NUHS is changing healthcare with AI and Spotfire: A behind-the-scenes tour with our CAO

I recently had the great pleasure of meeting with Dr. Kee Yuan Ngiam and the AI/LLM & Spotfire teams at the National University Health System in Singapore. They’re doing some super work with Spotfire to improve care and empower clinicians.

NUHS is ahead of the game when it comes to healthcare innovation. They’re putting data, AI, and predictive analytics to work in really cool ways. At the heart of their initiatives is the ENDEAVOR AI (EAI) platform—a core system that houses multiple AI tools and technologies to empower doctors to perform at the highest level, and improve the efficiency of care. These tools include enhancing diagnosis at admission, remote care management, predicting hospital stay duration, and optimizing discharge. This has led to significant improvements in patient care, productivity, cost savings, and optimized bed resources.

AI is changing the game in healthcare

NUHS has integrated Spotfire® software with their EPIC EHR system using TIBCO BW and HL7/XML protocols. Spotfire picks up data from resulting Kakfa topics for real-time visual data science prediction and at-rest database analytics. Patient and doctor notes are read in real-time, with EAI generating accurate predictions of patient length of stay (LoS) and forecasting future bed states and wait times while combining EMR data for things like medication use and pharmacogenomic risk.

NUHS has optimized cloud and on-premises environments to optimize the cost for computing and storage, with the LLMs running on-premises. NUHS launched its own secure healthcare generative AI model RUSSELL-GPT and developed specific prompts to train the AI model to answer questions with ~90 percent accuracy. Resulting discharges prior to 10 AM frees up beds and maximizes bed availability for incoming patients; it also helps doctors summarize patient notes and assess patients’ re-admission risk.

Dr. Ngaim and I compared notes regarding training and inferencing with LLMs. I showed him our latest Spotfire Copilot™ work, including our recent optimization of latency and cost using the Langsmith environment. He showed how to use the medical text summarization tool in NUHS’s intranet powered by RUSSELL-GPT to swiftly generate discharge notes for patients and answer follow-up questions. We will be putting our dev teams together to share these and other LLM learnings.

NUHS also uses advanced NLP methods to extract data and information from patient notes as input to predict hypoglycemia, high-risk pregnancy prediction, breast cancer, and pharmacogenomic risk with medication usage.

Spotfire plays a big part in this. As ENDEAVOUR AI pulls in live data throughout the health system, Spotfire reads this data and displays it in a command center format so that smarter decisions can be made in real time. EAI analyzes and displays the bed situation in near real-time across five institutions—National University Hospital (NUH), Ng Teng Fong General Hospital (NTFGH), Jurong Community Hospital (JCH), Alexandra Hospital (AH), and Jurong Medical Centre (JMC).

The Spotfire team is developing AI-based Spotfire® Accelerators as a way for our users to rapidly provide Spotfire solutions for high-value scenarios. One of the Spotfire Accelerators is for Hospital Management. This uses Spotfire to analyze historical patient data together with administrative data and train machine learning models to predict the LoS and re-admission risk for patients.

Event-driven AI at NUHS

NUHS is using event-driven AI to revolutionize the way hospitals operate, particularly in the area of predictive diagnosis at the point of care. TIBCO Integration technology enables real-time data ingestion from a number of source systems, especially the EPIC EMR. TIBCO Business Works invokes inbuilt HL7 and XML protocols to convert diverse upstream data formats into a JSON for downstream processing.

This platform manages multiple flows to handle different ADT schemas, ensuring that all necessary patient information is accurately extracted and integrated. TIBCO Integration pre-processes the data in-flight and combines this with historical data to provide a complete and current patient profile.

The result sets are analyzed using Spotfire visual data science technology, enabling physicians to make informed decisions swiftly. The visual data science analyses include the following areas:

Patient intake condition and medical history
Patient treatment data analysis and monitoring
Patient LoS prediction
Future bed states and wait times forecasting
EMR data analysis of medication usage and pharmacogenomics
Discharge data analysis and reporting
Readmission risk
Patient data analysis for case management and follow-up
Disease management and public health programs

Data scientists continuously refine these analyses and predictive models, ensuring their accuracy and efficacy. Behind the scenes, TIBCO Messaging ensures reliable data delivery at scale, streaming the data into TIBCO Streaming, where a set of transformations, aggregations, models, and rules are executed within the data streams. Inbuilt fan-out patterns allow data to be disseminated to multiple consumers simultaneously, ensuring that predictive models and analytics receive the necessary data in real time. This combination of messaging and real-time streaming data analysis ensures data integrity and analytics availability throughout the system. This enables the platform to run multiple large language models (LLMs) and natural language processing (NLP) models concurrently.

*NUHS Event-driven AI: High-level Architecture*

Note that TIBCO and Spotfire components are deployed in Kubernetes, offering benefits like scalability, resilience, and ease of management. This deployment strategy ensures the platform is agnostic, allowing it to be seamlessly moved to commercial clouds such as AWS and Azure without any changes. This flexibility ensures that healthcare providers can choose their preferred cloud environment while maintaining operational consistency and optimizing resource utilization.

A behind-the-scenes look at NUHS

Current and future work from NUHS also includes the management of robots for routine tasks like taking patients’ blood pressure and delivering medications and materials to their bedsides.

The nursing bot MISSI that can monitor and prevent patient falls.

While waiting to meet Dr. Ngiam, I noticed a group of folks using Spotfire. We got into a discussion about HbA1c diagnostics and safety data analysis.

Jingchuan Lin and I met with Jing Lin, the team’s manager, and discussed their use of IronPython scripts for controlling parameters in Python data functions for analytics inference.

Dr. Ngaim with Jing Lin, the team, and a Spotfire visual data science optimization of bed usage across five hospitals in Singapore.

NUHS in action at the Imagine AI Conference

Dr. Ngiam shared more about how NUHS uses Spotfire at the Imagine AI conference at Marina Bay Sands, December 5-6 2024.

Imagine AI brought together Healthcare and AI thought leaders from APAC and around the world. The keynote session was kicked off by Dr. Ngiam and Andrew Ng. Their conversation featured the rise of Agentic AI for addressing multi-faceted systems, where the agents work together on solutions comprising multiple workstreams. A panel with regional healthcare leaders from NUHS, Duke-NUHS, NHG, NUT, and SingHealth followed. Joint initiatives around foundation models and the use of AI to manage chronic diseases were discussed. This included the use of chatbots, risk identification across populations, and AI cancer screening.

Michael O’Connell with Dr. Ngiam, NUHS, at Imagine AI

Presenting on vision for Hospital Management

I presented our work on “Visual Data Science and Event-Driven AI” in the keynote session, combining data integration with visual data science for real-time insights and actions in healthcare operational systems. NUHS is using this combination to make a step change in the way hospitals operate at the point of care. For example, patient and doctor notes are analyzed with NLP algorithms and estimates of LoS; current and future bed state and wait times are updated in near real-time. In parallel, EMR data are updated with current medication to assess any pharmacogenomic risk. The operational data and analyses are updated using Spotfire visual data science technology, enabling physicians to make informed decisions at the point of care on a timely basis.

Along the way, I covered many of the healthcare applications outlined in this blog. I gave Spotfire demos on:

Hospital operations analyses with the Spotfire Copilot
Clinical safety data analyses
Real-time monitoring of vitals and diagnostics on patient populations

The solution I described includes real-time data ingestion and federation from the EPIC EMR and hospital operations systems using inbuilt HL7 and FHIR protocols to map diverse data into analysis-ready formats. Spotfire combines data from Kafka topics and at-rest systems in tandem, for visual data science analysis, prediction, and reporting.

Presentations by leading clinical, medical, and data science leaders followed the keynote session, including Gabriel Brat (Harvard Medical School), Hua Xu (Yale), Chuan Hong (Duke Medical School), Hua Xu (NUHS), Adam Dunn (Uni Sydney), Carl Moons (UMC Utrecht), and Asad Ali (Merck). Afternoon sessions with Cha Won Chui (Samsung Korea), Anirban Bhattacharyya (Mayo Clinic USA), and Justus Wolff (Syte Health, Germany). These sessions included data science and AI solutions for diagnostics, imaging, medical devices, drug discovery, genomics, clinical development, and digital health. Novel AI and data science systems including federated learning were discussed. Gabriel Brat described a Gen-AI system where the doctor acted as a copilot for the Gen-AI system vs the other way around.

Day 2 included speakers from NVIDIA, Intel, Huawei, HP, and PTC—featuring their latest work on Gen-AI, LLMs, foundation models, and agentic AI systems. Thought leaders from MIT, Emory, SInghealth, NUHS, Imperial College, Uni Birmingham, Duke-NUS, NUS, and NUHS wrapped up the conference. The technologies supported by the hardware and software systems were novel and broad based.

Our Spotfire presence at Imagine AI included our local Spotfire visual data science team and our GTM partner Delteq. Delteq creates solutions combining TIBCO and Spotfire products for applications in Life Sciences, Finance, Telco, and other industries.

I also met with Andy Ta (ADO and Director AI) and a team from Synapxe, who design and run software systems for the 70 or so healthcare centers across Singapore. It was fun to see live interactions with the Spotfire Copilot running on healthcare data. Singapore healthcare is highly innovative, and I can’t wait to see how AI continues to evolve at NUHS and the Singapore healthcare system.

NUHS (Lin Jing on left), Spotfire, and Delteq at Imagine AI

References

Acknowledgments

It was terrific to work with Jingchuan Lin on the Spotfire Customer Success team to visit NUHS. Thanks, Jingchuan for helping with this blog. And thanks to the NUHS team including Dr. Ngaim, Dr. James, and Lin Jing for the super collaboration.