HeLM - A Medical Discharge Platform

Inspiration

Our journey began when a team of physicians and nurses identified a critical gap in care across healthcare settings: discharge, the time when patients are sent home from the hospital.

This is a critical moment for the medical team to hand off care to the patient, but it can be complicated. Nurses often must introduce long, dense instruction packets to patients, and spend hours a day responding to patients’ concerns over text or call. Many of these concerns are repetitive and answerable straight from the information they received. The central motivation of HeLM was to make the discharge experience simpler, easier to navigate, and more efficient for care teams.

What it does

HeLM is a platform that allows providers to build robust educational chatbots in post-discharge settings by compiling document bases of protocols and patient info specific to their practice. Patients ask plain-language questions about their discharge instructions and receive clear, document-grounded answers.

HeLM is designed to help with routine topics like medications, wound care, activity restrictions, follow-up appointments, and warning signs. When the answer is not clearly supported by the documents, the platform can direct patients back to their care team or emergency services.

How we built it

We started by building a retrieval-augmented generation pipeline and a series of LLM agents to search the discharge document base and generate grounded responses. We expanded the system to make retrieval more structured and flexible, helping the chatbot find the most relevant information before generating an answer.

After several rounds of prototyping with clinicians, we improved the model with more robust safety features and built a front-end for providers to review patient responses and identify areas of confusion after discharge.

Challenges we ran into

The biggest challenge was balancing helpfulness with safety. Patients may ask urgent or personal questions, but the chatbot has to know when it can answer from the documents and when it should escalate.

We also had to think carefully about privacy and compliance, because medical teams each have specific protocols regarding use of patient data. We had to speak with people in healthcare regulation departments in the hospital to make this idea work.

Accomplishments that we're proud of

Simply put, our big accomplishment was getting this tool off the ground and into the hands of eager clinical teams -- trauma surgeons, weight loss specialists, hematologists, and more. We went through rounds of validation from doctors and nurses, and now we are gearing up for testing in patients. This project emerged directly from the experiences of these life-saving teams, and that is what we are most proud of.

What we learned

We learned that making the discharge process smooth for clinicians has wide-ranging benefits beyond the team’s efficiency itself. It would address crucial equity issues and make follow up visits less cumbersome for both doctors and patients.

We also learned that the hardest part of building safe, clinical tools is decision-making: knowing what to answer, what not to answer, and when to direct patients to human care. We went through more rounds of clinical validation than we imagined to validate this tool.

What's next for HeLM

We want to test the platform in real clinical environments and make the platform more safe for protected health information (PHI).

We believe HeLM could be a big next step to making progress in healthcare equity and quality, starting from the critical moment of post-care discharge.

Built With

• fly.io
• llama-index
• openai
• pinecone
• python