Using Kafka-Powered AI Models to Predict and Prevent Sepsis at City of Hope

This post details how City of Hope implemented a real-time data streaming infrastructure using Apache Kafka to power an AI model for predicting and preventing sepsis in bone marrow transplant patients. The architecture involves ingesting HL7 messages from Epic via Corepoint, converting them to JSON, and streaming them into Kafka topics (vitals, lab results, medication orders, ADT, substance administration). Microservices process this data, trigger the predictive model, and feed predictions back into Epic to alert clinicians. Confluent Cloud on Azure is used for scalability and reliability, with ksqlDB used for data inspection. The system leverages Kafka's immutable log for state retrieval if the model crashes.