Personalized Early Stage Alzheimer's Disease Detection: A Case Study of President Reagan's Speeches

Ning Wang, Fan Luo, Vishal Peddagangireddy, K. P. Subbalakshmi, Rajarathnam Chandramouli

Abstract

Early detection of Alzheimer’s Disease (AD) is vital for improving patient outcomes through timely intervention. This paper presents an unsupervised machine learning approach to identify early AD symptoms by monitoring longitudinal changes in an individual's speech. Using 98 public speeches by President Ronald Reagan from 1980 to 1989, the study employs linguistic feature extraction, t-SNE visualization, and anomaly detection algorithms. The findings suggest that Reagan exhibited detectable linguistic degradation between 1983 and 1987, years before his 1994 diagnosis. This methodology highlights the potential for personalized, unobtrusive cognitive health monitoring.

Introduction

Alzheimer’s Disease remains a significant global health challenge with no known cure, making early diagnosis the primary lever for enhancing quality of life. Conventional machine learning models for AD detection typically utilize supervised learning on clinical datasets, comparing patients against healthy controls in standardized tasks. However, these methods often require clinical visits and may not account for individual baseline variations.

This research addresses the need for personalized mental health monitoring. By treating an individual's own historical speech as a baseline, the authors utilize unsupervised learning to detect linguistic degradation as statistical anomalies. This approach shifts the focus from population-wide comparisons to intra-individual longitudinal tracking.

Methodology

The study frames early AD detection as an unsupervised clustering and anomaly detection problem based on longitudinal linguistic data.

Feature Engineering

The researchers extracted nine key linguistic biomarkers from speech transcripts, including:

  • Part-of-Speech (POS) Ratios: Specifically the pronoun-to-noun ratio, as word-finding difficulties often lead to increased pronoun usage.
  • Vocabulary Richness: Metrics such as Honore’s and Brunet’s measures.
  • Readability Scores: Including the Automated Readability Index.

Dimensionality Reduction and Visualization

To manage high-dimensional linguistic data, t-distributed stochastic neighbor embedding (t-SNE) was applied. This technique maps complex features into a 2D space, facilitating the visual identification of clusters that share AD-like characteristics.

t-SNE visualization of linguistic features across Reagan's speeches, colored by year. A distinct cluster (circled) emerges representing the mid-1980s.

Anomaly Detection

Two unsupervised algorithms were implemented to automate the detection of cognitive decline:

  1. One-Class SVM: Trained on early "normal" speech to flag subsequent deviations.
  2. Isolation Forest: Used to isolate anomalous data points across the entire timeline without a pre-defined training period.

Results

The analysis of President Reagan’s speeches yielded several key insights:

  • Linguistic Shifts: A notable shift in POS usage was observed. As cognitive function declines, patients often compensate for noun-finding difficulties by increasing pronoun frequency and using non-specific filler words.
  • Visual Clustering: The t-SNE analysis successfully grouped speeches from 1983 to 1987. These speeches were characterized by lower readability scores, higher pronoun-to-noun ratios, and increased repetition of high-frequency words.
  • Automated Detection: The Isolation Forest algorithm flagged 10 specific speeches between 1983 and 1988 as linguistic anomalies. Similarly, the One-Class SVM identified a cluster of anomalous speeches from 1984 to 1986.
Isolation Forest results identifying anomalous linguistic patterns (blue) relative to the baseline observations (white).

These results align with historical retrospective studies of Reagan's conversational patterns, suggesting that signs of AD were present during his presidency, well before his official announcement in 1994.

Discussion and Future Work

While promising, this case study has limitations. The pipeline was tuned specifically for Reagan's dataset, and its generalizability to other individuals requires further validation. Additionally, the use of prepared presidential speeches may mask spontaneous linguistic markers, as these texts are often influenced by speechwriters.

Future research aims to integrate this methodology into Internet-of-Things (IoT) devices, such as smart speakers. Such systems could provide low-cost, continuous, and unobtrusive early-warning signals for cognitive decline by automatically analyzing daily speech patterns against a personalized historical baseline.

Summary updated

BibTeX
@inproceedings{wang2020personalized,
  author = {Ning Wang and Fan Luo and Vishal Peddagangireddy and K. P. Subbalakshmi and Rajarathnam Chandramouli},
  title = {{Personalized Early Stage Alzheimer's Disease Detection: A Case Study of President Reagan's Speeches}},
  booktitle = {BioNLP},
  year = {2020},
  doi = {10.18653/v1/2020.bionlp-1.14},
}
Figure