Introduction
Alzheimer’s Disease (AD) remains a significant global health challenge with no known cure. Consequently, identifying and managing modifiable lifestyle factors—such as physical activity and diet—has become a primary strategy for prevention and delay of onset. While Electronic Health Records (EHRs) contain vast amounts of patient data, lifestyle information is frequently trapped within unstructured clinical narratives. Manually extracting this data is prohibitively labor-intensive. This study explores the use of weak supervision to train deep learning models, specifically BERT-based architectures, to automate the extraction of these critical lifestyle factors from clinical notes.
Methodology
The researchers developed a pipeline designed to overcome the "annotation bottleneck" associated with supervised learning in the medical domain. The approach utilizes weak supervision to generate training data without extensive manual labeling.
Study Design and Pipeline
The process begins with raw clinical notes from a large-scale repository. A rule-based NLP system (NLP-PIER) is used to filter and assign "weak labels" to sentences based on Unified Medical Language System (UMLS) Concept Unique Identifiers (CUIs). These weakly labeled sentences then serve as the training set for fine-tuning various BERT models.
Data and Models
The study utilized over 180 million clinical notes from the University of Minnesota’s Clinical Data Repository (CDR). Two primary case studies were conducted:
- Physical Activity: A dataset of 23,559 sentences (96.7% active, 3.3% inactive).
- Excessive Diet: A dataset of 886 sentences focusing on high fat, calorie, and salt intake.
Six BERT variants were evaluated, ranging from general-domain BERT Base to domain-specific models like PubMedBERT, BioBERT, and UMLS BERT. Performance was validated against a Gold Standard Corpus (GSC) consisting of 50 manually annotated sentences per category.
Results
The application of weak supervision proved highly effective. Domain-specific models consistently outperformed the general BERT Base model, demonstrating the importance of pre-training on biomedical and clinical corpora.
Performance in Physical Activity Classification
For the physical activity category, the PubMedBERT(Abs) model achieved the highest performance, reaching a Precision, Recall, and F1-score of 0.96.
Performance in Excessive Diet Classification
In the diet category, BioBERT achieved a perfect F1-score of 1.00. While impressive, the researchers noted that the smaller size of the diet dataset (886 sentences) might contribute to this result and warrants further investigation with larger samples.
Discussion and Conclusion
This research demonstrates that "noisy" labels generated by rule-based systems are sufficient to train high-performance transformer models for clinical NLP tasks. This methodology allows for the scaling of training sets to tens of thousands of examples with minimal human intervention.
However, challenges remain. Models occasionally struggled with complex linguistic contexts, such as distinguishing between a patient's current activity level and their future goals (e.g., "wants to increase exercise"). Future work will focus on improving the handling of negation and nuance, as well as expanding the pipeline to include other AD-related lifestyle factors like sleep quality and substance use. By automating the extraction of these factors, researchers can better leverage EHR data to understand and combat Alzheimer's Disease.
Summary updated