Accuracy-Latency Trade-offs Under Neural Network Compression in Safety-Critical Edge AI Applications: A Controlled Simulation Study

Abstract Neural network compression is indispensable for deploying AI on resource-constrained edge hardware across safety-critical domains including healthcare monitoring, prosthetic limb control, educational robotics, and autonomous navigation. Standard compression evaluation relies on overall accuracy on balanced benchmark datasets, which cannot capture the failure modes that determine safety in deployment: sensitivity-specificity divergence in imbalanced clinical classifiers, hard latency deadlines on embedded processors, and cross-seed instability in multi-class controllers. The present study characterises these domain-specific failure modes empirically and introduces evaluation tools and a pre-deployment procedure to address them. We conducted a controlled simulation study on four synthetic domain-representative datasets: clinical fall event detection via 3-axis accelerometry (binary classification, 18% fall prevalence), prosthetic gesture decoding via 4-channel surface electromyography (5 classes), intelligent tutoring system problem-type classification (30 features, 4 classes), and multi-class terrain state classification for legged robotics (200 features, 6 classes including deliberately confusable pairs). Post-training magnitude pruning and structured pruning were applied at six sparsity levels yielding parameter compression ratios (PCR) of 1.4x–9.5x, each condition replicated across five random seeds. Statistical analysis used paired t-tests (df = 4), one-way ANOVA, and 95% confidence intervals. We introduce two evaluation contributions: Youden's J (binary and macro-averaged one-vs-rest) as the primary metric in place of accuracy, and the Hardware-Task Feasibility Analysis (HTFA), a pre-deployment latency screening procedure. A pruning criterion ablation, comparing global magnitude, layer-wise magnitude, and random weight removal, isolates the mechanism underlying magnitude pruning's accuracy preservation. Four findings are reported. First, in clinical fall event detection, magnitude and structured pruning both preserved accuracy to 4.9x PCR with no statistically significant degradation (all p > 0.05 through 4.9x); degradation emerged only at extreme compression (9.5x: 98.7 ± 1.2%, p = 0.098). A class-imbalance ablation confirmed that sensitivity collapse at 9.5x is a sparsity threshold effect, Youden's J drops to 83.5–95.1% at 9.5x regardless of fall prevalence (10% or 18%), not a consequence of class imbalance. Second, in assistive prosthetic control, HTFA identified that Arduino Nano 33 BLE cannot meet the 10ms perceptibility deadline at any compression level because fixed processor overhead (11.0ms) exceeds the deadline independent of model size; projected latency ranges from 26.6ms (uncompressed) to 12.6ms (9.5x PCR), making hardware selection the required intervention. Third, the intelligent tutoring classification task proved highly compressible: magnitude pruning at 4.9x PCR produced non-significant accuracy degradation (1.7pp, p = 0.058), while terrain state classification for legged robotics degraded significantly at the same ratio (3.2pp, 95% CI [0.5, 5.9], p = 0.041). Structured pruning was markedly inferior to magnitude pruning in three of four domains. Fourth, the pruning criterion ablation demonstrated that random weight removal at 4.9x PCR reduces accuracy to 45.0 ± 13.5% versus 94.7 ± 2.3% for magnitude pruning, a 52.9 percentage point gap confirming that weight magnitude ordering, rather than parameter count reduction per se, is the active mechanism of accuracy preservation. Compression behaviour is domain-specific: the same PCR that is safe for intelligent tutoring classification (4.9x, non-significant degradation) is significantly harmful for terrain state classification in legged robotics, and no level of compression is sufficient for assistive prosthetic control on microcontroller hardware without a platform upgrade. Youden's J and HTFA provide practical, implementable alternatives to accuracy-only evaluation that capture these distinctions. The mechanistic ablation provides the strongest direct evidence to date that weight magnitude ordering is the operative principle in magnitude pruning. Replication on real sensor datasets (MobiAct, NinaPro DB1) and direct hardware latency measurement are essential next steps before these guidelines can be applied in production deployments.