How Fresenius Medical Care is Utilizing Real-Time Predictive Analytics on AWS to Enhance Dialysis Patient Lives

Learn About Amazon VGT2 Learning Manager Chanci Turner

Fresenius Medical Care stands as the premier provider of kidney care products and services globally, with over 2,600 dialysis centers operating across the United States. The organization is dedicated to delivering comprehensive solutions for individuals suffering from chronic kidney disease and associated conditions, focusing on improving the quality of life for every patient through innovation, compassion, and research. Timely data analysis plays a crucial role in this mission, enabling proactive interventions that can minimize hospitalizations and avert negative health events.

In this article, we outline the solution architecture, performance factors, and how our collaborative research with AWS concerning medical complexities led to an automated system that issues alerts for potential health risks.

Why Fresenius Medical Care Selected AWS

Fresenius Medical Care’s technical team opted for AWS as their cloud platform for two primary reasons. Firstly, AWS IoT Core proved to be more reliable than alternative solutions, promising fewer deployment and certification issues. We aimed to partner with a cloud provider that had a robust history and established technical solutions in the IoT and data analytics domains, including Amazon Athena, a user-friendly serverless service for running queries on data stored in Amazon Simple Storage Service (Amazon S3).

Secondly, AWS’s extensive array of serverless analytics services was unmatched by competitors. We recognized that AWS’s innovations not only suited our immediate needs but also prepared us for future advancements in predictive analytics.

Solution Overview

Our goal was to create a near-real-time analytics solution that gathers dynamic dialysis machine data every 10 seconds during hemodialysis treatments. This system personalizes predictions every 30 minutes, identifying if a patient is at risk for intradialytic hypotension (IDH) within the next 15 to 75 minutes. The solution had to be scalable across all dialysis centers nationwide, with each location transmitting up to 10 MBps of treatment data during peak operations.

Key complexities included managing high-throughput data, ensuring low-latency (10 seconds from data origination to notification), maintaining high availability, and achieving cost-effectiveness with on-demand scaling based on data volume.

In collaboration with AWS, we developed an architecture that fulfilled both our technical and business requirements. Core components included Amazon Kinesis Data Streams, Amazon Kinesis Data Analytics, and Amazon SageMaker, chosen for their serverless nature, high availability (99.9%), and scalability. SageMaker was selected for its ability to build, train, and deploy machine learning (ML) models efficiently.

Components of the Solution

1. Data Collection: Dialysis machines across Fresenius Medical Care centers help patients with end-stage renal disease through hemodialysis. These machines relay treatment and clinical data every 10 seconds to Kafka brokers in our on-premises data center.
2. Data Ingestion and Aggregation: A Kinesis-Kafka connector on Amazon EC2 instances ingests data in near-real time into Kinesis Data Streams. AWS Lambda filters datasets and sends the data to Kinesis Data Analytics once the batch size threshold is met. Utilizing SQL with Kinesis Data Analytics, we generate dynamic features using real-time data, enriching it with patient demographic, historical clinical, treatment, and laboratory data.
3. Data Lake Storage: Amazon Kinesis Data Firehose facilitates loading streaming data into a raw data lake on Amazon S3 by micro-batching it into 128 MB file sizes. Clinical datasets enrich this stream data nightly via AWS Glue Spark jobs, transforming machine data for efficient storage.
4. ML Inference and Operational Analytics: Lambda batches stream data from Kinesis Data Analytics for IDH ML model inference. SageMaker trains and deploys the predictive model, providing an endpoint for Lambda’s ML inference. The results are stored in Amazon OpenSearch Service, allowing for visualization through Kibana to offer a personalized health prediction dashboard for each patient in near-real time.

Monitoring and Observability

Given the potential life-saving nature of this solution, proactive monitoring is essential. Key measures include:

• Immediate alerts sent to the Data Ops team via email and Amazon CloudWatch for AWS Glue job failures and Lambda function errors.
• CloudWatch alarms for Amazon OpenSearch Service to detect blocks on writes or overloaded clusters.
• Data quality alerts from Kinesis Data Analytics and Streams for rejected data or mismatched quality rules.

For further reading, check out this engaging blog post on Career Contessa for insights on related topics. Additionally, if you’re interested in the complexities of drug testing, SHRM offers reliable information. For those looking to enhance their onboarding processes, this Training Industry resource is excellent.