AI-Driven Clinical Decision Support for Lung Physicians
Ongoing UX research and design within a multi-partner EU clinical AI consortium, translating validated models into something physicians can actually use.
Overview
Joining this EU-funded clinical AI consortium meant coming in at a point where strong psychological research on physician attitudes toward AI was already in place, but the product work was just beginning. The insights existed; a UX foundation still needed to be built.
My first responsibility was to build that bridge: define an early direction and create interface concepts physicians could meaningfully respond to. Working across a multi-partner environment combining clinical, academic, and industry perspectives, I focused on turning existing research into a practical foundation while staying open to what evaluations would reveal.
From a UX perspective, the challenge was to support physicians' decision-making under time pressure, fragmented data, and limited transparency in AI reasoning.
Research
A non-linear, research-informed and evaluation-driven workflow
This project evolved through overlapping phases shaped by existing research, tight timelines, and limited physician access. My focus was to create structure within these conditions while keeping the process flexible and responsive to new insights.
Requirements Analysis
Extracted product-relevant insights from existing psychological research: what physicians trust, what makes them reject AI suggestions, and where the literature had gaps.
Concept Exploration
Used AI tools (Base44, Lovable, LLMs) to rapidly surface structural and interaction directions. Given the project stage and timeline, this replaced the conventional low-to-high fidelity progression, prioritising evaluable output quickly. It also reflects how UX teams increasingly work, and an expectation I take seriously.
Interaction and Workflow Design
Defined the core experience logic to accommodate different physician decision-making styles without forcing a single path.
Interface Design in Figma
Created initial screens focused on information hierarchy and trust signalling, concrete enough for physicians to meaningfully respond to.
Early Physician Evaluations
Reviewed early screens remotely with lung disease physicians for directional feedback, not formal usability testing.
Figma screens available on request. Project name and branding withheld (ongoing, not yet publicly launched).
Design
Two modes for two types of doctors
How might we support different clinical decision styles without overwhelming or underserving any physician?
"Some cases only need a quick summary, but others require digging into the details before I can make a call."
User Pain Point
Physicians vary significantly in how much information they need before acting on AI output. A single layout either overwhelms those who need a quick overview or underserves those who require deeper evidence before deciding.
UX Solution
Quick Mode presents a high-level view for fast orientation. Deep Analysis exposes the AI's reasoning chain, evidence, and contributing factors in full. Both decision-making styles are supported without forcing either into a compromise.
The physician decides, the AI suggests
How might we provide AI guidance while keeping clinical judgment clearly in the physician's hands?
"AI can be helpful, but the final judgment has to stay mine."
User Pain Point
Physicians are wary of systems that feel authoritative. Since they don't input data directly, the interface had to carry the message that clinical judgment stays theirs.
UX Solution
The system presents multiple treatment paths rather than a single "correct" one, each with a High / Medium / Low Match indicator. When data is outdated or insufficient, the system flags the gap and recommends additional testing. Language throughout uses "AI Suggested" rather than "Recommended," reinforcing that the system offers support, not directives.
Explainability that earns trust
How might we help physicians validate AI reasoning without increasing cognitive load?
"If AI suggests something, I need to understand what it's based on before I can trust it."
User Pain Point
A suggestion without justification is clinically unusable. Physicians need to trace the reasoning before they can act on it, especially in complex diagnostic pathways.
UX Solution
A structured AI Reasoning section within Deep Analysis surfaces contributing factors, interpretation layers, and the evidence behind each suggestion. In Quick Mode, lighter-weight contextual panels provide just enough rationale. Neither version requires the physician to interrogate the system, but both make it possible.
When the AI is uncertain, say so
How might we make AI transparency clear enough for clinicians to rely on appropriately?
"How can I feel confident in a result if I can't tell whether the underlying data is solid?"
User Pain Point
Reasoning transparency is not enough on its own. Physicians also need to know how confident the AI is, and how well a specific patient matches the model behind a suggestion.
UX Solution
Transparency signals are embedded throughout: an AI Confidence bar communicates the system's certainty level, a Patient Matching score shows how closely the patient maps to specific treatment models, and links to external clinical resources allow traceability. Together, these help physicians calibrate when to lean on the system and when to investigate further.
Key Findings
What was evaluated and what we learned
I conducted semi-structured sessions with 2-3 physicians across three lung disease specialties (oncology, ILD, and infectious diseases) from health institutions in Portugal and Israel. Each session ran approximately 90 minutes: an opening hour exploring clinical workflows, decision-making patterns, and attitudes toward AI support tools, followed by a 25-minute interface walkthrough assessing comprehension, fit with clinical needs, and overall direction. Session transcripts were analysed thematically to identify patterns that directly shaped subsequent design decisions. The interface was also presented to consortium partners and the EU Horizon board committee, providing an additional layer of stakeholder feedback.
Key Contributions
Establishing the UX groundwork
Takeaways
The biggest learning: physician decision-making is far more variable than any single interface can accommodate. The dual-mode structure is a direct response to that — not a design preference, but a research finding made concrete.
Working across a clinical consortium (researchers, physicians, engineers, programme managers) also reinforced something about the role of design in multi-stakeholder projects: it's often the only shared language everyone can evaluate. A screenshot is accessible in a way that a research paper or a requirements document isn't. That's not a trivial thing.