Predictive analytics helps Fresenius Medical Care anticipate dialysis complications

The solution needed to scale to all of Fresenius’s dialysis centers, with each location sending 10MBps of treatment data at peak times. A low-latency, time-sensitive solution of 10 seconds from data origination from dialysis machines and medical sensors to reporting and notification was critical.

In addition, systematic and automated monitoring and alerting mechanisms were necessary to help the team spot problems and resolve them quickly. The solution uses CloudWatch alerts to send notifications to the DataOps team when there are failures or errors, while Kinesis Data Analytics and Kinesis Data Streams are used to generate data quality alerts.

“Using an agile approach, we prioritized features to deliver a minimal viable prototype over a six-month period,” Waguespack says. “Our primary challenge was in our ability to scale the real-time data engineering, inferences, and real-time monitoring to meet service-level agreements during peak loads (6K messages per second, 19MBps with 60K concurrent lambda invocations per second) and throughout the day (processing more than 500 million messages daily, 24/7).”

Fresenius’s machine learning model uses electronic health records comprising intradialytic blood pressure measurements and multiple treatment- and patient-level variables. The team trained and validated the model using observational data from 42,656 hemodialysis sessions in 693 in-center hemodialysis patients. In the training cohort, the model was optimized to generate an IDH alert between 15 and 75 minutes before an IDH event.

Transforming dialysis

Waguespack says the project was new ground for Fresenius, requiring the organization to explore measures to protect health information in the cloud and the role AI can play in a clinical setting. Each of those were associated with blockers, real and perceived.

“It was imperative for us to gain full partnership from all our stakeholders by creating absolute alignment focused on quality improvement, complete transparency in our work, and showing the utmost integrity by living up to our own expectations,” he says.

He notes that success required the IT organization to become more agile, embrace failing fast, and internalize that learning is a deliverable as valuable as adding a new feature.

“This shift in attitude and expectations needed to come top down and bottom up,” he says. “Top down, to provide the support and space to change. Bottom up, from those experienced in an agile approach and able to model behavior day in and day out. This shift could be seen in the words we used, the way we celebrated learning and progress, and the respectful and supportive nature of the team.”

The IDH tool has not yet been evaluated or cleared for use by the US Food and Drug Administration (FDA), but Zhang says the team recently published its findings in a top peer-reviewed kidney journal. While further clinical studies are required to validate whether prediction of IDH followed by timely, appropriate preventative measures translates into lower IDH rates and improved patient outcomes, Zhang says the model’s high performance in the validation cohort is promising. Waguespack adds that the project has been another step in Fresenius Medical Care’s ongoing digital transformation.

“The opportunity to predict IDH during a dialysis treatment is one of several building blocks to transform our company into the world of the Internet of Things, big data, and artificial intelligence,” he says. “Building on the success of this initiative, we continued our journey to collect terabytes of data from novel sources in a modern data platform. From here, we continue to iterate on the process and technology to effectively manage our data so that it can enable continued innovation, including machine learning for image classification apps, genomic research, large language models, and beyond.”