A breakthrough AI model analyzes diverse data to reveal faster Alzheimer’s progression in Black females, offering a powerful tool for early detection and personalized care.
Study: Prediction and clustering of Alzheimer’s disease by race and sex: a multi-head deep-learning approach to analyze irregular and heterogeneous data. Image Credit: STEKLO / Shutterstock
In a recent study published in the journal Scientific Reports, researchers in the United States developed a deep learning multi-head model designed to analyze the progression of Alzheimer’s disease (AD) in cognitively normal individuals. This model uses advanced neural network architectures, including convolutional neural networks (CNN) and long-short-term memory (LSTM), to process multimodal data and predict AD progression. They found four primary progression clusters—slow, moderate, rapid converters, and non-converters—with patterns varying by race and sex. The model significantly outperformed single-modal models and identified critical, diverse predictors of disease progression.
Background
AD affects cognition, memory, and behavior, with early detection critical to optimizing treatment outcomes. Current studies largely target individuals with mild cognitive impairment (MCI) or advanced symptoms, while little research explores asymptomatic individuals with normal cognition who are at risk for AD progression. This underrepresentation is particularly evident for Hispanic/Latino and Black/African American people, who face a higher AD risk but are less often included in clinical trials. Factors contributing to this gap include mistrust, socioeconomic barriers, and logistical challenges, all of which complicate data collection across diverse populations. Machine learning is increasingly used in neurology for disease progression prediction, which can enable earlier detection and intervention. However, conventional models struggle to manage the complex and high-dimensional nature of biomedical data. Scalable deep learning models are better suited to identify patterns across diverse risk factors, potentially enhancing AD risk prediction and supporting equitable, inclusive AD research. With these challenges in mind, researchers in the present study developed a multi-head deep learning model to predict AD progression, identifying key predictors and clustering participants to capture population-level heterogeneity.
About the Study
The researchers utilized comprehensive, high-dimensional data from the National Alzheimer’s Coordinating Center (NACC) to build a multi-head predictive deep learning model. This model was uniquely structured to handle large datasets, combining biomedical and imaging data to enhance AD progression prediction, clustering, and feature extraction. Extensive preprocessing was applied to refine the data and mitigate potential biases, including feature selection, balancing methods, and handling missing values, resulting in the inclusion of 6,110 participants and 447 features across various domains. About 63% of participants were female, 87% were White (including 5% Hispanic/Latino), and 13% were Black or African American. Most participants were over 61 years of age at their initial visit, providing a substantial dataset for this age group.
The multi-head neural network model used multimodal data and fusion strategies to learn and differentiate complex interactions. This approach uniquely combines CNN, LSTM, and XGBoost models to capture longitudinal data and predict disease progression with high accuracy. Further analyses included SHAP-based feature extraction and clustering, identifying top predictors of disease progression such as Clinical Dementia Rating, stroke, depression, and diabetes, and validating population heterogeneity.
Results and Discussion
Approximately 61% of the participants were missing the risk gene apolipoprotein E4 (APOE4) allele, while 27% had 1 copy and 3% had 2 copies. The multi-modality, multi-head, early-fusion, two-layer CNN model outperformed 14 other single-modal models in key metrics, including accuracy, precision, recall, and F1 score, demonstrating strong pattern recognition and prediction of imbalanced classes.
Participants were classified into four clusters based on progression rate: non-converters (80%), slow converters (14%), moderate converters (4%), and rapid converters (2%). Black/African American participants, particularly females, showed faster disease progression and more variability than White participants, with non-converters remaining the largest group across sex and race. Notably, Black/African American females in the rapid converter group bypassed the MCI stage, progressing directly to dementia, whereas White females tended to experience an intermediate MCI phase. This trend highlights how race and sex can influence AD progression, with Black/African American females showing both earlier disease onset and greater variability across all clusters. MRI-based predictors, while influential, ranked lower in predicting specific transitions to MCI/dementia.
Conclusion
In conclusion, this study successfully demonstrated that a multi-head deep learning model can effectively address AD’s complexity, uncovering significant heterogeneity across progression clusters and offering new insights into how demographics influence disease progression. Black/African American participants, particularly females, showed earlier and more variable disease progression compared to White participants. In the future, this approach may support the development of targeted screening methods for those at highest risk.
Journal reference:
- Chang, C. Y., Slowiejko, D., & Win, N. (2024). Prediction and clustering of Alzheimer’s disease by race and sex: A multi-head deep-learning approach to analyze irregular and heterogeneous data. Scientific Reports, 14(1), 1-12. DOI: 10.1038/s41598-024-77829-1, https://www.nature.com/articles/s41598-024-77829-1