What if clinicians had an IDE?

This started as a LinkedIn post about our paper with Ted Melnick and Andrew Loza at Yale. Here is the longer version.

As a software engineer, I take for granted what my IDE does: autocomplete that predicts my intent, inline warnings before I ship a bug, hover tooltips with exactly the context I need, all without breaking my flow.

Meanwhile, clinicians navigate fragmented screens, hunt for critical information, and catch errors only after the fact.

Ambient AI is a breakthrough, but not a complete fix

Ambient AI listening has emerged as a major advance in clinical documentation. By pairing passive listening, speech-to-text transcription, and large language models, ambient systems can generate notes nearly in real time and capture details that traditional documentation loses. That reduces the burden of the electronic health record (EHR) and lets clinicians focus more fully on their patients. It also substantially reduces after-hours "pajama time" and supports clinician well-being.

But it is not a complete fix. Documentation errors, the propagation of inaccuracies, and persistent challenges with reviewing and retrieving information remain.

A programming-inspired approach

To address these gaps, our paper proposes a comprehensive, programming-inspired approach to EHR workflow redesign, modeled after the Language Server Protocol used in modern software development. It is the protocol that lets an editor offer the same intelligent help, such as completions, warnings and context on hover, for any language.

If intelligent coding environments made developers so much more effective, imagine what the same infrastructure could do for clinical decision-making.

I'm hoping this sparks the right conversations.

The paper

Accuracy without compromise: a programming-inspired solution for EHR workflow redesign in the generative AI era. Edward R. Melnick, Felix P. Moser, Andrew J. Loza. Discover Artificial Intelligence, 2025. Read it here, or find it with my other work on the publications page.