Wearable sensor devices are now commonplace, making possible comprehensive, dense assessments of multi-dimensional biosignals we collect from people in everyday settings. “In many large observational cohorts, participants' physical activity profiles are tracked via accelerometers or smart mobile devices. We measure activity as often as every five seconds, over a prolonged period of time,” says Haochang Shou, PhD. These prolific data carry implications for a wide range of outcomes in patients—including, notably, for their mental-health status.
But, Dr. Shou points out, these data can pose significant statistical challenges. “They often present with complex multivariate structures and are variable over time, study design and choice of devices,” she says. Clinicians who treat complex conditions such as mood disorders could learn much from accurate analyses; but it’s difficult for researchers to glean information that can be applied with patients. Hence the mission of Dr. Shou and colleagues at the NIMH Mobile Motor Activity Research Consortium for Health (mMARCH)—an international collaborative network of mental health studies that use sensor data. These researchers have undertaken a series of projects that address the challenges—allowing us to better understand how an individual's activities align with his or her clinical outcomes.
“We work to develop novel statistical methods that model sensor data from multiple scales and angles, in the contexts of hypothesis testing, regression and prediction. Our methods respect within-subject correlations of repeated observations and extract time-varying and distributional activity features that link with clinical endpoints, such as participants' behavioral and cognitive functions,” says Dr. Shou. In the context of mental health, this allows the researchers to identify activity “signatures” associated with moods such as sadness or anxiety—information that can offer valuable foresight for clinicians.