From Normal Cognition to Dementia: Using Natural Language Processing to Identify Cognitive Stages from Clinical Notes

Samad Amini, You Cheng, Colin Magdamo, Ioannis Ch. Paschalidis, Sudeshna DasAlzheimer 2024 · DOI

TL;DR

This paper introduces a novel Natural Language Processing (NLP) approach to automatically identify cognitive stages (normal cognition, Mild Cognitive Impairment (MCI), and dementia) from electronic health record (EHR) clinical notes. By leveraging a transformer-based language model, they achieve high accuracy in classifying these stages, offering a scalable solution to improve dementia diagnosis and management.

What Problem Does This Paper Address?

Accurately and efficiently diagnosing cognitive impairments like dementia and MCI is crucial for timely intervention and patient care. However, traditional methods often rely on manual review of clinical notes, which is time-consuming, subjective, and difficult to scale across large patient populations. This paper addresses the urgent need for automated, reliable methods to detect these cognitive impairments by analyzing the rich, unstructured text data found in EHRs. The goal is to bridge the gap between the wealth of information in clinical notes and the need for systematic, data-driven cognitive staging.

Methods

The researchers employed a two-pronged NLP strategy using the Universal Sentence Encoder (USE), a transformer-based language model, to classify cognitive stages.

  1. Random Sampling: This method involved extracting sentences containing relevant keywords (related to dementia and daily living activities) and randomly batching them. Each batch was then processed by the USE to generate 512-dimensional embeddings.
  2. Encounter-based Sampling: This approach grouped sentences by their specific encounter types and keyword categories. This allowed for the creation of 24 distinct embeddings per patient, capturing a more contextualized representation of their cognitive status.

These methods were designed to classify patients into three categories: normal cognition, MCI, and dementia, using three binary classification tasks.

Datasets

The study utilized an EHR dataset from Massachusetts General Brigham, encompassing clinical notes from a 2-year period (2017-2018) across 12 types of patient encounters. The dataset included:

  • 531 subjects with normal cognition
  • 153 subjects with MCI
  • 229 subjects with dementia

This dataset appears to be a newly curated benchmark for this specific NLP task, focusing on the nuanced classification of cognitive stages within real-world clinical notes.

Key Insights and Ideas

  • Leveraging Unstructured Data: The paper demonstrates the significant potential of unstructured clinical notes, often overlooked or manually processed, as a rich source of information for cognitive staging.
  • Contextual Embeddings Matter: The comparison between random and encounter-based sampling highlights the importance of contextualizing information. Grouping sentences by encounter type and keyword categories (Encounter-based Sampling) generally yielded better results, suggesting that the context of a note is crucial for accurate interpretation.
  • Transformer Models for Clinical Text: The successful application of the Universal Sentence Encoder (USE) showcases the power of advanced transformer-based language models in understanding complex medical narratives and extracting clinically relevant features.
  • Scalability Potential: The automated nature of the NLP approach promises a highly scalable solution, which is critical for widespread adoption in healthcare systems.

Key Results

The NLP-based classification methods demonstrated strong performance across the three classification tasks:

  • Normal Cognition vs. Dementia:
    • Random Sampling: AUC 97.8%
    • Encounter-based Sampling: AUC 93.6%
  • Normal Cognition vs. MCI:
    • Random Sampling: AUC 81%
    • Encounter-based Sampling: AUC 74.6%
  • Normal Cognition vs. Cognitive Impairment (MCI + Dementia merged):
    • Random Sampling: AUC 92.4%
    • Encounter-based Sampling: AUC 85.4%

Notably, the proposed NLP approaches significantly outperformed a baseline model (based on the number of relevant sentences) by a margin of 6-13% in AUC across these tasks.

Limitations and Future Work

  • Limitations: The paper does not explicitly detail limitations, but common challenges in EHR analysis, such as data sparsity, variations in clinical documentation styles, and potential biases in the data, are implicit. The performance differences between the two sampling methods also suggest that the model's sensitivity to specific contextual cues might be a limitation.
  • Future Work: The authors suggest that their highly accurate, fully automated approach has significant scalability potential and could transform hospital practices in managing clinical notes. Future work could involve expanding the dataset, exploring other advanced NLP models, and integrating this system into clinical workflows for real-time decision support. Further investigation into the specific features that drive classification accuracy within different encounter types could also be beneficial.

Summary updated

Abstract

Our study harnessed NLP techniques for enhanced dementia staging accuracy using EHR clinical notes. We introduced a highly accurate, fully automated approach with scalability potential, promising to transform hospital practices in managing clinical notes.
BibTeX
@inproceedings{amini2024from,
  author = {Samad Amini and You Cheng and Colin Magdamo and Ioannis Ch. Paschalidis and Sudeshna Das},
  title = {{From Normal Cognition to Dementia: Using Natural Language Processing to Identify Cognitive Stages from Clinical Notes}},
  booktitle = {Alzheimer},
  year = {2024},
  doi = {10.1002/alz.089228},
}
Figure