📄 On Automatic Detection of Cognitive Decline

Summary

This position paper by Fabio Tamburini asks how language technology can catch cognitive decline, from Mild Cognitive Impairment (MCI) to early Alzheimer's disease, soon enough for intervention and population-wide screening. It reviews work on Digital Linguistic Biomarkers (DLBs), quantifiable properties of speech and writing that range from how varied a person's vocabulary is and how complex their sentence structure is to fluency, semantic coherence and acoustic characteristics. The author argues that the dominant setup is poorly suited to the problem. In that setup, a classifier labels a person from language samples collected on one occasion. Decline unfolds continuously over many years, and individual factors like cognitive reserve shift when symptoms surface. A one-off snapshot therefore struggles to separate pathology from a person's own baseline. The paper instead proposes longitudinal analysis that follows each individual's linguistic behaviour across decades. To support this, the author compiles and examines a long-term corpus of spontaneous speech from well-known public figures. The detailed design and findings of that experiment fall in a portion of the text not provided here. The work matters because it shifts DLB research from cross-sectional accuracy toward tracking change within each person.

Key points

• Context figures cited from the Alzheimer's Association (2025):
  • About 7.2 million Americans aged 65+ have Alzheimer's dementia, a number projected to approach 13.8 million by 2060.
  • MCI affects roughly 12–18% of people over 60, and 10–15% of them progress to dementia each year.
• In shared tasks (ADReSS/ADReSSo and the PROCESS Grand Challenge at ICASSP-2025), the best systems reached F1 scores of roughly 70–80% in 2- or 3-class settings based on clinical judgment.
• Across the literature, separating Alzheimer's patients from healthy controls often exceeds 90% accuracy. Distinguishing MCI from controls, the clinically more valuable case, falls to around 75–80%.
• Cognitive reserve, inferred through proxies such as education, occupation and questionnaires like CRIq, can postpone diagnosis while pathology continues. This weakens single-time-point classification.
• The proposed alternative is a longitudinal corpus of spontaneous speech from famous individuals spanning decades, used to study linguistic change within each person over time.