PulseImpute: A Novel Benchmark Task and Architecture for Imputation of Physiological Signals

Maxwell Xu, Alexander Moreno, James Matthew Rehg

Keywords: missingness, imputation, mHealth, sensors, transformer, self-attention

Abstract: Providing care for patients with chronic diseases is one of the biggest drivers of the nation’s rising healthcare costs, but many of these diseases are linked to mutable health behaviors. Mobile health (mHealth) biophysical sensors that continuously measure our current conditions provide the framework for a personalized guidance system for the maintenance of healthy behaviors. However, this physiological sensor data is plagued with missingness due to insecure attachments, wireless dropout, battery, and adherence issues. These issues cripple their rich diagnostic utility as well as their ability to enable temporally-precise interventions. While there is a sizable amount of research focusing on imputation methods, surprisingly, no works have addressed the patterns of missingness, quasi-periodic signal structure, and the between subject heterogeneity that characterizes physiological signals in mHealth applications. We present the PulseImpute Challenge, the first challenge dataset for physiological signal imputation which includes a large set of baselines' performances on realistic missingness models and data. Next, we demonstrate the potential to address this quasi-periodic structure and heterogeneity with our Dilated Convolution Bottleneck (DCB) Transformer, a transformer architecture with a self-attention mechanism that is able to attend to corresponding waveform features in quasi-periodic signals. By utilizing stacked dilated convolutions with bottleneck layers for query and key transformations, we visually demonstrate that the kernel similarity in the attention model gives high similarity to similar temporal features across quasi-periodic periods. We hope the release of our challenge task definitions and baseline implementations will spur the community to address this challenging and important problem.

One-sentence Summary: We present PulseImpute, a benchmarking challenge for the imputation of biophysical signals, and propose a novel self-attention module for attending over quasi-periodic signals.