Software Reference Architecture for Real-Time Mobile Digital Phenotyping: Evaluation of System Designs.

Background: Digital phenotyping-the use of continuous data streams from digital devices such as smartphones to assess behavioral, psychological, and physiological states-holds transformative potential for health monitoring and personalized care. However, real-time analysis of large multimodal data often exceeds mobile devices' computational resources, leading most platforms to rely on sequential processing and cloud-based computation. Objective: We propose the Stanford Screenomics platform as a software reference architecture that uses a modular design to integrate parallel processing and edge computing, enabling scalable, real-time digital phenotyping on smartphones. Methods: Two prototype apps were developed: one following the parallel, on-device architecture (Stanford Screenomics platform) and another based on a traditional sequential, cloud-based design (traditional). Both processed identical multimodal data streams at the same intensity; only the location and sequence of computation differed. In two 48-hour experiments, performances were compared across four load profiles: low (≈10 MB/min), medium (≈30 MB/min), heavy (≈40 MB/min), and very heavy (≈60 MB/min). In the first experiment, offline resource performance was assessed under continuous simulated smartphone use. Virtual users completed six tasks in a fixed five-minute sequence: watching YouTube (Google LLC), reading eBooks, browsing TikTok (ByteDance Ltd), web surfing, listening to Spotify, and scrolling Instagram Reels (Meta). Minute-by-minute measurements of CPU usage (%), RAM usage (MB), battery drain (%/h), and data loss (%) were collected. Descriptive statistics (mean±SD) summarized performance, and independent t tests compared architectures. Data loss trajectories were analyzed to determine whether growth was linear or exponential under increasing load. In the second experiment, end-to-end phenotyping latency was evaluated over stable Wi-Fi. Five key-stage timestamps per trial tracked local writes, preprocessing, memory parsing, phenotype analysis, and intervention delivery. Total phenotype update time per trial was the primary outcome, and latency differences between architectures were analyzed using linear mixed-effects models, with IQRs reported to capture variability across load conditions. Results: The Stanford Screenomics platform consistently demonstrated lower CPU usage (3.9%-14.6% vs 10.5%-26.9%) and RAM usage (97-132  MB vs 101-155  MB) than the traditional, with reduced battery drain (0.9%-2.1%/h vs 1.4%-3.2%/h). Data fidelity was higher in the Stanford Screenomics, with shallow linear data loss (0.4%-1.5%/h) compared to exponential growth in the traditional (2%-7.1%/h), achieving up to 9.4× greater data retention under very heavy load. The Stanford Screenomics completed phenotype updates in 0.90 seconds under low load and 9.32 seconds under very heavy load, compared to 30.1-398.1  seconds for traditional, representing 34-43×faster processing with substantially narrower variability (IQR 0.3-6 s vs 11  s-5  min). Conclusions: These results demonstrate that the Stanford Screenomics platform architecture enables real-time, on-device digital phenotyping with high fidelity and low latency. This validated prototype architecture establishes a resilient foundation for the next generation of scalable, reliable, and context-aware deployment of real-world mobile health interventions on mobile devices.