TL;DR
Researchers at Harvard and Massachusetts General Hospital have applied Clinical BERT and pre-trained medical embeddings to Electronic Health Records (EHRs) to identify distinct subtypes of Alzheimer’s Disease and Related Dementias (ADRD). By analyzing both structured diagnostic codes and unstructured clinical notes, they identified three consistent patient clusters defined by comorbidities like chronic pain, diabetes, and cardiovascular disease, suggesting that dementia is far more heterogeneous than a single diagnosis suggests.
What Problem Does This Paper Address?
Alzheimer’s Disease (AD) affects over 50 million people globally, yet our "one-size-fits-all" approach to treatment has largely failed. For decades, the "amyloid hypothesis"—the idea that protein plaques are the primary cause—has dominated research, but drugs targeting these plaques often show limited efficacy.
The authors argue that the problem lies in heterogeneity. "Alzheimer’s" may actually be an umbrella term for several distinct biological pathways (etiologies). While previous studies have tried to find these subtypes using expensive RNA sequencing or brain imaging, this paper tackles the challenge using Electronic Health Records (EHRs). EHRs are a goldmine of data but are notoriously "noisy" and difficult to analyze because they include both rigid billing codes (ICD-9) and messy, unstructured doctor’s notes.
Methods
The study employs unsupervised learning to find patterns in patient data without pre-defined labels. The researchers took a two-pronged approach:
- Structured Data (ICD-9 Codes): Instead of simple "one-hot encoding" (which treats every medical code as an independent category), they used pre-trained 300-dimensional embeddings. These embeddings capture the semantic relationships between diseases (e.g., "hypertension" is mathematically closer to "heart failure" than to "broken toe").
- Unstructured Data (Clinical Notes): They utilized Clinical BERT, a transformer model fine-tuned on medical text. To turn long neurologist notes into a single mathematical vector, they used a novel attention-weighted averaging technique. By calculating the entropy of the model's attention heads, they could automatically give more weight to "informative" words while ignoring filler text.
- Clustering: Both data types were clustered using Hierarchical Agglomerative Clustering with Ward’s linkage. The researchers then performed enrichment analysis (using Chi-squared tests and TF-IDF) to see which medical conditions and words defined each cluster.
Datasets
The study used data from the Mass General Brigham (MGB) Research Patient Data Registry.
- Initial Cohort: ~20,000 candidate patients.
- Final Cohort: 6,470 patients after filtering for those over age 50 with sufficient follow-up and substantial clinical notes (at least 512 characters).
- The dataset included structured ICD-9 diagnostic codes and unstructured progress notes written by neurologists.
Key Insights and Ideas
- The "Orthogonal" Discovery: When visualizing the clusters (using UMAP), the researchers found that the main differences between patient groups were not driven by the presence of an Alzheimer’s code itself. Instead, the clusters were defined by comorbidities. This suggests that the "background" health of a patient significantly shapes their dementia phenotype.
- The Pain-Diabetes Axis: One of the most distinct clusters was heavily enriched for Type 2 Diabetes (T2D) and chronic pain (fibromyalgia, back pain). While the link between T2D and AD is known, the strong association with chronic pain is a more novel finding that warrants further investigation as a specific disease trajectory.
- Text Validates Codes: The clusters found in the clinical notes (unstructured) mirrored those found in the ICD codes (structured). For example, the "Cardio/Aging" cluster appeared in both modalities, proving that transformer models can extract high-quality clinical signals directly from raw text.
- Movement and Gait: The notes-based clustering was particularly good at identifying a "Hydrocephalus/LBD" cluster, characterized by words like "tremor" and "gait," which are often clinical precursors or complications of specific dementia types like Lewy Body Disease.
Key Results
The researchers identified three primary clusters in both modalities:
- ICD-9 Clusters:
- Chronic Pain/T2D (N=508): High prevalence of diabetes and pain-related codes.
- Other Neuro (N=4,350): The largest group, primarily focused on neurodegenerative conditions with fewer general medical interactions.
- Cardio/Aging (N=1,612): Older patients with high rates of hypertension and cardiovascular disease.
- Notes Clusters: Identified similar "Cardio/Aging" and "Other Neuro" groups, but also a specific Hydrocephalus/LBD group (N=1,469) enriched for movement disorders.
- Statistical Significance: All clusters showed highly significant differences (p < 0.001) in age of onset, sex distribution, and prevalence of specific dementia types (AD vs. FTD vs. LBD).
Limitations and Future Work
- Single-Center Study: The findings rely on data from one hospital system (MGB). The authors acknowledge the need for an independent validation cohort to ensure these clusters exist in other populations.
- Billing Bias: ICD codes are generated for billing, not research, which can introduce noise.
- Modality Fusion: Currently, the codes and notes are clustered independently. The authors suggest that future work should use multi-modal deep learning (like joint autoencoders) to fuse these data sources into a single, more powerful representation of the patient.
- Temporal Dynamics: The current model is a "snapshot." It doesn't explicitly account for the order in which symptoms appear, which is crucial for staging disease progression.
Summary updated