AUTHOR=Iakovakis Dimitrios , Hadjidimitriou Stelios , Charisis Vasileios , Bostantjopoulou Sevasti , Katsarou Zoe , Klingelhoefer Lisa , Reichmann Heinz , Dias Sofia B. , Diniz José A. , Trivedi Dhaval , Chaudhuri K. Ray , Hadjileontiadis Leontios J. TITLE=Motor Impairment Estimates via Touchscreen Typing Dynamics Toward Parkinson's Disease Detection From Data Harvested In-the-Wild JOURNAL=Frontiers in ICT VOLUME=5 YEAR=2018 URL=https://www.frontiersin.org/journals/ict/articles/10.3389/fict.2018.00028 DOI=10.3389/fict.2018.00028 ISSN=2297-198X ABSTRACT=

Parkinson's Disease (PD) is a neurodegenerative disorder with early non-motor/motor symptoms that may evade clinical detection for years after the disease onset due to their mildness and slow progression. Digital health tools that process densely sampled data streams from the daily human-mobile interaction can objectify the monitoring of behavioral patterns that change due to the appearance of early PD-related signs. In this context, touchscreens can capture micro-movements of fingers during natural typing; an unsupervised activity of high frequency that can reveal insights for users' fine-motor handling and identify motor impairment. Subjects' typing dynamics related to their fine-motor skills decline, unobtrusively captured from a mobile touchscreen, were recently explored in-the-clinic assessment to classify early PD patients and controls. In this study, estimation of individual fine motor impairment severity scores is employed to interpret the footprint of specific underlying symptoms [such as brady-/hypokinesia (B/H-K) and rigidity (R)] to keystroke dynamics that cause group-wise variations. Regression models are employed for each fine-motor symptom, by exploiting features from keystroke dynamics sequences from in-the-clinic data captured from 18 early PD patients and 15 healthy controls. Results show that R and B/H-K UPDRS Part III single items scores can be predicted with an accuracy of 78 and 70%, respectively. The generalization power of these trained regressors derived from in-the-clinic data was further tested in a PD screening problem using data harvested in-the-wild for a longitudinal period of time (mean ± std : 7 ± 14 weeks) via a dedicated smartphone application for unobtrusive sensing of their routine smartphone typing. From a pool of 210 active users, data from 13 self-reported PD patients and 35 controls were selected based on demographics matching with the ones in-the-clinic setting. The results have shown that the estimated index achieve {0.84 (R), 0.80 (B/H−K)} ROC AUC, respectively, with {sensitivity/specificity : 0.77/0.8 (R), 0.92/0.63 (B/H−K)}, on classifying PD and controls in-the-wild setting. Apparently, the proposed approach constitutes a step forward to unobtrusive remote screening and detection of specific early PD signs from mobile-based human-computer interaction, introduces an interpretable methodology for the medical community and contributes to the continuous improvement of deployed tools and technologies in-the-wild.