Suggestions
Joshua Kelly
CTO @ Flexpa & YC Founder (W20)
Joshua Kelly is the CTO and co-founder of Flexpa, a healthcare technology company focused on patient-directed access to health data.12 He has been in this role since July 2020.4 Kelly has 8 years of experience in open source health data development, working on projects like FHIR and the "quantified self" movement.12
Prior to Flexpa, Kelly was the CTO at Universe, which was acquired by Live Nation/Ticketmaster in 2015. At Universe, he led product and technology efforts, growing the company to over $100M in gross transaction value.1 He was also the Founder and CEO of Affiga (YC W20), which was later acquired by OpenUnit (YC W20).1
Kelly has a background in blockchain technology, winning the Open Zeppelin Bounty at ETHDenver in March 2018 for his project using the Solidity library.3 He also participated in the ETH San Francisco hackathon, winning the BloomID Bounty in October 2018.3
At Flexpa, Kelly is part of a founding team that includes Andrew Arruda (CEO) and Thomas Hamilton (COO).1 The company's name stands for "Flexible Patient Access," reflecting their mission to provide patients with access to their health information and claims data.2
Highlights
There's really exciting research happening in the FHIR <> LLM space, like this FHIR-GPT paper.
Two quick thoughts on areas for further research, based in part on work Flexpa has done recently:
- Tool use / function calling is under-explored and extremely powerful
Giving an LLM access to the FHIR $validate operation and instructing it to input a FHIR resource can be effective at conversion tasks - and in particular testing conversions to Profiles
Similarly, giving access to a FHIR Terminology Service can be helpful
In this paper, the authors use few-shot examples and an in context lookup of SNOMED codes - $validate and Terminology Services available as function calls are likely to be much more effective
- FHIR Schema is promising because providing raw StructureDefinitions in context consumes a lot of tokens
With FHIR Schema, we can provide complete, succinct definitions of Resources and Profiles in context
Combined with tool use, an LLM can iteratively build transforms of unstructured data into structured data and receive feedback guiding its output
Software is going to melt into everything and nothing all at once
Asked Claude-3.5 to implement an example patient application using the Good Faith Estimates provided by the Patient Cost Transparency Guide (nitty gritty healthIT interop spec)
And it one-shotted a React app
I gave it PDFs of these pages:
- https://lnkd.in/gfZ3aKwd
- https://lnkd.in/gguHUQPk
Which are the technical specifications for how this is supposed to work
The idea here is that patients should be able to see, up front, what some expensive healthcare service is going to cost
It made this nice little demo app of how this will work in the future once these APIs are live and in production
An upcoming MRI Scan appointment shows an estimate link - and we can see the expected cost breakdown in the form of an Advanced Explanation of Benefits
