Please join us online for a talk by . He will present 鈥淲hen Did the Stress Begin? Statistical Learning for Passively Sensed mHealth Data in Late-Life Mood Disorders鈥� on Wednesday, September 30, at 1:30 pm Central Time. For access to this seminar, contact series administrator Cierra Streeter.
Samprit Banerjee, PhD, is Associate Professor of Biostatistics in Population Health Sciences and in Psychiatry at Weill Cornell Medicine. His methodological work concerns passively sensed and ecological momentary assessment data from smartphones and wearables: pre-processing of commercial-device data, change-point and segmentation methods for nonstationary behavioral time series, and semi-supervised and transfer learning for individualized prediction. He collaborates broadly on randomized clinical trials in late-life mental health and on large-scale observational research in aging, and has published more than 140 peer-reviewed articles. He directs the methods and data cores of several federally funded centers;, he has served on three FDA advisory committee panels, an NIH standing study section, and five data and safety monitoring boards;, and he reviews for PCORI methods grant cycles.
Abstract:
Passively sensed data from smartphones and wearables offer a way to monitor behavior continuously in the natural environment of patients with mood disorders, for whom the timing of emotional stress episodes is clinically consequential and rarely observed. These data are difficult in two ways: unknown device non-use produces hidden missingness and downward bias, and the association between behavior and self-reported stress is nonstationary, present only within limited windows of time. After a brief account of 2SpamH, which treats device non-use as a missing-data problem with unobserved labels, this talk focuses on two methods for locating those windows within a single patient, developed on data from the Weill Cornell ALACRITY Center. A co-segmentation algorithm shifts the unit of analysis from the time point to a data-driven segment, validates candidate segments through sequential tests of association and distributional change, and uses the resulting segment labels to predict future stress states, outperforming methods that ignore segmentation. Hotspots are then defined as intervals over which the mean and variance of a passive sensing series and of the stress series shift jointly, and are estimated by extending moving-sum statistics within and across series without specifying a functional relationship between them; two complementary aggregation rules are studied by simulation and applied to ALACRITY data. Together the methods give an interval-valued, individualized answer to the question of when a stressful period began.