A Touchscreen-Based Framework for Accurate and Transferable Cognitive Prescreening Using Drawing Behavior Encoding

Overview

Cognitive decline affects more than 55 million people worldwide, yet early screening still depends on clinical infrastructure that limits access. In the United States, an estimated 7.4 million cases of mild cognitive impairment in adults age 65 and older remain undiagnosed. This research presents a novel smartphone-based framework for cognitive prescreening that is accurate, accessible, and transferable across datasets and neurological conditions. Adults aged 60 and older (N=185) completed four touchscreen drawing tasks through a custom web application that recorded touch trajectories, timing, and pressure. After demographic filtering and data quality review, 134 subjects were retained for model development using 5-fold subject-level cross-validation. To capture both what was drawn and how it was produced, I designed a renderer that converts each drawing into an image where velocity is encoded as color and pressure as stroke width. Across 16 rendering strategies, this representation performed best. This study is the first in smartphone-based cognitive prescreening to use this combined rendering strategy and validate it through a systematic renderer comparison. Using a fixed image embedding with a low-dimensional linear readout produced the strongest model, achieving AUC 0.968 ± 0.008, 96.8% sensitivity, and 90.3% specificity. Robustness testing across validation splits, bootstrap resampling, and backbone variants showed that the result was stable and reproducible. The same learned representation also generalized beyond the original cohort without retraining, achieving AUC 0.836 on an independent clinician-verified mild cognitive impairment dataset of 918 subjects and AUC 0.911 on a Parkinson’s disease dataset of 204 spiral drawings in a zero-shot transfer test. These findings show that short touchscreen drawing tasks can capture robust and transferable biomarkers of neuro-motor and cognitive impairment.