A Multi-Stage Feature Pipeline on Timestamped Speech Transcriptions for Dementia Assessment

This paper details our approach to the "Prediction and Recognition of Cognitive Decline through Spontaneous Speech (PROCESS)" Grand Challenge at ICASSP 2025, which focuses on early dementia detection using speech alone. The challenge comprises two tasks: (1) classifying individuals as healthy controls (HC), with mild cognitive impairment (MCI), or with dementia, and (2) predicting Mini-Mental State Exam (MMSE) scores. We focused on the latter, developing a robust pipeline that leverages a diverse set of linguistic and temporal features extracted from transcribed speech. Our ensemble model scored at the top for the MMSE prediction task with a root mean squared error (RMSE) of 2.46 on the held-out test set while maintaining diagnostic transparency and real-world clinical applicability.