Scaling-Out Longitudinal Clinical Analytics with Dataflow Processing

摘要

There are two key ingredients in supporting high-frequency and continuous clinical assessment of patient populations at scale: first, the availability of validated metrics of disease progression which reliably capture the longitudinal variations of symptoms; and second, the ability to compute these metrics on the fly over multiple concurrent streams of sensor data captured at home or in the community. In this paper, we describe the design, development and validation of PDkit, a comprehensive data science toolkit for Parkinson's Disease, and explore the dataflow paradigm as a means to provide salable performance. Our aim is to contribute towards the development of robust clinical outcome measures for therapeutic trials and to support longitudinal investigations of disease mechanism through the analysis of data collected from wearables and smartphones. The PDkit is released as open source and offers a succinct interface for interactive collaborative data exploration. Moreover, it enables the composition of data processing pipelines for tremor, tapping, bradykinesia and gait tests with the view to support horizontal scalability over common Cloud infrastructures on production workloads. Specifically, we report on our early experiments executing PDkit pipelines using Apache Beam, a unified dataflow multi-runtime stream processing engine. Our long-term aim is to provide the PD research community with the tools needed to individually tailor treatment plans and to empower patients to become more involved in their own care.