Artificial intelligence for dementia—Applied models and digital health

Donald M. Lyall, Andrey Kormilitzin, Claire Lancaster, José Sousa, Fanny Petermann‐Rocha, Christopher Buckley, Eric L. Harshfield, Matthew H. Iveson, Christopher R. Madan, Ríona Mc Ardle, Danielle Newby, Vasiliki Orgeta, Eugene Tang, Stefano Tamburin, Lokendra Thakur, Ilianna Lourida, David J. Llewellyn, Janice M. RansonAlzheimer 2023 · DOI

TL;DR

This review explores how artificial intelligence (AI) and digital health technologies are revolutionizing dementia research, from early risk prediction and diagnosis to tracking disease progression. By integrating diverse data sources like electronic health records and wearable sensors, AI offers the potential for more accurate, personalized, and accessible dementia care.

What Problem Does This Paper Address?

Dementia, a complex and multifaceted neurodegenerative disease, poses a significant global health challenge. Accurately predicting an individual's risk, diagnosing the condition early, and monitoring its progression are crucial for effective management and treatment. However, traditional methods often rely on subjective assessments, are time-consuming, and can be inaccessible. This paper addresses the need for more robust, objective, and scalable approaches by examining the application of AI and digital health technologies in dementia research. The authors highlight the challenge of integrating diverse and complex data sources to build comprehensive models that can overcome the limitations of existing methods.

Methods

The paper reviews the application of machine learning (ML) models and digital health technologies across three key areas: dementia risk prediction, diagnosis, and prognosis.

  • Applied Models: The review focuses on ML approaches, which use algorithms to learn patterns from data and make predictions. This includes techniques like logistic regression, decision trees, Bayesian networks, neural networks, support vector machines (SVM), and gradient boosting machines (GBM). The paper emphasizes the integration of "multi-omics" data (genetics, biomarkers, etc.) and electronic health records (EHRs) to improve model accuracy.
  • Digital Health Technologies: The authors discuss the use of various digital tools for data collection, including:
  • Wearable sensors: Accelerometers and other devices to track physical activity, gait, and sleep patterns.
  • Digital assessments: Online or app-based cognitive tests that can be administered remotely.
  • Natural language processing (NLP): To extract valuable information from unstructured text in EHRs.
  • Virtual reality (VR): For ecologically valid cognitive assessments.

The paper highlights the trend towards "deep phenotyping," which involves collecting rich, multi-modal data to create more detailed individual profiles.

Datasets

The paper discusses the use of various datasets in the context of ML for dementia research:

  • Longitudinal Cohort Studies: Examples include the Health and Retirement Study, Framingham Heart Study, and the Alzheimer's Disease Neuroimaging Initiative (ADNI). These studies provide long-term data on individuals.
  • Large-Scale Datasets: UK Biobank is frequently mentioned as a valuable resource for its extensive data on a large population.
  • Electronic Health Records (EHRs): Datasets from systems like the Secure Anonymised Information Linkage (SAIL) databank in Wales and platforms like South London and Maudsley NHS Foundation Trust Clinical Record Interactive Search (SLAM CRIS) are discussed for their potential in large-scale analysis.
  • "E-cohorts" and Registries: Initiatives like the Brain Health Registry and Dementias Platform UK (DPUK) are creating large, internet-based registries for participant recruitment and data collection.

The paper emphasizes the need for transparent, accessible, and harmonized data sharing across these diverse sources.

Key Insights and Ideas

  • The Power of Multi-Modal Data: Integrating diverse data sources (genomics, biomarkers, EHRs, sensor data) significantly enhances the predictive power of AI models compared to single-source approaches.
  • Digital Health as a Game-Changer: Wearable sensors and digital assessments offer objective, continuous, and scalable ways to collect data that was previously subjective or difficult to measure, providing richer insights into cognitive function and disease progression.
  • Moving Beyond Primary Prediction: AI is enabling a shift from simply predicting dementia risk to understanding secondary outcomes like treatment response and physical health in individuals with dementia.
  • The Promise of NLP for EHRs: Natural language processing can unlock the wealth of information hidden within unstructured clinical notes, providing a more comprehensive view of patient history and care.
  • "Bench to Bedside" Integration is Key: While AI models show great promise in research settings, translating these findings into practical clinical applications requires a concerted effort to ensure trust, accessibility, and regulatory approval.

Key Results

  • Improved Predictive Power: Machine learning approaches have demonstrated improved predictive power compared to standard clinical risk scores in predicting dementia. For instance, one study showed ML models achieving 92% accuracy in predicting incident dementia over 2 years, outperforming existing models like CAIDE and BDSI (around 80%).
  • High Accuracy in Specific Tasks: Ensemble ML models using speech features and memory tests achieved 97.2% accuracy in distinguishing participants at high and low risk of dementia. Another model using RF and GBM algorithms achieved 87% accuracy, 99% specificity, and 76% sensitivity in predicting dementia risk.
  • Diagnostic Subtyping Potential: While classifying dementia subtypes remains challenging (40-50% accuracy across multiple ML methods), there's significant room for improvement. Deep feed-forward neural networks have shown promise in subtyping rare dementia variants like PPA and FTD with 80% accuracy.
  • Early Detection of Risk Factors: Digital mobility markers (gait, physical activity) are showing associations with cognitive decline and dementia, with potential for early detection.

Limitations and Future Work

  • Data Accessibility and Harmonization: A significant challenge is the need for transparent, accessible, and harmonized data sharing across diverse datasets and cohorts.
  • Interpretability and Trust: The "black-box" nature of some AI models can hinder clinical trust and decision-making. Future work should focus on developing more interpretable models.
  • Population Generalizability and Equity: Algorithms trained on specific datasets may not generalize well to diverse populations, raising concerns about data equity.
  • Causality: Identifying causal risk factors remains challenging, as many identified factors are part of a broader "protective lifestyle" or reflect disease processes.
  • Sample Size and Covariance: Small sample sizes in some studies and the covariance among risk factors can complicate ML modeling.
  • Future Directions: The authors recommend:
  • Emphasis on transparent and reproducible analyses.
  • Integration of multi-omics data and EHRs.
  • Movement towards secondary outcomes and clinical cohorts.
  • Development of bespoke tools for naturalistic "real-world" data.
  • Focus on developing scalable, accessible, and ethically sound digital health solutions.
  • Collaboration between researchers, clinicians, patients, and regulators to ensure clinical integration.

Summary updated

Abstract

Such approaches will benefit also from improvements in remote data measurement, whether cognitive (e.g., online), or naturalistic (e.g., watch-based accelerometry).
BibTeX
@inproceedings{lyall2023artificial,
  author = {Donald M. Lyall and Andrey Kormilitzin and Claire Lancaster and José Sousa and Fanny Petermann‐Rocha and Christopher Buckley and Eric L. Harshfield and Matthew H. Iveson and Christopher R. Madan and Ríona Mc Ardle and Danielle Newby and Vasiliki Orgeta and Eugene Tang and Stefano Tamburin and Lokendra Thakur and Ilianna Lourida and David J. Llewellyn and Janice M. Ranson and Janice M. Ranson},
  title = {{Artificial intelligence for dementia—Applied models and digital health}},
  booktitle = {Alzheimer},
  year = {2023},
  doi = {10.1002/alz.13391},
}
Figure