D. Devatha, J. Xiao
Dementia affects more than 55 million people worldwide, and its progressive decline is difficult to track using infrequent in-person assessments, which can miss subtle changes between visits and add to clinician burden. One widely used measure, the Mini-Mental State Examination (MMSE), is administered intermittently and remains subject to inconsistent scoring judgment, limiting early detection. Prior speech-based machine learning approaches have largely focused on cross-sectional classification rather than longitudinal cognitive forecasting. We introduce a longitudinal, patient-level framework that combines transformer-derived semantic speech representations with longitudinal speech-change features and clinical history to forecast a patient's future MMSE score from their history of prior visits. To our knowledge, this is the first framework to unite transformer-derived speech encoding with longitudinal forecasting of cognitive severity, rather than single-visit classification alone. We evaluate this framework on longitudinal transcripts from the DementiaBank Pitt Corpus using a LightGBM gradient-boosted regression model, validated with patient-grouped cross-validation to prevent identity leakage between training and evaluation folds. The model forecasts future MMSE scores with high accuracy and stability across folds (R^2 = 0.840 +- 0.015, RMSE = 2.75, MAE = 2.02, Pearson r = 0.917), with transformer-derived speech representations contributing meaningful predictive signal alongside clinical history. Ablation analysis demonstrated that transformer-derived speech representations provided complementary predictive information beyond clinical variables. These results establish speech as a viable longitudinal digital biomarker of cognitive decline, offering a low-burden complement to intermittent clinical assessment that could enable earlier detection and more frequent monitoring, supporting better-timed care decisions for patients with dementia.