Search papers, labs, and topics across Lattice.
This paper benchmarks zero-shot LLMs against supervised tabular models for cognitive impairment (CI) detection using speech transcripts in English, Slovene, and Korean. They find that while LLMs offer competitive zero-shot performance, supervised tabular models leveraging engineered linguistic features and embeddings achieve superior results, especially with feature fusion. Furthermore, the benefit of few-shot learning with embeddings is language-dependent, highlighting the continued importance of structured linguistic features in low-resource CI detection.
Despite the rise of LLMs, structured linguistic features and simple fusion-based classifiers still outperform them in low-data cognitive impairment detection across multiple languages.
We evaluate cognitive impairment (CI) classification from transcripts of speech in English, Slovene, and Korean. We compare zero-shot large language models (LLMs) used as direct classifiers under three input settings -- transcript-only, linguistic-features-only, and combined -- with supervised tabular approaches trained under a leave-one-out protocol. The tabular models operate on engineered linguistic features, transcript embeddings, and early or late fusion of both modalities. Across languages, zero-shot LLMs provide competitive no-training baselines, but supervised tabular models generally perform better, particularly when engineered linguistic features are included and combined with embeddings. Few-shot experiments focusing on embeddings indicate that the value of limited supervision is language-dependent, with some languages benefiting substantially from additional labelled examples while others remain constrained without richer feature representations. Overall, the results suggest that, in small-data CI detection, structured linguistic signals and simple fusion-based classifiers remain strong and reliable signals.
