TL;DR
This study compares two distinct approaches (hypothesis-driven and data-driven machine learning) to identify dementia risk factors in a large real-world cohort. Both methods converged on age, depression, and low education as major risk factors, while surprisingly finding a protective effect of higher Body Mass Index (BMI). This highlights the importance of considering multiple risk factors simultaneously for accurate dementia risk assessment.
What Problem Does This Paper Address?
Dementia, particularly Alzheimer's disease, poses a significant global health challenge with immense economic costs. While numerous risk factors have been identified, most research examines them in isolation. This is problematic because older adults often present with multimorbidity, meaning multiple co-existing conditions. Analyzing risk factors individually can fail to capture their complex interactions and cumulative effects, potentially leading to an incomplete understanding of dementia risk. This paper aims to bridge this gap by comparing two analytical paradigms to stratify dementia risk factors, considering their simultaneous impact in a real-world setting.
Methods
The study employed two primary methodologies:
- Hypothesis-Driven Approach: This involved traditional statistical methods, specifically survival analysis using Kaplan-Meier curves and Cox proportional hazards regression. Participants were stratified into quartiles based on their total count of risk factors, and hazard ratios (HRs) were calculated to quantify the impact of individual risk factors on dementia conversion.
- Data-Driven Approach: This utilized a machine learning (ML) framework, specifically the XGBoost algorithm, for prediction. A nested cross-validation (NCV) approach was implemented for robust model evaluation, along with hyperparameter optimization (HPO). To enhance interpretability and understand feature contributions, the SHAP (SHapley Additive exPlanations) framework was employed. This approach allowed for the simultaneous evaluation of multiple risk factors and their interactions.
A key aspect of the data-driven approach was the seed averaging technique, where the model was trained multiple times with different random seeds to ensure stable performance metrics and SHAP values, mitigating the impact of randomness.
Datasets
The study utilized data from the National Alzheimer's Coordinating Center (NACC) Uniform Data Set (UDS), a prospective cohort dataset collected from 2005 to 2023 across 46 Alzheimer's Disease Research Centers (ADRCs) in the United States. This dataset includes sociodemographic characteristics, medical history, behavioral symptoms, and cognitive and functional status of participants. The analysis included 9606 participants.
Key Insights and Ideas
- Convergent Validity: The most striking insight is the strong agreement between the hypothesis-driven and data-driven approaches. Despite using different analytical paradigms, both methods consistently identified the same core set of significant risk factors, lending strong credibility to their findings.
- Multimorbidity Matters: The study strongly advocates for evaluating risk factors simultaneously, moving beyond isolated analyses. This is crucial for understanding the complex interplay of conditions in older adults and for more accurate risk stratification.
- Unexpected Protective Factor: The finding that higher Body Mass Index (BMI) was protective against dementia conversion was unexpected and warrants further investigation. This challenges some prior assumptions and highlights the nuanced relationship between weight and cognitive health.
- Clinical Utility: The identified key risk factors (age, depression, low education, BMI) are readily assessable during a single clinical consultation, making the findings directly applicable to clinical practice for rapid risk assessment.
Key Results
- Significant Risk Factors: Both models consistently identified age, depression, and low education as major predictors of dementia conversion.
- Protective Factor: Higher BMI was found to be protective, associated with a 37% decrease in the hazard ratio of dementia conversion.
- Predictive Performance (Data-Driven): The ML model achieved a mean ROC AUC of 0.732 and a mean PR AUC of 0.208, indicating moderate discrimination performance. The model achieved an accuracy of 0.643, sensitivity of 0.702, and specificity of 0.637.
- Hypothesis-Driven Results: Older age significantly increased the hazard ratio of dementia diagnosis by 469%. Less educational attainment was associated with a 131% increase in HR, and depression with an 87% increase. Hypertension showed a smaller but significant increase (23%).
- Prevalence: Key risk factors identified included older age (71%), hypertension (46%), and hyperlipidemia (50%).
Limitations and Future Work
- External Validation: The study acknowledges the lack of external validation in independent cohorts and the primary reliance on a predominantly White population from specialized research centers. This limits the generalizability of the findings.
- Unavailable Risk Factors: Three modifiable risk factors from the Lancet Commission (air pollution, physical exercise, and social isolation) were not available in the NACC dataset.
- Binary Classification: The use of binary classifications for some risk factors may not fully capture the nuances of their impact.
- Future Directions: The authors suggest future studies should:
- Include external validation in diverse populations, including those from the Global South.
- Evaluate the cumulative impact of risk factors over time and their influence on cognitive outcomes.
- Incorporate more granular variables (e.g., blood pressure, plasma glucose levels).
- Investigate the complex interplay and potential causal pathways between depression and dementia.
- Further explore the protective role of BMI and its potential confounding factors.
Summary updated