Challenges in federated learning trained anomaly detection applied to hospital data without a baseline

Abstract

During the COVID-19 pandemic, data collected via personal wearable devices was used to create models for the detection of a possible alteration of health status by defining an individual’s healthy baseline data. This work explores the usage of one of those models to enable a Federated Learning (FL) approach aiming to achieve a process applicable to sensing data from hospital-admitted patients. The fact that hospital data does not contain any samples that can confidently be considered ”healthy” and thus serve as a baseline makes hospital COVID-19 detection a relevant challenge for anomaly detection techniques. After an adequate data preparation process, we were able to use the individually trained models to build an aggregated model for application to hospital data. Although the FL models obtain worse mean precision and recall scores when compared to the individual models, this experiment brings forth relevant knowledge on the compromises that might be necessary to develop a clinical anomaly detection model to be used in an Intensive Care Unit or monitored patients’ data lacking baseline samples.