New strides in wearable sensor technology

Much like the early descriptions of the Shaking Palsy by James Parkinson, which were derived from 6 case studies, 3 of which were observations of people on the streets of London, the assessment and quantification of Parkinson’s disease (PD) symptoms is usually composed of a discrete snapshot of symptom presentation over a short epoch. Yet it is well appreciated that both the motor and nonmotor symptoms that make up the parkinsonian syndrome can vary greatly on a minuteto-minute or hour-to-hour time scale. This poses a considerable challenge for the development and testing of new treatment interventions because many of the most debilitating symptoms of advanced PD, such as dyskinesia, on–off fluctuations, and freezing of gait, are highly episodic and thus are not easily or reliably captured during a single clinical or laboratory visit. For the most part, randomized controlled trials of therapeutic interventions have relied on daily diaries to provide a temporal profile of the incidence and severity of symptoms. Yet this method is fraught with compliance and bias problems and lacks sufficient temporal resolution to capture transient episodes of impairment. Recent advances in wearable sensor technology are reshaping the way we can monitor and assess patient status around the clock. Wearable devices are now commercially available that contain gyroscopes, inertial sensors, accelerometers, global positioning systems, or electromyographic (EMG) sensors. Developments in microelectronics have resulted in miniaturization of many of these sensors, so that 1 or more can be embedded into microelectromechanical or systemon-chip systems that can provide real-time signal processing (for a review, see Patel et al, 2012). Furthermore, some systems can be mounted on flexible circuit boards or embedded in textiles (e-textiles), thus allowing for greater comfort and unrestricted movement. When this technology is combined with Moore Law’s growth in the capacity to store and wirelessly transmit information, our capacity to monitor patients on a 24/ 7 basis will likely have few, if any, limitations. The issue that is more likely to slow the implementation of advanced patient monitoring and assessment technologies in the field of movement disorders is the lack of validated data-processing methods that provide measures of spontaneous changes in impairment with high construct validity. In other words, we lack algorithms that can derive valid and reliable signatures of movement impairment from a large data set obtained in freely behaving and interacting individuals. Part of the problem is that the dependent variables that have been shown to discriminate abnormal patterns of movement in people with PD from controls with high sensitivity and specificity were obtained in a laboratory setting (eg, Robichaud et al, 2009; Delval et al, 2010). Many of the wearable sensors that are currently available were developed to capture a single or small subset of motor features (eg, tremor) and were tested and validated using scripted activities. The article by Roy and colleagues in this issue provides an example of the next generation of patient monitoring and assessment technologies that combine multicomponent wearable sensor systems with advanced algorithms for post hoc feature extraction and classification. In this case, triaxial accelerometer and EMG signals were obtained in freely moving individuals with PD. The signals were used as input to dynamic neural networks (DNNs), a form of machine learning, which have the capacity to learn time-dependent relationships between the signals. Episodes of tremor or dyskinesia that were identified using the DNN algorithms could subsequently be categorized as being mild, moderate, or severe using Bayesian classifier methods, thus providing an ordinal measure comparable to clinical ratings but with 1-second time resolution. The main finding was that the DNN algorithms could identify episodes of tremor and dyskinesia with high sensitivity (94.9%) and specificity (97.2%). Although these impressive results were obtained in only a limited sample of patients and were restricted to the classification of tremor or dyskinetic events, these findings demonstrate the utility and viability of this approach. Algorithms that have the capacity to classify changes in movement speed -----------------------------------------------------------*Correspondence to: Dr. Colum D. MacKinnon, Department of Neurology, University of Minnesota, 717 Delaware Street SE, Minneapolis, MN 55455, USA; cmackinn@umn.edu