A co-pilot for the diagnostic unknown
A concept exploring how a medical LLM could simplify rare disease assessment by turning complex clinical data capture into an intuitive, conversational experience.
Overview
Rare diseases present an unusual challenge in everyday clinical practice because clinicians are trained to prioritize common explanations first. As the saying goes, “when you hear hoofbeats, think horses, not zebras.” However, for patients with rare conditions, the zebra is exactly what may be hiding in plain sight.
Rare Disease Co-Pilot is a concept for an AI-powered clinical assistant designed primarily for pediatric clinicians, while also supporting primary care providers, genetic counselors, and genetic specialists.
The concept explores how a medical LLM could act as a clinical thinking partner - helping clinicians capture relevant patient information, identify areas worth investigating, and consider rare disease candidates without requiring deep expertise in rare disease.
Project Snapshot
Client
Rare Disease Co-Pilot (Design concept)
Year
2025
Company stage
Enterprise
Role
Senior Product Designer
Responsibilities
The opportunity - Rare disease expertise, without requiring an expert
The challenge wasn't simply giving clinicians access to more information. Medical information is already abundant.
The harder problem was helping a busy clinician understand what to look for next when a patient's presentation might point beyond the common diagnoses they encounter every day.
At the same time, the experience had to work within the realities of clinical practice: limited patient time, varying levels of technical literacy, and devices that may not always offer ideal performance.
That led to a simple design question:
Could AI make rare disease expertise feel less like searching a database and more like consulting a knowledgeable colleague?
This question became the foundation for the interaction model.
The concept - A conversation instead of a form
Rather than building another structured clinical data-entry tool, I designed Co-Pilot around an ongoing conversation.
A clinician selects an existing patient case and starts sharing what they observe. The medical LLM can ask contextual follow-up questions while continuously building an understanding of the case.
Clinical information isn't limited to text. Through a single Add action, clinicians can contribute different forms of evidence - patient photos, movement or behavioral videos, sounds, phenotypes, symptoms, and other relevant clinical data.
The Co-Pilot uses these inputs to guide the examination, surface areas worth assessing further, and eventually suggest syndrome candidates for further investigation.
The goal isn't to replace clinical judgment. It's to help clinicians know where to look next.
Key design decisions
Designing for consultation, not search
The most important decision was making Co-Pilot feel like a conversation.
Traditional rare disease tools can require clinicians to know what they're searching for before they begin. That creates a problem for clinicians without specialist knowledge of rare diseases.
A chat-like interaction changes that relationship.
Instead of asking the clinician to navigate complex medical databases or complete extensive forms, Co-Pilot can progressively ask for the information it needs based on what has already been shared.
The interaction is intentionally familiar: less like Googling a syndrome, more like consulting a peer.
Where the concept could go
Rare Disease Co-Pilot was created to explore a future direction for AI-assisted clinical assessment - particularly what happens when we separate capturing clinical evidence from the administrative burden of clinical documentation.
The concept suggests that a sophisticated medical AI doesn't necessarily require a sophisticated-looking interface.
By reducing the experience to a conversation and a simple way of adding clinical evidence, the interface can stay out of the clinician's way. At the same time, the intelligence behind it handles the complexity.
For me, that was the central design principle:
The more complex the intelligence behind the product becomes, the less of that complexity the clinician should have to manage.
One entry point for multimodal clinical evidence
Rare disease assessment isn't based on text alone.
Facial morphology, movement, behavior, vocal characteristics, phenotypes, and symptoms can all contribute useful clinical signals. Instead of creating separate workflows for each type of information, I consolidated them behind a single Add interaction.
This keeps the primary interface deliberately sparse while allowing the underlying clinical dataset to become increasingly sophisticated.
The simplicity is also practical: the concept needed to remain usable across clinicians with different levels of technical confidence and across clinical environments with varying device capabilities.
