Improving Patient Outcomes With Accelerated Graph Databases

The volume of data that healthcare providers collect is increasing and overwhelming the approaches currently being used to analyze it. At the same time, the ability to connect across tables and business entities and identify hidden relationships and patterns offers tantalizing breakthroughs in patient care and outcome improvements as well as efficiencies and cost savings for the providers.

Every day, more and more enterprises are using graph databases to explore and analyze their connected data. Graph databases deliver deeper analytics on connected data, treating connections and relationships between data as first class citizens, while relational databases and other NoSQL databases are blind to connected data. Awareness, interest, evaluation and adoption of graph databases continues to outpace other database technologies, and Gartner forecasts accelerated growth for graph: “By 2025, graph technologies will be used in 80% of data and analytics innovations, up from 10% in 2021, facilitating rapid decision making across the enterprise.” 1

World’s Largest Healthcare Company and Graph Databases

Tigergraph is the world’s fastest and most scalable graph platform. In this blog post, we discuss how Tigergraph uses Xilinx hardware acceleration and graph analytics libraries to help the world’s largest healthcare company deliver integrate data from over 200 sources to create a full longitudinal health history of every member in order to streamline their call center, and provide real-time care path recommendations for their members.

This customer provides health care services, health benefit plans, and insurance and financial services. The organization currently runs the largest connected healthcare graph database asset in the United States, with over 10B+ vertices and 50B+ edges and contains data from over 100M members and houses 18 months of current data from members, claims, clinical interactions, providers, phone calls, house calls and more. This database contains over 1.2TB of data and supports over 33,000 online users on various applications.

The customer deployed TigerGraph’s Patient360 solution to provide validated care path recommendations quickly and efficiently as possible via call center agents. By delivering better and more efficient guidance in real-time, their goal was for a 20 minute call to be 10% shorter leading to better customer satisfaction as well as $100M in call center savings.

One of the queries was to calculate patient similarity. This use case also is known as a medical twin. A doctor who is treating a patient wants to analyze other patients with characteristics like their own to determine a treatment regimen. To do that the doctor is looking for the most similar patients that match their patient’s medical history and who had successful outcomes. Let’s dive into how this works.

The Importance of Cosine Similarity

Similarity calculations are critical for all types of recommendation engines: consider the utility (and delight) in YouTube recommending a video you’d never seen from your favorite musical artist, or Netflix recommending a movie that becomes your new favorite, or Amazon recommending a product that magically fits both your taste and the very next stage of your home remodel. TigerGraph uses a cosine similarity algorithm to deliver the same sort of magical results to their Member Journey solution for their customer.

Let’s briefly sketch out what cosine similarity is and how it’s an algorithm that’s of great use in patient care path recommendations. Property-based similarity is used to find the most similar items in a graph by comparing properties and structure, and these features of an item are represented as numbers that can be stored in a matrix as an array. Allergies, procedures, immunizations, and conditions are all examples of the types of properties that would be in the mix for this use case. From the counts and weighting of the numbers in that array a vector is created. After that, the target vector (the one just created) is compared that against the population vectors to find the closest matches.

Tigergraph blog vector graphic-01

So why “cosine” similarity? As the angle between two vectors diminishes, the cosine of that angle approaches 1. When the angle between two vectors is zero, the cosine is 1 (cos(0)=1). On the other hand, when the vectors are in orthogonal directions, the cosine is 0 (cos(90)=0). The closer the cosine gets to 1, the more similar the two patient histories are!

Accelerating Care Paths

Clearly, the ability to search through millions of patient records across hundreds of properties, return the most similar records, and derive a care path is represents huge benefits for patient health. The ability to do it fast adds in the benefits of patient peace-of-mind and patient satisfaction as well as saving a lot of money by increased call center efficiency. The more patients you have, the harder it is to do these things in real time. When you have 100 million patients, it’s extremely difficult to return results in time for those results to fit into the natural flow of a human conversation. Results in a few minutes- which TigerGraph is able to achieve in a conventional CPU-based computing architecture- is still objectively impressive, but it fails to meet the needs of this use case.

That’s where Xilinx comes into the picture. Xilinx is the world leader in FPGAs- semiconductors based around a matrix of configurable logic blocks. Among other advantages, FPGAs are massively parallel, meaning that they can perform several computations at the same time. This makes FPGAs ideal for accelerating computationally intensive workloads. Xilinx Alveo Accelerator cards are standard PCIe devices that make FPGA co-processing easy to deploy into industry-standard servers, and Alveo U50 cards were used in this use case. Xilinx Vitis libraries make Alveo acceleration easily and flexibly accessible to applications like TigerGraph using common high-level language

That’s where Xilinx comes into the picture. Xilinx is the world leader and inventor of FPGAs. The architecture of the Xilinx FPGA allows to adapt "Custom Fit" to match the unique needs of each computationally intensive workload like Cosine Similarity used in a patient recommendation engine.

The Xilinx Alveo U50 is a PCIe-based FPGA accelerator card that can be deployed into industry-standard servers. The card features massively parallel FPGA processing horsepower to compute the cosine similarity algorithm with fast access to High Bandwidth Memory (HBM2) that stores the patients' records locally for fast processing. This helps achieve greater than 300x performance improvement “speed up” to complete a query search vs a CPU-based implementation. 

The result of moving patient similarity queries from a CPU-based architecture to Xilinx Alveo cards was transformative. Query response times dropped from one minute to 50 milliseconds. This enabled the provider to meet their objective goals for call times and savings, as well as creating more personal and natural interactions between the patient and representative with results available to the representative within the cadence of a normal human conversation.


For more information on Xilinx, TigerGraph, and this particular solution, visit

Also, join TigerGraph in a live webinar to learn how Xilinx turbocharges Customer Journey/360 performance for healthcare organizations. Register now


1“Top Trends in Data and Analytics for 2021” by Rita Sallam, Distinguished VP and Fellow, Gartner