Abstract
Early detection of Alzheimer’s Disease (AD) is paramount for improving patient quality of life and managing disease progression. This study presents a personalized, unsupervised machine learning framework designed to detect early-stage AD through longitudinal linguistic analysis. Using President Ronald Reagan’s speeches from 1980 to 1989 as a case study, the research employs feature engineering, t-SNE visualization, and anomaly detection algorithms. The results identify a significant cognitive shift between 1983 and 1987, demonstrating that linguistic biomarkers can signal neurodegeneration years before clinical diagnosis.
Introduction
Alzheimer’s Disease represents a global health crisis with no current cure. The "holy grail" of AD research is early detection, yet traditional clinical assessments are often infrequent and stressful for patients. While most machine learning research focuses on distinguishing AD patients from healthy controls in general populations, this paper addresses the challenge of personalized detection. Because language is highly individual, a person's own historical data serves as the most accurate baseline for identifying cognitive decline. This study explores whether unsupervised models can detect these subtle shifts in a high-profile longitudinal dataset: the public speeches of Ronald Reagan.
Methodology
The study utilized 98 speech transcripts from the Reagan Library, covering his two terms in office. The methodology followed a three-pronged unsupervised learning approach.
Feature Engineering and Selection
Researchers initially extracted 11 linguistic features categorized into Part-of-Speech (POS) ratios, vocabulary richness, and readability. POS features included pronoun-to-noun ratios and the frequency of verbs and adverbs. Vocabulary richness was measured using metrics like Honore’s Statistic and the Sichel Measure, while readability was assessed via the Automated Readability Index (ARI). To ensure model efficiency and reduce redundancy, a correlation analysis was performed to prune the feature set.
Visualization and Anomaly Detection
To interpret the high-dimensional linguistic data, the researchers used t-Distributed Stochastic Neighbor Embedding (t-SNE) to map the speeches into a 2D space. This allowed for the visual identification of clusters with similar linguistic patterns. For automated detection, the study employed:
- One-Class SVM: Trained to define the "normal" linguistic profile of the subject and flag deviations.
- Isolation Forest: An algorithm that isolates anomalies by randomly partitioning data, particularly effective when "normal" historical data is limited.
Key Results and Insights
The analysis successfully identified a clear timeline of cognitive decline that aligns with historical suspicions regarding Reagan's health during his second term.
The Timeline of Decline
Both the t-SNE visualization and the anomaly detection algorithms (SVM and Isolation Forest) identified a distinct cluster of "abnormal" speeches between 1983 and 1987. This finding is significant as it predates Reagan’s official AD diagnosis in 1994 by nearly a decade.
Linguistic "Leakage"
A primary indicator of decline was the "pronoun-to-noun" shift. Patients in the early stages of AD often experience the "tip-of-the-tongue" phenomenon, struggling to find specific nouns and substituting them with generic pronouns. The model captured this shift, alongside a reduction in unique vocabulary and lower readability scores. Interestingly, these findings in scripted speeches mirror results from previous studies of Reagan’s unscripted interviews, suggesting that cognitive decline "leaks" into even highly prepared oratory.
Discussion and Limitations
This research highlights the potential for non-invasive, continuous monitoring of cognitive health. By integrating such algorithms into smart home technology, clinicians could receive early warning alerts based on natural conversation.
However, the study notes several limitations. Presidential speeches are often collaborative efforts involving speechwriters, which may dilute the purely individual linguistic signal. Furthermore, as a single-subject case study, these results require validation across broader demographics. Future work will focus on applying these models to spontaneous speech and exploring deep learning architectures to detect subtler semantic changes.
Summary updated