InterSystems IRIS – Letting the light in

InterSystems is a private company, founded by Phillip Ragon in 1978 and still wholly owned by him. It is headquartered in Cambridge Massachusetts, with 39 offices globally and 1,950 employees. The company has just over a billion dollars in revenue, more than a thousand enterprise customers for its IRIS software platform and 800 partners. The company provides a high-performance data platform that is targeted at application developers, and applications and application frameworks that all leverage the core capabilities of the data platform targeted to healthcare, financial services and supply chain industry challenges. Historically the majority of its customers have been application software providers, however the company has been successful expanding its presence with end user organisations. InterSystems software manages 1 billion patient records globally, processes 2 billion equity trades daily and tracks 20 million shipping containers in real-time. InterSystems claim a 99.9% customer renewal rate with 99.8% of customers “completely satisfied”. These high satisfaction levels seem to be consistent with high ratings on publicly available rating services like TrustRadius. InterSystems has its own start-up incubator and has made 16 investments in various start-ups so far.

At the heart of the IRIS data platform is a core database engine that has a key-value structure but with “projections” above this that allow it to deal with SQL, documents, images, JSON files and other formats without any data duplication. The core data platform software has in-database support for processing analytics, vector search, time series and more, such as Python processing. There is a suite of connectors to allow it to ingest data from a wide range of database and document sources. The IRIS platform can be deployed either in a public cloud like AWS or a private cloud, a managed service or on-premises, hybrid and multi-cloud, and supports microservices and containers via Kubernetes. The software can run on a little Raspberry Pi processor right up to huge SMP servers.

IRIS can be viewed as a smart data fabric, exposed through data services. Data fabric is a modern data architecture that spans data silos, aimed at minimising the need to physically move large amounts of data around in order to satisfy user requirements. In IRIS you can either virtualise data or physically integrate it, as the use case demands. Data Fabric Studio is a tool to help users bring data together from many sources without the need for programming skills. Another widely adopted IRIS capability, that is also available as a dedicated SaaS offering, allows healthcare standard formats to be converted from one format to another. IRIS provides the ability to develop and operationalise machine learning models directly within the platform using simple SQL-like syntax rather than being restricted to the domain of highly skilled developers or data scientists.

At the core of IRIS is a data engine, a key-value store that is persistent and multi-threaded. On top of this you can have various data models including relational, objects, documents, time series, vectors and more. IRIS supports both row and columnar structures, as well as documents and bitmaps (used for certain indexes). Projections are used to map between these various data models and their ideal storage form. For example, there may be sensor data that is projected as relational data. The object-relational mapping is useful for complex healthcare data like the FHIR standard (Fast Healthcare Interoperability Resources). There is optimised connectivity for JDBC and Kafka and a deep level of support for Python as a native language, which allows easy deployment of generative AI. Integrated support for vector representations and vector math means that it natively supports GenAI use cases, as well as retrieval-augmented generation using any LLM without requiring a separate vector database. IRIS has adapters for LLAMA and LangChain and more, in terms of its broad AI support.

A practical example of IRIS deployment is medical data, which is one of their biggest use cases. The “Epic Systems” healthcare EMR software, based on IRIS, is deployed across thousands of US hospitals, managing 325 million patient records. This is also deployed at some UK hospitals like St Ormond Street and University College London. This software connects patient and other healthcare data for insurers, pharmacies, healthcare providers and patients. One statewide deployment in the US manages 300 million records from 8,000 healthcare facilities and 18,000 healthcare providers. Another example is the life sciences company Biostrand in Belgium, which uses vector search for protein sequencing analytics.

IRIS’ native vector search makes it well suited to retrieval augmented generation (RAG) with large language models (LLMs) where the knowledge gleaned from training data a mainstream LLM was trained on is supplemented with organisation-specific data, such as hospital records or customer service history, that is stored on IRIS. RAG can reduce (though never eliminate) LLM hallucinations if it is based on high-quality data, and can improve the quality of LLM responses, in conjunction with careful prompt engineering.

Three of the five largest US healthcare insurers use InterSystems, and two-thirds of UK health trusts also use them. They also have partners selling laboratory systems and other solutions based on IRIS. TrakCare is a healthcare system that is fully owned by InterSystems, deployed in 29 countries. This covers patient care, diagnostic services and more. One use case is at the Saudi Arabia King Khaled Eye Specialist Hospital, where it is deployed to try and predict patients’ likelihood of not keeping appointments. By prioritising repeated reminders directed at those problem cases, this approach resulted in a 40% reduction in no-shows.

One interesting case study is Tricore Labs, who used the FHIR standard to integrate data to find 34,000 previously undiagnosed Hepatis C patients. In another case, an AI is used to scan a large number of patient messages on a heavily used patient portal to highlight urgent ones like those with chest pain and flag them for clinical attention.

One AI use case is in listening to audio of patient interviews with physicians. The AI analyses the audio and does not just transcribe the session but comes up with suggestions. For example, the AI summarises patient conditions as stated during the interview and lists suspected conditions, as well as suggested actions like ordering further specific tests. An intriguing aspect was its explainability. For example, the AI could justify its suggestions by bringing up specific parts of the audio when asked e.g. when a patient said they were a heavy smoker the AI highlighted this clip as a reason for a further specific further test.

InterSystems is an interesting company that has built a billion-dollar revenue business primarily through marketing itself to application solution builders, although more recently also directly to businesses. Having a foundation of a non-relational database would have been a very difficult sell to large enterprises in the 1990s, and a hurdle for several data platform companies that tried to compete directly against the relational juggernaut (witness the fate of Software AG’s Adabas). However, by being embedded within broader vertical solutions, this problem was largely avoided by InterSystems. Ironically, in more recent years there has been an explosion of different database types, from graph databases to vector databases and assorted NoSQL tools. The IRIS platform is now the database that can speak its name in a more diverse data platform world.