Depression screening via speech is promising for elderly patients, but many also live with mild cognitive impairment (MCI). The new arXiv preprint warns that the two conditions produce overlapping acoustic changes, so a model that has not been trained to separate them may conflate one with the other.
The authors argue that classifiers are especially prone to learning features of cognitive decline rather than depression itself. This could lead to inaccurate diagnoses and inappropriate treatment plans.
To address this, they suggest collecting concurrent clinical assessments—measuring both depression and cognitive status at the same time—so the model can learn which speech patterns belong to which condition. The paper is at arXiv:2610.26890.