"Metriport’s Medical API was essential for getting that extra context on a patient. The fact that there was more than just EMR data in there, things such as PDFs, TIFs, and JPEGs — that was huge. If a very important procedure happened, you need to know about it. I mean, how would you how else would you get that information?"
Back in June, Metriport was fortunate enough to have been an official sponsor of the inaugural XPC Hackathon, a healthcare + AI focused virtual event. With over 150 participants, the hackathon spanned across multiple tracks, including care insights, patient empowerment, and care delivery. Being the official supplier of medical data for the event, Metriport provided both de-identified patient clinical documents via our Medical API, as well as health and fitness data from wearables via our Devices API.
When it was all said and done, we were blown away at how many groups ended up using Metriport in their AI-based applications. While there were too many great projects to feature all of them, we took a bit of time to catch up with some of the ones that stood out to us during the event.
What did you end up building, as the official winner of the XPC Hackathon?
- We built Recco, a tool to streamline the medication reconciliation process for providers, based on large language models (LLMs) and patient medical data. With Recco, clinicians can query a patient’s medical record for their medication history and any personalized prescription-related questions (i.e., medication interactions, drug allergies, etc.). We built it using Langchain, GPT Index, and Metriport.
What main problem was your team trying to solve in healthcare?
- The US healthcare system spends up to $1.8B annually due to preventable adverse drug events and there are 251,000 deaths each year in the United States due to medical errors — one of the leading causes of these problems is due to incomplete or poorly performed medication reconciliation, which is the process of identifying an accurate list of a patient’s past and present medications, including name, dosage, frequency, and route.
How did Recco use Metriport?
- Metriport streamlined the process of parsing through a patient’s medical records over time and from many facilities, in order to identify and access the most relevant medication information in a structured format. We used Metriport’s Medical API to query for a patient’s consolidated medical record and get a list of their medication prescriptions from their record. We then used this information to create a custom dataset that was used to augment our GPT Index model.
If you weren’t able to use Metriport, how would you have accessed this data?
- We would have used the raw text from the EHR for the language model’s dataset. We suspect that using this dataset for our model would not have led to the level of granularity and accuracy on a large variety of queries that we were able to achieve with our tool using Metriport’s structured data.
When choosing vendors, how important is open-source software to you when making a decision?
- We appreciated being able to use Metriport’s open-source software, which is at a more affordable price point than similar tools. Metriport’s open source nature also allowed us a lot of flexibility and room to experiment as we designed and iterated on our solution (i.e., whether to use purely medication data, or both medication data and the entire medical record). We especially appreciated the Metriport’s team hands-on support in helping us utilize and customize their API and API queries for our specific use case.
You can check out the demo video for Recco here.
What did you end up building for the hackathon?
- What we built was basically an extended version of Gray’s chart review tool, which does semantic search over a patient's medical records. We went from doing search over a document that comes in via fax, to doing semantic search over patient data available in the Metriport sandbox.
That sounds really cool. Can you walk us through the full pipeline?
- Patient's data came in via a simulated fax, and we would then figure out who the patient was using AI.
- We then pinged Metriport for what information they had on that patient, and pulled that data.
- We then ingested that data through our pipeline, and then were able to do a search over the combination of patient data that came in via fax, as well as all the data that came in via Metriport.
How was it working with this data you had available?
- Metriport’s Medical API was essential for getting that extra context on a patient. The fact that there was more than just EMR data in there, things such as PDFs, TIFs, and JPEGs — that was huge. If a very important procedure happened, you need to know about it. I mean, how would you how else would you get that information?
Where did the inspiration for this come from for this?
- A few months ago, my grandma was in the hospital for a few days, and I was helping her sort through her medical records. She was there for four days and it was like a 200 page document! After I started going through all her records, I started to realize just how much information there was. I started talking to doctors and realized that these are specialists who care about very specific things about their patients, especially if a patient is elderly, or they have some sort of chronic illness. So as more and more data becomes available through things like Metriport and other interoperability tools, there'll be more and more information. So the big question is really, what's relevant here?
Anything else you would add?
- I think the Metriport channel was amazing, there was a great amount of collaboration and people asking questions. It was a glimmer of open source collaboration!
Peep the demo for Grays AI here.
Who is your team, and what did you end up building at the XPC Hackathon?
- Our team is Triage.ai, inclusive of 2 MDs, 2 engineers, and 2 repeat founders. We built a comprehensive 360 degree triage tool that streamlines the Emergency Room (ER) intake process. We incorporate a continuously improving data model, pulling in standardized data from medical health records, and leverage an LLM to aid timely triage decisions, ultimately reducing ER wait times and improving quality of care.
What main problem was your team trying to solve in healthcare?
- The Emergency Department intake and triage process is manual, and painful for clinicians, slowing time to disposition and treatment for patients. Basic guidelines for triage exist, but most decisions are a gray area and left to clinical judgment, patient’s past history, and collaboration between ER providers and consultants. This results in the Emergency Room becoming the #1 bottleneck of the hospital.
How did Metriport help you with your development?
- We leveraged Metriport’s Medical API to collect FHIR-compliant clinical data for Emergency Care patients. These served as the basis for synthetic data generation, which fueled our few-shot prompt based approach. Metriport provided exceptional examples of raw clinical data sourced from the Medical API, which led to the creation of test patients with matching demographics and queried using the Consolidated Endpoint. We were then able to pull specific FHIR resources used to fill out the patient information and journey for the triage use case. We opted to go with Metriport for ease of data access and immediate integration.
What do you like about open-source software?
- Open-source is quite useful as it allows us to understand the underpinnings of the software and ensure the data is being sent to us via means we’re confident in. Otherwise, we’re often at the mercy of the vendor itself and need to be more conscious of usage.
Anything else you would like to add about the hackathon?
- Dima was an exceptional help to fluidly integrating in the API and sourcing useful data. While we ended up generating much more synthetic data to complement our development, it would’ve been far more difficult without the ease of Metriport’s data.
Watch the full Triage AI presentation video here.
What did your team end up building for the XPC Hackathon?
- We built a website that allows users to view their lab results after logging in. There were two main use cases for AI on the site:
1) With the application of NLP, users can have emotional support from the system that scans the report and offer some common Q&A regarding the keywords of the report.
2) An AI chatbot that provides follow ups to users. For example, users can find out solutions to reduce their blood sugar, based on their specific eating habits.
How did you use Metriport in HealthTrack View?
What sort of data did you end up using?
- We ended up using FHIR (JSON) data provided by Metriport, which specifically included test results, vitals, as well as patient demographic information, which we used to power our application.
Overall, how was the hackathon experience for you?
- I really enjoyed the hackathon! It was a perfect combination for what I am trying to do as my career, since I am currently learning to code and I come from a medical background, having worked in medical device sales. I will be participating in any coming AI + healthcare-focused hackathons in the future!
Check out HealthTrack View's presentation here.
Special thanks to Paulius Mui, and the whole hackathon advisory board for running such an awesome and successful event. We can't wait for the next one!
- Metriport Team