Towards Early Dementia Detection: Fusing Linguistic and Non-Linguistic Clinical Data

Joseph Bullard, Cecilia Ovesdotter Alm, Xumin Liu, Qi Yu, Rubén A. ProañoNAACL 2016 · DOI

Abstract

Early detection of Alzheimer’s disease and related dementias is critical for effective clinical intervention and patient planning. While structured data like biomarkers and cognitive scores are standard, unstructured clinical notes contain rich, often untapped diagnostic signals. This study investigates the fusion of linguistic features derived from doctor's notes with structured clinical data using Natural Language Processing (NLP) techniques. Our results demonstrate that fusing these data sources significantly improves classification accuracy, particularly in identifying early-stage cognitive impairment, where linguistic data serves as a vital bridge when specialized cognitive tests are unavailable.

Introduction

The "holy grail" of geriatric medicine is the early detection of dementia. Current diagnostic pipelines often rely on expensive or invasive structured data, such as MRI scans and cerebrospinal fluid biomarkers. However, clinical text notes are a ubiquitous byproduct of patient care, capturing nuanced clinician observations that structured codes often omit. This research explores whether NLP can transform these routine notes into a predictive engine for early dementia detection.

Methods

The study utilized data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI-2), involving 679 subjects categorized into four stages: Normal (NL), Early Mild Cognitive Impairment (EMCI), Late Mild Cognitive Impairment (LMCI), and Alzheimer’s Disease (AD).

Researchers compared three linguistic representations:

  1. Bag-of-Words (BOW): Simple word frequency counts.
  2. TF-IDF: Weighted counts highlighting unique terms.
  3. Latent Dirichlet Allocation (LDA): Topic modeling to group related words into interpretable themes.

Two fusion strategies were tested to combine text with 22 structured clinical fields (biophysical markers and cognitive scores):

  • Vector Concatenation ($\cup$): Joining feature sets into a single vector.
  • Posterior Probability Composition ($\oplus$): A Bayesian approach combining confidence scores from separate models.

Key Insights: The Power of Clinical Text

A major finding is that text notes are potent stand-alone predictors. Models using only text often rivaled those using structured biophysical data. This suggests that "soft" clinical observations are dense with diagnostic signals.

Furthermore, LDA topics provided human-readable clinical context. For instance, high-ranked topics included terms directly reflecting the clinician's evaluative process, such as "impression," "MMSE," and "amnestic."

Table 2a: Top 10 words for high-ranked LDA topics (K=60), illustrating the clinical relevance of discovered themes.

Results

The models were evaluated on a 4-class "Standard" problem and a 2-class "Early Risk" problem (NL vs. EMCI).

Standard Task (4-class)

Fusing text with structured data (including cognitive scores) reached an accuracy of ~65% in leave-one-out cross-validation (LOOCV), a significant jump over the 27.5% majority-class baseline.

Early Risk Task (NL vs. EMCI)

The linguistic data was most impactful in the "Early Risk" scenario. Fusing TF-IDF features with structured data yielded a peak accuracy of 83.1%. Notably, adding text to structured data without cognitive scores boosted accuracy from 67.6% to 76.8%, proving that linguistic data captures unique information missing from biophysical markers.

Performance results for the Early Risk (NL vs. EMCI) task using Leave-one-out Cross-validation (LOOCV).

Conclusion

Fusing unstructured clinical notes with structured data significantly enhances the detection of early cognitive decline. While LDA topics offer interpretability, simple vector concatenation often performs as well as more complex Bayesian fusion. Future work will explore supervised LDA and lexical similarity to further refine these predictive models for real-world clinical application.

Summary updated

BibTeX
@inproceedings{bullard2016towards,
  author = {Joseph Bullard and Cecilia Ovesdotter Alm and Xumin Liu and Qi Yu and Rubén A. Proaño},
  title = {{Towards Early Dementia Detection: Fusing Linguistic and Non-Linguistic Clinical Data}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2016},
  doi = {10.18653/v1/w16-0302},
}
Figure