How Can AI Boost Language Analysis in Individuals with Cognitive Deficits?

In recent years, scientists have started using audio recordings from neuropsychological tests to better detect signs of possible cognitive deficits. The catch? There is still no single, shared vocabulary to describe language difficulties, and clinicians often use different terms. This makes it hard to train AI algorithms that can spot the patterns.

On top of that, the available databases are usually small or unbalanced: few patient examples and many healthy ones. When data are scarce, machine‑learning models struggle to deliver reliable results.

To overcome these hurdles, a research team tried a new "data augmentation" strategy (i.e., artificially increasing the number of examples). They took the test texts from the Pitt Corpus and enriched them with typical symptoms of subjects with cognitive deficits, using large language models (LLMs) like ChatGPT.

The outcome was striking: models trained on these augmented texts classified subjects with 87.4% accuracy, an F1‑score of 87.4%, and similarly high precision and recall. In practice, AI distinguished the two groups better than traditional augmentation methods.

The study shows that, when data are limited, inserting typical linguistic traits of patients into texts can significantly boost model performance. It’s a step forward toward faster, less invasive diagnoses thanks to AI.

The proposed method relies on a textual augmentation pipeline driven by Large Language Models (LLMs) pre‑trained on massive linguistic corpora. In practice, the original neuropsychological test texts are analyzed to pinpoint syntactic and lexical structures typical of subjects with cognitive deficits. Then, the LLM generates alternative versions of the texts, inserting known linguistic symptoms—such as more frequent pauses, use of generic words, agreement errors, or lexical simplifications. This approach differs from traditional augmentation techniques like random synonym replacement or back‑translation because it embeds specific clinical knowledge.

Statistically, the influx of synthetic data reduces over‑fitting, enhancing classifier generalization. Classification models employed—typically Support Vector Machines or deep neural networks—benefited from a more balanced class distribution, improving their ability to learn discriminative patterns. The reported metrics (accuracy 0.8742, F1‑score 0.8744, precision 0.8736, recall 0.8781) indicate near‑identical performance across all indicators, a sign of a well‑balanced model.

A crucial aspect is the clinical validation of the inserted symptomatology. The authors collaborated with neuropsychologists to define a set of "linguistic features" typical of cognitive deficits, based on clinical linguistics studies. These features include reduced syntactic complexity (shorter sentences), decreased lexical variety, and increased morphosyntactic errors. The LLM was then fine‑tuned on a small corpus of clinical texts to ensure that the generated modifications were realistic and consistent with patient profiles.

Finally, the work highlights the potential of LLMs in low‑resource scenarios. When large datasets are unavailable, using generative models to create targeted synthetic data can bridge the gap between training data needs and real‑world availability, paving the way for more reliable decision‑support systems in neuropsychology.