Ai Driven Healthcare Systems: Edge Cloud Architecture for Precision Medicine A Computational Research Study with Original Analysis

Abstract

The ai models have allowed the combination of clinical history, imaging, laboratory values, and continuous sensor signals to enable precision medicine. Latency has limited deployment, privacy has been an issue, and the operational cost of storing models in various care environments. There is adoption of edge cloud architecture to be able to distribute computation such that time critical inference has been performed near the patient with population scale training and governance performed in the cloud. This research study has carried out an original computational analysis in order to measure the trade off between edge and cloud inference latency and to determine the privacy preserving training using federated learning. An openly available proxy of the precision risk stratification tasks has been made on the Breast Cancer Wisconsin Diagnostic dataset. A central logistic regression baseline has been developed and assessed on a held out test set. An implementation of federated learning using FedAvg style of aggregation has been simulated on 5 clients and convergence behavior across 25 rounds has been measured. Single sample batch inference has been used to measure edge inference latency and cloud inference compute latency has been used to measure batch inference and then added to an explicit assumed network round trip overhead. Findings have demonstrated high predictive performance of centralized learning that is highly discriminative with low error of calibration. It has been demonstrated that Federated learning reaches similar discrimination and accuracy drops slightly with the simulated client split. The results of latency have revealed that compute time per sample has been lower when cloud batching has been used, whereas total per sample latency has been dominated by network overhead when remote inference has been assumed. The implication of precision medicine on systems have been brought up, and an implementative evaluation methodology has been presented to be deployed in future studies. It has utilized a publicly available benchmark clinical proxy dataset and simulated a cross silo federated learning environment to measure convergence and performance.