Computer keyboard interaction as an indicator of early Parkinson’s disease L. Giancardo1,* , T. Arroyo-Gallego2 , I. Butterworth1 , C. S. Mendoza1 , P. Montero3 , M. 4,5 ´ Matarazzo4,5 , and A. Sanchez-Ferro
arXiv:1604.08620v1 [cs.HC] 28 Apr 2016
1 Madrid-MIT
´ Consortium, Research Laboratory of Electronics, Massachusetts Institute of Technology, M+Vision Cambridge, MA, USA 2 Universidad Politecnica ´ de Madrid 3 Movement disorders unit, Hospital Clinico San Carlos, Madrid, Spain 4 HM Hospitales - Centro Integral en Neurociencias HM CINAC, Mostoles, ´ Madrid, Spain 5 Instituto de Investigacion ´ Hospital 12 de Octubre (i+12), Madrid, Spain *
[email protected]
ABSTRACT Parkinson’s disease (PD) is a slow progressing neurodegenerative disease with motor symptoms from its initial stages. Objective measurements of motor symptoms are of vital importance for diagnosing, monitoring and developing disease modifying therapies. This is particularly true for the early stages of the disease when neuroprotective treatments could stop the death of dopaminergic neurons. Current medical practice has limited tools to monitor at-risk patients with high frequency and without adding additional burdens to the routine of patients. In this paper, we present data indicating that the daily interaction with computer keyboards can be used to measure motor symptoms in the early stages of PD. We explore a solution that monitors the normal use of a computer without any change in the hardware and converts it to a PD motor index. This is achieved by the automatic discovery of patterns in the time series of key hold times, i.e. the time required to press and release keys, by an ensemble regression algorithm. The diagnosis performance of our algorithm with early PD patients and controls is comparable to performance reported for trained physicians.
Introduction Parkinson’s disease (PD) is the second most prevalent neurodegenerative disorder in the western world1 . Subtle motor signs can precede the clinical diagnosis by several years and continue throughout the course of the disease, however they often go unnoticed particularly in the early stages2–4 . After this point of diagnosis, the patient follows a progressive course leading to severe disability and death after an average of 7 and 16 years respectively5, 6 . A number of drugs are available for symptomatic relief, including Levodopa, dopamine agonists or MAO-B inhibitors7 . These types of treatments administered by a specialist significantly lowered the risk of hip fractures, admissions to skilled nursing facility and increased survival rates8 . An accessible way to precisely quantify PD motor symptoms in the patient’s home would bring significant benefits to therapy management, better diagnosis and potentially earlier detection of the symptoms, thus enabling the development of new therapies9, 10 . The current standard to quantify PD motor symptoms is the Unified Parkinson’s Disease Rating Scale part III (UPDRS-III)11 , a compound clinical score that evaluate various aspects of the disease, such as rigidity, resting tremors, speech and facial expression among others. This test requires trained medical personnel, attendance of the patient in the clinic, and it cannot be administered with a high frequency. This limits the ability of running longitudinal clinical studies measuring symptoms with a time resolution lower than 3 months12 . Digital technologies for objectively quantify PD motor symptoms exist and new ones are being developed. The oldest and most validated is arguably finger tapping, where subjects are asked to intermittently press buttons as fast as possible for a given time13 . More recently, wearable inertia measurement units (IMUs) have been employed to measure information about gait, posture, tremors, bradykinesia (slow movements) and dyskinesias (involuntary movements)14, 15 . Typically, multiple sensors are applied on various areas of subject’s body who is then asked to perform a particular task. IMUs can also be found in modern smartphones, which has motivated attempts to combine finger tapping and IMUs in a single device, in some cases also including voice utterances tests16 . In the last 30 years, typing cadence (also known as keystrokes dynamics) has been studied by various research groups and employed commercially mainly as a way to replace or strengthen passwords17, 18 . Applications to the biomedical field are almost non-existent, one exception are Austin et al. who used the typing speed in login sessions to evaluate sensory-motor
speed in healthy subjects19 . In our previous work20 , we showed how to use the interaction with keyboards to detect a state of psychomotor impairment regardless of the typing speed, language used or text typed. In this work, we demonstrate the ability to distinguish PD patients at the early stage of the disease from matched healthy controls by monitoring their natural interactions with standard keyboards. We record the hold time (HT) occurring between pressing and depressing a key while operating an unmodified word processor, then we convert these variables to the numerical neuroQWERTY index (nQi) employing a novel algorithm. The system automatically learns by example the PD typing patterns by comparing the PD subjects with a control group with similar typing skills and education. Our approach does not require information of the text content and works by collecting the time series of HT events, which typically accounts for around 100 milliseconds21, 22 . Fig. 1(a) shows how the neuroQWERTY index is computed. First, the HT time series is converted into matrix form, then an ensemble regression algorithm identifies the likelihood of PD patterns and generates a single numerical score for each matrix column representing 90 seconds of typing. The results obtained are compared to the clinical ground truth (the Unified Parkinson’s Disease Rating Scale part III) and two quantitative motor tasks used in clinical studies for evaluating the PD progression and typing skills of our cohort.
Results Fig. 1(b-c) show examples of the feature matrices derived from the hold time probability occurring during ∼10 minutes of typing. Each column xt of the matrix accounts for 90 seconds of typing and it comprises of an upper part containing the HT probability density estimation and the bottom part containing 5 descriptors of the HT variance (see Methods). By visually inspecting the matrices, a skewness towards higher value in the HT probability densities and higher values for the variance descriptors are visible. The nQi scores computed with our algorithm are shown at the bottom of the matrices. nQi scores makes the evaluation of the HT readily quantifiable, even for patients with very mild symptoms. The algorithm to generate the nQi score from xt starts with Principal Component Analysis (PCA), an unsupervised eigenvectors-based multivariate approach to detect the subspace components with the highest variance23 . Fig. 2(a) shows the projection of 570 xt vectors from 18 early PD and 13 matched controls on the first two PCA components. It can be seen that only the samples coming from the Parkinson’s group tend to go toward higher values in the PCA space, while the samples from controls appear much less disperse. Fig. 2(b) confirms this observation by showing the two distributions on the first PCA component only. These effects do not appear to be measurable by typing speed. In Fig. 3, the joint typing time series for early PD and control groups are shown. The nQi scores consistently partition the two groups with higher values for PDs and lower values for controls while typing speed, a common measure of typing skills, does not allow to distinguish them. These graph are generated using 90 seconds non-overlapping time windows to compute both nQi and typing speed in order to show the progression through time. When each 90 seconds window is independently used to classify a subject as early PD or CNT we obtained a Area under the Receiving Operating Characteristic Curve (AUC) of 0.86 (0.83-0.88 95% CI, p-value