Free-Text Smartphone Keystroke Dynamics for Cognitive Impairment Pre-Screening with LLM-Assisted Text Analysis.

Abstract

To evaluate whether free-text smartphone keystroke dynamics, augmented by LLM-assisted text analysis, can support low-burden pre-screening for mild cognitive impairment (MCI) under more natural typing conditions than fixed-text copy tasks. We developed a framework that combines timing-based keystroke features with LLM-assisted text-derived features while reducing reliance on bespoke language-specific natural language processing. Using free-text smartphone typing data from aKorean-speaking cohort comprising cognitively normal (CN) participants and participants with protocol-confirmed MCI, weaggregated participant-level features across repeated sessions and evaluated three proposed digital biomarker configurations under a common XGBoost classifier. The proposed free text keystroke-derived biomarkers showed meaningful utility for participant-level cross-sectional MCI pre-screening. Timing-based biomarkers and LLM-assisted biomarkers achieved ROC-AUCs of 0.737 and 0.703, respectively, and their combination improved the ROC-AUC to 0.771. Free-text smartphone keystroke derived biomarkers can capture clinically relevant signals associated with cognitive impairment pre-screening, with timing-based biomarkers providing a strong core signal and LLM-assisted analysis contributing complementary information when automated text processing is acceptable. This work presents a practical biomedical engineering route toward low-burden adjunctive pre-screening for cognitive impairment on personal smartphones using ecologically collected typing behavior. Because the present evaluation was conducted in a Korean-speaking cohort, broader multilingual generalizability remains to be established.