Language Models Detect Alzheimer's Through Meaning
Alzheimer's disease alters how a person processes and produces language. Current language models can identify these alterations by examining texts, such as descriptions that patients make of images. However, there is a risk that these systems rely on superficial patterns of the text and not on the real semantic deterioration, which would limit their value for diagnosis. 🔍
An Approach to Isolate Real Meaning
To verify if the models capture the underlying meaning, the original texts are transformed. Their syntax and vocabulary are altered, but their semantic content is preserved. Although superficial metrics indicate that the text is very different, semantic similarity scores remain high. When evaluating the models with these transformed texts, their ability to detect Alzheimer's persists, with only slight variations. This indicates that the models do use semantic indicators and not just the superficial form of language.
Key Findings of the Method:- The texts are modified to change their structure but preserve their meaning.
- The models' ability to classify remains stable, suggesting they detect semantic deterioration.
- This process allows filtering out spurious correlations and focusing on what really matters.
Even when the words change completely, the blurred message reveals the problem.
Verbal Descriptions Do Not Reconstruct the Visual Image
The study also explores whether a verbal description contains enough details for a generative model to reconstruct the original image. The results show that visual elements are largely lost. When these regenerated images are used to create new descriptions, noise is introduced and the accuracy for classifying Alzheimer's decreases. This confirms that the key information resides in the language, not in an imperfect visual recreation.
Implications of the Visual Finding:- The visual information degrades when passing through a textual description.
- Using regenerated images as a source introduces noise and reduces diagnostic accuracy.
- The language is the main biomarker, more reliable than trying to recreate the visual scene.
Towards a More Accurate Diagnosis
This approach allows validating that AI models capture the weakening of meaning in language, an early sign of Alzheimer's. By confirming that they do not depend on superficial artifacts, their potential clinical utility is improved. The technique underscores that, even when completely altering the words, the loss of semantic coherence remains as a detectable signal. 🧠