Medical AI has demonstrated specialist-level diagnostic accuracy, yet these capabilities remain largely inaccessible in resource-constrained rural settings where bandwidth is scarce, compute is limited, and clinical decision-making requires integrating heterogeneous modalities. We introduce a cloud--edge collaborative architecture that addresses these constraints: lightweight, domain-specific mode
A cloud–edge collaborative architecture addresses the inaccessibility of specialist-level medical AI in rural settings by deploying lightweight, domain-specific models on local edge hardware to transform raw medical data into compact structured outputs. A cloud-based Large Language Model (LLM) then synthesizes these outputs into clinical summaries, enabling multimodal integration without transmitting bandwidth-intensive raw data.
This hybrid system achieves 95–99% diagnostic tool recall and 0.87–0.90 oracle accuracy while transmitting only ~6.5 KB of structured evidence to the cloud, which is three orders of magnitude less than cloud-only baselines. It maintains bandwidth-invariant latency (25–38 seconds) across varying network conditions and demonstrates superior factual grounding compared to cloud-only approaches, which often suffer from hallucinations and high latency due to raw data transmission.
The architecture ensures privacy and reliability by keeping raw images and signals local, allowing for graceful degradation if cloud connectivity fails. By decoupling perception (edge) from reasoning (cloud), the system provides a scalable solution for resource-constrained environments where traditional cloud-based AI fails due to scarce bandwidth, limited compute, and unreliable internet connectivity.
This paper targets a central deployment bottleneck in medical AI: high-accuracy diagnostic models are often impractical in resource-constrained rural settings because they require reliable bandwidth, substantial compute, and integration across heterogeneous clinical data. The authors propose a cloud–edge collaborative architecture for multimodal clinical screening that partitions the inference workload between local edge devices and a central cloud. In this design, lightweight, domain-specific models run on the edge to handle latency-sensitive and bandwidth-sensitive processing, while the cloud can provide heavier computation, aggregation, or refinement where connectivity permits. The system is explicitly motivated by clinical decision-making in settings where raw multimodal data may be costly, slow, or infeasible to transmit.
The key contribution is an architectural approach to making specialist-level screening more deployable in low-resource environments. Rather than assuming a single monolithic model with unrestricted data access, the work emphasizes practical constraints such as intermittent connectivity, limited on-device compute, privacy-sensitive clinical data, and the need to fuse multiple modalities into a usable screening signal. This matters because rural clinics often lack both specialist access and the infrastructure needed to run state-of-the-art medical AI; a cloud–edge design could improve triage, support local clinicians, and enable scalable screening without requiring hospital-grade resources.