0377 The Modular Actigraphy Platform for Raw Wearable Sensor Data Processing and Sleep Estimation

Abstract Introduction Adequate sleep is essential for childhood development, yet wearable-based assessments often depend on proprietary algorithms. To enhance reproducibility, the field is moving toward open-source solutions. We developed the Modular Actigraphy Platform (MAP), an open-source tool that makes processing raw data easier and efficient. This study evaluates MAP’s performance using pediatric sensor data. Methods MAP is deployed on Google Cloud Platform and comprises four core modules: (1) data pre-processing, (2) non-wear detection, (3) sleep scoring, and (4) physical activity scoring. Each module is Docker-containerized, with Google Kubernetes Engine providing orchestration and Argo to manage workflow execution. The pipeline incorporated the GGIR algorithm and the Monitor Independent Movement Summary (MIMS) algorithm to process raw data for sleep and physical activity estimation. To demonstrate performance, we processed tri-axial acceleration wrist-worn data from ActiGraph and GENEActiv devices and compared MAP’s outputs to offline data processing benchmarks. Testing included unit, integration, and system checks, followed by alpha and beta testing. The goal of alpha testing was to evaluate MAP in a development environment. Alpha testing used 17 CHOP-generated files (4.9 GB, 1–14 days). Beta testing assessed functionality and scalability on real-world datasets from four cohort studies (686 files, each with up to 4 days of data at a 30–50 Hz sampling rate). Results MAP passed unit testing for stability, functionality, and security. Preprocessing was the most computationally demanding module, requiring 41–61% of the total processing time. Beta testing required up to 60 CPUs and 120 GiB memory for preprocessing and 40 CPUs/100 GiB for the other modules. The memory-to-CPU ratio was critical for avoiding crashes. Performance evaluations showed that MAP was faster than offline processing. The offline GGIR pre-processing was 1.6 to 2.9 times slower compared to MAP. MIMS processing was most resource-intensive, requiring up to 20 CPUs and 500 GiB of memory. For MIMS unit processing, MAP processed 243.3 GB in 444 minutes versus 6007 minutes offline (2.4–14× faster). Conclusion Raw actigraphy data increasingly support sleep estimation but require advanced expertise. MAP reduces barriers by offering an accessible, efficient cloud-based solution for deriving sleep metrics from raw sensor data. Support (if any)