JMIR Publications

ABSTRACT

Background:

Remote digital phenotyping has expanded the scalability of cognitive neuroscience studies. However, the integrity of millisecond-level response time (RT) data relies on the client-side graphics pipeline. When local Graphics Processing Units (GPUs) become unavailable, operating systems and web browsers silently transition to software-based Central Processing Unit (CPU) renderers. The extent to which these software fallbacks corrupt behavioral metrics remains unquantified.

Objective:

To characterize the technical constraints of software-based rendering architectures and evaluate their systemic impact on the data validity of remote, web-based cognitive tasks.

Methods:

We evaluated behavioral outcomes and technical paradata across two studies using our custom Adaptive Cognitive Evaluation-Explorer (ACE-X) platform. Study 1 utilized a naturalistic longitudinal sample (N = 864,702 trials; n = 277 participants) to observe real-world performance under the legacy software-based Google SwiftShader renderer. Study 2 employed a controlled, within-subjects experimental design (N = 4,089 trials; n = 74 participants) on Windows machines running Google Chrome to isolate hardware acceleration (native GPU) against software-based rendering (Windows Advanced Rasterization Platform (WARP)). Statistical profiling was conducted using stratified outlier removal and linear mixed-effects models (LMMs) with log-transformed RTs.

Results:

In Study 1, software rendering with SwiftShader introduced a massive, statistically significant delay, increasing baseline reaction times by 171.23% (β = 0.9978, P < .001), yielding an average hardware penalty of 515 ms (816 ms CPU vs 301 ms GPU). Study 2 experimentally validated this behavior, showing that WARP significantly inflated reaction times by 39.60% (β = 0.3336, P < .001), yielding a baseline penalty of 107 ms (377 ms CPU vs 270 ms GPU). Software rendering increased visual frame instability (FPS (frames per second) Coefficient of Variation) by over 1.5 standard deviations (P < .001). Furthermore, the integration of random slopes demonstrated that individual participant reaction times varied heterogeneously in response to this hardware-induced jitter (P < .001).

Conclusions:

Software-based rendering pipelines act as destructive technical artifacts in digital research, introducing profound, non-uniform delays and visual stutters that mask true psychophysiological signals. Because high individual heterogeneity renders uniform post-hoc linear corrections mathematically invalid, researchers collecting high-resolution timing data on varying hardware must actively capture graphics paradata and exclude software-rendered sessions. Ultimately, these mitigation strategies must be balanced with health equity considerations, as systematic data exclusion risks underrepresenting populations with restricted access to optimized hardware or stable device configurations.