AI Driven Microbiome Diagnostics: Edge Cloud Architecture and Federated Learning for Infection Risk Stratification

Abstract

Microbiome diagnostics Ai based diagnostics have been investigated to aid infection risk stratification through learning behaviors on microbial community profiles. Precision microbiology has demanded quick results, dependable calibration and privacy conscious learning within the laboratories. To achieve low latency inference at sample processing points, edge cloud architecture has been taken with centralized training, monitoring, and model governance being kept in the cloud. In this research study, a new computational assessment has been done in relation to a synthetic microbiome relative abundance dataset that has been created in regards to a Dirichlet based community model to simulate compositional microbial profiles. The risk labels of infection have been created based on community dependent prevalence and dysbiosis such as the perturbations of selected groups of taxa. The training of a clear baseline has been done by logistic regression and a simulation of federated learning has been conducted by using FedAvg style aggregation approach on 6 clients. The test set has been assessed on performance, and edge style single sample inference and cloud style batch inference and inference latency have been measured, assuming explicit network overhead to model deployment. Findings have indicated heavy discrimination of centralised learning of ROC AUC of about 0.97 and accuracy of over 0.93. It has been demonstrated that federated learning can converge to similar ROC AUC, but with a small decrease in accuracy. The result of the latency has indicated that the network overhead has dominated the total remote inference time even when the efficient compute is achieved when batches are taken. Clinical microbiology workflow implications have been outlined such as privacy preserving collaboration, model monitoring and deploying in small connectivity laboratories.