Aidbox FHIR Platform for the development of Immunix for precision medicine
Introduction
Sepsis is a life-threatening condition characterized by an overactive immune response to infection. The WHO estimates that it accounts for 1 in 5 deaths worldwide. In the United States alone, sepsis treatment costs nearly $27 billion annually, making it one of the largest cost drivers for U.S. hospitals.
Sepsis is difficult for providers to detect due to the varied presentation in patients. However, new technologies can enhance early prediction. Machine learning (ML) models are particularly effective when rapid assessment of multiple parameters is required. Prenosis developed an AI/ML based Software as a Medical Device for the diagnosis and prediction of sepsis that analyzes over 20 biological parameters.
About Prenosis
Prenosis Inc. is an artificial intelligence company enabling precision medicine in acute care. Prenosis spent nearly a decade building a biological and clinical dataset, which is recognized as the largest ever generated for acute patients suspected of serious infection. Leveraging this comprehensive dataset, Prenosis developed the Sepsis ImmunoScore -- the first FDA-authorized AI/ML tool for sepsis risk diagnosis and prediction. It integrates with existing EHR systems and provides ML-based insights to help providers diagnose sepsis faster and initiate treatment sooner.
Challenge
In 2022, Prenosis completed its R&D stage and faced the challenge of developing an AI/ML product that would meet FDA requirements and pass SaMD (Software as a Medical Device) certification.
Prenosis needed a backend to aggregate and normalize data from EHRs and medical devices, perform calculations, and deliver results back to the EHR. FHIR was the natural choice for clinical data normalization. While most modern EHRs have FHIR capabilities, their APIs are limited:
- They rarely allow writing data back (effectively read-only).
- They don't support complex queries, such as advanced filtering, aggregations, or calculations.
- None of the major EHR systems currently support FHIR subscriptions, and they allow only limited load on their APIs.
- EHR systems' FHIR APIs differ slightly from each other, with variations even between implementations from the same vendor.
These limitations led Prenosis to adopt the Aidbox FHIR platform, which provided:
- A comprehensive FHIR backend with both read and write capabilities.
- Additional APIs, including SQL on FHIR, for advanced data processing.
- An HL7 v2 adapter to receive real-time updates from the EHRs.
- An EHR-independent data store offering a unified data model and APIs.
- The ability to handle data that doesn't fit FHIR using custom FHIR-like resources.
- Robust security features, including HIPAA compliance and high availability.
Solution
Aidbox receives information about admissions and transfers to the ICU via an HL7 v2 ADT feed using the Aidbox HL7 v2 to FHIR connector. All necessary clinical information is requested via the FHIR API and stored in Aidbox as FHIR resources. Each hospital has a slightly different configuration, and data normalization is sometimes required.
Prenosis defined additional custom FHIR-like resources to store service data:
- The GatherDataConfiguration custom resource records information about resource type, search parameters, and polling periods.
- LabSampleItem and other custom resources flag remnant plasma samples of interest from patients recruited into the study.
- Box and BoxSet custom resources track specimen packaging and delivery.
Clinical data is stored in Aidbox in the FHIR format and complemented with indexes for performance optimization. Sepsis ImmunoScore's custom logic, developed in Python, serves data to the ML Engine. The ML Engine, developed in R, processes data and sends the prediction back to the FHIR database in just 14 seconds.
The Sepsis ImmunoScore is stored as a Composition FHIR resource containing Observations. Prenosis uses a SMART on FHIR practitioner-facing application called Immunix to display the score to clinicians.
Feedback from the Developer
We've been working as part of the Prenosis product team, gathering requirements, defining clinical workflows, and prototyping a UX solution. Together, we created a clean and informative UI that is seamlessly embedded into the software providers use daily. A great UX comes with a cost. This interface requires advanced FHIR server capabilities including custom resource definitions and SQL on FHIR.
Custom Resources allowed us to persist all Prenosis data, including non-FHIR data, in the same database. While the majority of healthcare data was stored as FHIR resources, the application needed to store system configuration and workflow-related information.
While the FHIR API is excellent for interoperability, ML-based applications require more powerful tools for calculations. Aidbox doesn't fall short here, providing SQL API for all data. Prenosis could be one of the first companies to bring SQL on FHIR to production.
Finally, we have implemented a no-code data gathering tool to simplify the EHR integration process and applied an infrastructure-as-code approach to easily scale the solution.
Compliance, Security, QMS, FDA Clearance
Obtaining FDA clearance requires substantial preparatory work. A significant portion of these regulatory requirements has been met using Aidbox platform features:
- Audit Logs allowed Prenosis to obtain all logs simply by sending them to the FHIR Platform endpoint, creating only one integration point with Elasticsearch.
- Prenosis utilizes native support for OAuth 2.0 and Azure Active Directory integration within Aidbox, allowing single sign-on for hospital end users.
- The Aidbox Access Control module enabled Prenosis to develop and implement granular Access Policies.
Results
Prenosis success:
- Built the MVP of Sepsis ImmunoScore within 3 months.
- Successfully ran pilots with three U.S. hospitals.
- Received FDA clearance.
Aidbox FHIR platform allowed Prenosis to:
- Reduce development time by approximately 50% by utilizing off-the-shelf configurable APIs and FHIR storage.
- Meet regulatory requirements and obtain FDA approval.
- Ensure security and HIPAA compliance out of the box.
Sepsis ImmunoScore integration outcome:
- Connected with three large hospitals running Epic EHR.
- Over 75,000 patients, 200,000 encounters, and 6 million observations processed.
- Over 5,000 plasma samples processed and tracked over the past year.
- Nearly 1,200 patients recruited into a clinical study with Sepsis ImmunoScores calculated.
After two decades in healthcare technology and nearly a decade of working with FHIR, I've used and evaluated numerous FHIR servers. Aidbox stands out as best-of-breed, offering cost-effectiveness, robust features, and exceptional support.