Causal Modeling of Inference Latency with Model Compression Strategies in Mobile Health Applications

Abstract

The integration of deep learning models into mobile health applications has revolutionized continuous monitoring and point-of-care diagnostics. However, deploying complex neural networks on resource-constrained mobile edge devices introduces significant inference latency, which can be detrimental in time-critical medical scenarios. While model compression strategies such as quantization, pruning, and knowledge distillation are routinely employed to mitigate latency, standard benchmarking methodologies often fail to capture the true latency reduction due to environmental confounders like thermal throttling, battery degradation, and dynamic background operating system workloads. This paper introduces a comprehensive causal modeling framework to accurately assess the inference latency of various model compression strategies in mobile health applications. By conceptualizing model compression as a causal treatment and inference latency as the outcome, we construct a structural causal model to identify and adjust for systemic confounders using inverse probability weighting and propensity score matching. Our methodology is validated through extensive experiments on commercial mobile hardware using medical datasets, including electrocardiogram classification and dermatological image analysis. The causal analysis reveals that raw observational data significantly overestimates the latency benefits of unstructured pruning while underestimating the efficacy of integer quantization due to hidden thermal interactions. These findings provide a robust paradigm for mobile health developers to evaluate compression trade-offs, ensuring reliable and rapid artificial intelligence deployment in critical healthcare environments.