Multilingual Cognitive Impairment Detection in the Era of Foundation Models

TL;DR

This study compares zero-shot large language models and supervised tabular models for multilingual cognitive impairment detection from speech transcripts, highlighting the effectiveness of structured linguistic features and fusion methods in small-data scenarios.

Contribution

It provides a comprehensive evaluation of multilingual CI detection methods, emphasizing the strengths of supervised models with linguistic features over zero-shot LLMs in low-data contexts.

Findings

Supervised models outperform zero-shot LLMs in multilingual CI detection.

Engineered linguistic features combined with embeddings improve classification accuracy.

Few-shot learning benefits vary across languages, depending on available labeled data.

Abstract

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.