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

Area of Science:

• Biomedical Engineering
• Digital Health
• Cognitive Science

Background:

• Mild cognitive impairment (MCI) poses a growing public health challenge.
• Early detection of MCI is crucial for timely intervention and management.
• Current pre-screening methods can be burdensome and may not reflect real-world cognitive function.

Purpose of the Study:

• To evaluate the efficacy of free-text smartphone keystroke dynamics, enhanced by LLM-assisted text analysis, for low-burden MCI pre-screening.
• To assess the utility of typing behavior under natural conditions for identifying cognitive impairment.
• To develop a practical digital biomarker for MCI detection on personal smartphones.

Main Methods:

• Developed a framework combining timing-based keystroke features with LLM-assisted text-derived features.
• Utilized free-text smartphone typing data from a Korean-speaking cohort (cognitively normal and MCI participants).
• Aggregated participant-level features and evaluated biomarker configurations using an XGBoost classifier.

Main Results:

• Free-text keystroke-derived biomarkers demonstrated significant utility for MCI pre-screening.
• Timing-based biomarkers achieved an ROC-AUC of 0.737; LLM-assisted biomarkers achieved 0.703.
• The combination of both biomarker types improved the ROC-AUC to 0.771.

Conclusions:

• Smartphone keystroke dynamics, including timing and LLM-analyzed text, can capture signals relevant to cognitive impairment.
• This approach offers a practical, low-burden method for adjunctive MCI pre-screening using everyday smartphone use.
• Further research is needed to establish multilingual generalizability.