Dementia Language Models: a generalizable and controllable representation of cognitive impairment

Abstract

We introduce Dementia Language Models (DeLMs)--generalizable and controllable representations of cognitive impairment through language--alongside an evaluation framework for establishing their validity and clinical grounding. DeLMs created by fine-tuning large language models on a small clinical corpus successfully generated patient-like narratives across unseen tasks, received predicted Mini-Mental State Examination (MMSE) scores in the impaired range, and produced narratives that neurologists identified with accuracy comparable to real transcripts. The models' internal representations, as well as their non-linguistic decision-making, supported mild cognitive impairment detection in unseen cohorts. The effect was controllable: moving from Healthy toward Dementia in weight space progressively worsened language and predicted MMSE scores while increasing dementia probability. DeLMs could support clinician training, hypothesis generation, and scalable experimentation, reserving patient involvement for where it is truly needed.