Electronic medical records (EMRs) supports the development of machine learning algorithms for predicting disease incidence, patient response to treatment, and other healthcare events. But insofar most algorithms have been centralized, taking little account of the decentralized, non-identically independently distributed (non-IID), and privacy-sensitive characteristics of EMRs that can complicate data collection, sharing and learning. To address this challenge, we introduced a community-based federated machine learning (CBFL) algorithm and evaluated it on non-IID ICU EMRs. Our algorithm clustered the distributed data into clinically meaningful communities that captured similar diagnoses and geological locations, and learnt one model for each community. Throughout the learning process, the data was kept local on hospitals, while locally-computed results were aggregated on a server. Evaluation results show that CBFL outperformed the baseline FL algorithm in terms of Area Under the Receiver Operating Characteristic Curve (ROC AUC), Area Under the Precision-Recall Curve (PR AUC), and communication cost between hospitals and the server. Furthermore, communities' performance difference could be explained by how dissimilar one community was to others.

BACKGROUND: A number of recent studies have reported an association between intraoperative burst suppression and postoperative delirium. These studies suggest that anesthesia-induced burst suppression may be an indicator of underlying brain vulnerability. A prominent feature of electroencephalogram (EEG) under propofol and sevoflurane anesthesia is the frontal alpha oscillation. This frontal alpha oscillation is known to decline significantly during aging and is generated by prefrontal brain regions that are particularly prone to age-related neurodegeneration. Given that burst suppression and frontal alpha oscillations are both associated with brain vulnerability, we hypothesized that anesthesia-induced frontal alpha power could also be associated with burst suppression. METHODS: We analyzed EEG data from a previously reported cohort in which 155 patients received propofol (n = 60) or sevoflurane (n = 95) as the primary anesthetic. We computed the EEG spectrum during stable anesthetic maintenance and identified whether or not burst suppression occurred during the anesthetic. We characterized the relationship between burst suppression and alpha power using logistic regression. We proposed 5 different models consisting of different combinations of potential contributing factors associated with burst suppression: (1) a Base Model consisting of alpha power; (2) an Extended Mechanistic Model consisting of alpha power, age, and drug dosing information; (3) a Clinical Confounding Factors Model consisting of alpha power, hypotension, and other confounds; (4) a Simplified Model consisting only of alpha power and propofol bolus administration; and (5) a Full Model consisting of all of these variables to control for as much confounding as possible. RESULTS: All models show a consistent significant association between alpha power and burst suppression while adjusting for different sets of covariates, all with consistent effect size estimates. Using the Simplified Model, we found that for each decibel decrease in alpha power, the odds of experiencing burst suppression increased by 1.33-fold. CONCLUSIONS: In this study, we show how a decrease in anesthesia-induced frontal alpha power is associated with an increased propensity for burst suppression, in a manner that captures individualized information above and beyond a patient’s chronological age. Lower frontal alpha band power is strongly associated with higher propensity for burst suppression and, therefore, potentially higher risk of postoperative neurocognitive disorders. We hypothesize that low frontal alpha power and increased propensity for burst suppression together characterize a “vulnerable brain” phenotype under anesthesia that could be mechanistically linked to brain metabolism, cognition, and brain aging.