A design pattern for detecting cognitive decline (healthy vs mild cognitive impairment vs dementia) from short spontaneous-speech recordings by fusing many complementary, deliberately heterogeneous representations rather than betting on a single model. Audio from several clinical elicitation tasks (picture description, phonemic and semantic fluency) is denoised and transcribed, then a broad pool of feature families is extracted in parallel: knowledge-based acoustic and prosodic descriptors and pause-based biomarkers, hand-crafted and LLM-derived linguistic descriptors from the transcripts, and neural embeddings of both speech (speaker/paralinguistic embeddings) and text (transformer language-model embeddings). Each family is paired with a classifier to form many single systems, and the strongest, most complementary systems are combined by late fusion - a logistic-regression meta-classifier over their soft outputs. The core lesson is that complementarity, not any one representation, drives performance: fusing interpretable pause and macrolinguistic features together with neural embeddings, drawn across all tasks, beats every single system and stays robust under severe class imbalance, where the minority (dementia) class is the hardest to recover. The contribution is the recipe and the empirical evidence for which representations carry independent signal, not a new model architecture.
This system-description paper reports a submission to the PROCESS Challenge 2024, whose task is three-class classification of spontaneous speech into healthy control, mild cognitive impairment (MCI),…