FIDM model reads cognition from speech, boosting cognitive screening accuracy

Inverse decision modeling has long been used to infer latent properties of decision-making from observed behavior, but existing methods focus almost exclusively on action trajectories. That leaves crucial signals unexamined in verbal tasks—hesitations, repetitions, verbal production, and interaction dynamics. Researchers from The Chinese University of Hong Kong present the Factorized Inverse Decision Model (FIDM), a new framework that decomposes each individual's task-execution likelihood into an action factor and an effort factor, each governed by separate individual-specific parameters. This factorization allows the model to selectively estimate distinct cognitive dimensions from the same raw transcript.

FIDM works by using a language model to convert raw verbal transcripts into structured task-execution traces, on which factorized inference is performed. The team validated FIDM on data from 400 older adults engaged in a grocery-shopping dialog task designed for cognitive screening. Controlled recovery experiments confirmed that FIDM can selectively estimate the intended factors, and semi-synthetic matched conditions showed it preserves action-execution distinctions even when aggregate behavioral summaries are identical. In cognitive-status classification, FIDM provided information complementary to clinical scores, trajectory summaries, and frozen language representations, yielding consistent gains across every baseline evaluated in the binary setting. The paper, available on arXiv (2608.09222), positions FIDM as a promising tool for dementia screening and broader cognitive assessment from natural conversation.