[Paper Review] SAIA: Split Artificial Intelligence Architecture for Mobile Healthcare System

Di Zhuang, Nam V. Nguyen+2 more|arXiv (Cornell University)|Apr 25, 2020

IoT and Edge/Fog Computing64 references4 citations

TL;DR

SAIA proposes a split artificial intelligence architecture for mobile healthcare systems that dynamically routes data between lightweight embedded AI on devices and powerful networked AI in the cloud based on real-time conditions. By using a meta-information-based decision unit with tunable hyperparameter ε, SAIA achieves superior effectiveness and efficiency over baselines on skin lesion and nail fungus datasets, maintaining performance even during communication disruptions.

ABSTRACT

As the advancement of deep learning (DL), the Internet of Things and cloud computing techniques for biomedical and healthcare problems, mobile healthcare systems have received unprecedented attention. Since DL techniques usually require enormous amount of computation, most of them cannot be directly deployed on the resource-constrained mobile and IoT devices. Hence, most of the mobile healthcare systems leverage the cloud computing infrastructure, where the data collected by the mobile and IoT devices would be transmitted to the cloud computing platforms for analysis. However, in the contested environments, relying on the cloud might not be practical at all times. For instance, the satellite communication might be denied or disrupted. We propose SAIA, a Split Artificial Intelligence Architecture for mobile healthcare systems. Unlike traditional approaches for artificial intelligence (AI) which solely exploits the computational power of the cloud server, SAIA could not only relies on the cloud computing infrastructure while the wireless communication is available, but also utilizes the lightweight AI solutions that work locally on the client side, hence, it can work even when the communication is impeded. In SAIA, we propose a meta-information based decision unit, that could tune whether a sample captured by the client should be operated by the embedded AI (i.e., keeping on the client) or the networked AI (i.e., sending to the server), under different conditions. In our experimental evaluation, extensive experiments have been conducted on two popular healthcare datasets. Our results show that SAIA consistently outperforms its baselines in terms of both effectiveness and efficiency.

Motivation & Objective

• To address the challenge of deploying deep learning in mobile and IoT healthcare devices with limited computational resources and unreliable connectivity.
• To design a hybrid AI system that leverages both local embedded AI and remote networked AI, ensuring functionality during communication outages.
• To develop an intelligent decision unit that adaptively selects between client-side and server-side AI processing based on data characteristics and system conditions.
• To optimize the trade-off between accuracy, latency, energy consumption, and communication overhead in mobile healthcare applications.
• To evaluate the architecture on real-world healthcare datasets, demonstrating robustness and superiority over existing approaches.

Proposed method

• The system employs a four-component architecture: data pre-processing (object detection, semantic segmentation, feature extraction), embedded AI (lightweight DNN classifiers), networked AI (fusion of multiple advanced DNNs), and a core decision unit.
• The decision unit is a lightweight machine learning classifier trained on meta-features (e.g., image quality, complexity, and contextual cues) to predict whether to process a sample locally or remotely.
• The decision unit uses a tunable hyperparameter ε to balance confidence thresholds, enabling dynamic adaptation to varying network conditions and resource availability.
• Meta-features include image resolution, texture complexity, and lesion characteristics, which are extracted before classification to inform routing decisions.
• The embedded AI uses compact models like MobileNet or EfficientNet, while the networked AI employs ensemble models for state-of-the-art accuracy.
• The system supports both cloud-dependent and offline operation, ensuring resilience in contested environments such as satellite link disruptions.

Experimental results

Research questions

• RQ1 How can a mobile healthcare system maintain high diagnostic accuracy when cloud connectivity is unreliable or unavailable?
• RQ2 What criteria should guide the routing of a medical image sample between local embedded AI and remote networked AI?
• RQ3 Can a lightweight meta-information-based decision unit effectively balance accuracy, latency, and communication cost in real-time mobile healthcare applications?
• RQ4 How does the proposed SAIA architecture compare to traditional cloud-only or client-only AI approaches in terms of performance and efficiency?
• RQ5 To what extent can the hyperparameter ε in the decision unit be tuned to optimize system behavior under varying network and device conditions?

Key findings

• SAIA consistently outperforms baseline systems in both effectiveness and efficiency across two large-scale healthcare datasets: skin lesion and onychomycosis (nail fungus) classification.
• The meta-information-based decision unit with hyperparameter ε enables adaptive routing, reducing unnecessary data transmission without sacrificing diagnostic accuracy.
• The system maintains high performance even during simulated communication outages by relying on the embedded AI, demonstrating resilience in contested environments.
• The use of a lightweight decision unit significantly reduces computational overhead on the client side while enabling intelligent workload distribution.
• The fusion of multiple advanced DNNs on the server side achieves state-of-the-art accuracy, which is not possible with standalone lightweight models on the device.
• Experimental results show that SAIA reduces communication load and energy consumption compared to traditional cloud-only or split-DNN approaches, especially under bandwidth-constrained conditions.