showed statistically significant relationships between stroke and various modifiable factors (p < 0.05), including alcohol consumption, work stress, and physical activity. This project highlights the potential implementation of AI in primary stroke prevention using socioeconomic and behavioural data. In addition to clinical settings, this model can be applied on a larger scale in public health and insurance planning with further refinement and an expanded dataset.

Integrating Artificial Intelligence and Behavioural Risk Profiling for Early Stroke Prevention

Accurately detecting Alzheimer’s disease (AD) in its early stages is critical for intervention yet remains a persistent clinical challenge. Although volumetric MRI biomarkers like hippocampal atrophy are well-established, emerging research suggests that radiomic texture features may offer insights into neurodegeneration. This study investigates the relative predictive value of volumetric, demographic, and gray-level co-occurrence matrix (GLCM)-based texture features for early AD classification using the OASIS-2 dataset. T1-weighted MRI scans and clinical metadata from 120 participants (48 demented, 72 nondemented, excluding converted) were analyzed. Radiomic features, including GLCM texture metrics and first-order statistics, were extracted from skull-stripped sagittal slices. Volumetric estimates were derived from brain masks, and demographic variables included age, sex, education, and socioeconomic status (SES). Multiple Random Forest classifiers were trained using combinations of feature groups. Hyperparameter optimization was performed using grid search with 5-fold cross-validation, and class imbalances were addressed by the utilization of Synthetic Minority Oversampling Technique (SMOTE). The best performing model was the combination of volumetric and demographic features, achieving an accuracy of 79.2% and an AUC of 0.8430. GLCM texture features alone yielded lower predictive performance (accuracy: 57.5%, AUC: 0.5784) but provided marginal benefit when added to volume and demographic data (accuracy: 73.33%, AUC: 0.7935). Confusion matrix and Receiver Operating Characteristic (ROC) analyses confirmed the superiority of multimodal models. However, texture features may have been limited by slice-based extraction and modest sample size. These findings support the clinical utility of integrating structural MRI with demographic profiling for scalable, non-invasive AD screening.

Machine Learning Driven Discovery of MRI-based Alzheimer’s Biomarkers

DNA methylation is an epigenetic modification that involves the addition of methyl groups to DNA at specific sites, often affecting gene expression. These patterns can be influenced by genetic and environmental factors, including aging. Specific DNA methylation sites have been linked to various diseases, such as Alzheimer’s disease and cancer, making them promising epigenetic biomarkers for age and disease prediction. This study aims to develop machine learning models to predict chronological age and classify Alzheimer’s disease using DNA methylation profiles derived from blood samples. Two datasets were made from publicly available studies: the age prediction dataset included 9 studies with a total of 4,303 samples (median age 54 years) and the Alzheimer’s disease dataset which included 2 studies comprising of 823 samples, 247 of whom were from individuals diagnosed with Alzheimer’s. The final age prediction model achieved a mean absolute error (MAE) of 3.01 years and a median absolute error (MedAE) of 2.40 years. The Alzheimer’s classification model attained an accuracy of 79%, with a false positive rate of 13.9% and a false negative rate of 36.6%. Both models identified sites associated with relevant genes. The age model highlighted ELOVL2, FHL2, KLF14, and TRIM59, which are commonly associated with aging, while the Alzheimer’s model identified SCGN, FKBP5, GATA4, ORMDL3, IRF3, and SPIDR, all of which have been linked to Alzheimer’s disease in previous studies. These findings support the potential of DNA methylation as an epigenetic biomarker for both age and disease prediction.

Downloads

Next from Aging Powerfully: AI Blueprint for Chronic Disease Management and Social Determinants of Longevity

Deep Learning Insights into Social Determinants of Chronic Disease and Longevity

Similar lecture

GrayKD: Distilling Better Knowledge from Black-box LLM via Multi-rationale Injection

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES