Dementia Prediction via Hierarchical Attention in Notes
Original Title: Clinical Manifestations
Journal: Alzheimer's & dementia : the journal of the Alzheimer's Association
Overview
The clinical interview is the primary diagnostic gateway for identifying dementia, serving as a screening phase to determine if a patient requires intensive neurological evaluation. While large language models excel in general text processing, their utility in analyzing unstructured medical records for cognitive assessment remains under-explored. This research evaluates a deep learning framework designed to predict Alzheimer’s disease solely from clinical notes. The study used a dataset of 1,387 clinical notes collected from medical centers in South Korea, including 542 Alzheimer’s cases and 845 normal controls. Notes were structured into ten categories with up to 22 question-and-answer pairs each. The researchers implemented a hierarchical attention mechanism operating across sentence-level and category-level layers. The system was trained on 80% of the data and tested on 20%. The proposed model achieved an accuracy of 0.74 and a macro F1-score of 0.72. These results outperform standard models like ChatGPT-4o and Claude-3.5-sonnet, which recorded accuracies of 0.61 and 0.56, respectively.
Novelty
The technical contribution lies in a hierarchical attention architecture designed for the multi-layered structure of clinical interviews. Unlike linear models, this approach uses a Sentence-Level Attention layer to contextualize question-and-answer pairs via DistilBERT. A Category-Level Attention layer then aggregates these responses into a unified representation. A distinct feature is the integration of a context quality metric. This metric evaluates sentence specificity and topic relevance, allowing the model to prioritize informative responses while de-emphasizing vague dialogue. This weighting ensures the final context vector reflects diagnostically significant interview portions. The architecture remains compatible with various language models, providing a flexible framework for future integration with advanced encoders.
Potential Clinical / Research Applications
This methodology offers potential for enhancing dementia screening in primary care and specialized clinics. By automating note analysis, the system helps identify patients for prioritized diagnostic procedures, such as amyloid PET imaging or specialized cognitive testing. In research, this framework can analyze longitudinal records to identify linguistic shifts preceding clinical diagnosis. Furthermore, quantifying textual quality provides a metric for evaluating documentation thoroughness. This could lead to digital tools providing real-time feedback to medical staff, ensuring collected data is sufficiently detailed for accurate assessment.