Q&A: Designing for Patient-Specific Queries

Building a holistic information retrieval system for clinicians.

– min read
Role
Researcher · Design Lead
Timeline
6 months
Collaborators
Product · Machine Learning · Engineering · Clinical
Methods
Competitor analysis · Interviews · Data review · Design principles · Prototyping · Interaction design
The context

Scaling the product

Suki's reputation was built on ambient documentation: listen to a visit, write the note. As the world moved toward ChatGPT and Gemini, conversational chat-window interfaces became the default expectation for an AI assistant. The pressure on the team was to become more of a full-stack assistant ourselves, with our own information retrieval and Q&A system.

The brief

"Build our own chat window"

This was a vague brief in a crowded space. It was important to understand what our users wanted to query, and what our edge would actually be. We looked at two things in parallel.

Competitor landscape
What are other AI assistants doing?
Clinician questions
What do physicians actually ask?
Competitive landscape

What's out there?

Category Knows the patient? Knows the medicine? What it can't answer
DAX, Ambience,
Abridge, Heidi
Ambient scribes
Partial
Encounter context and note-level EHR data only. No question-answering against the chart.
No
No access to guidelines, literature, or drug databases.
Anything outside the live encounter including historical trends, drug safety against this patient's profile, evidence-based recommendations.
UpToDate,
OpenEvidence
Clinical knowledge
No
Completely patient-blind. No EHR access, no individual context of any kind.
Yes
Deep, cited, evidence-based answers from curated editorial content and peer-reviewed literature.
Anything specific to the patient in front of them, including their history, their labs, their current medications, how their profile changes the answer.
ChatGPT, Gemini,
Perplexity
General chat
No
No EHR access. No patient context of any kind.
Partial
Broad but unverified. No clinical trust model, no HIPAA compliance.
Anything requiring clinical trust, patient specificity, or cited evidence.
Table 01 – A comparison of AI assistant categories by patient and medical knowledge coverage
Question snippets

Clinical questions

A set of real clinician questions, surfaced through shadowing and interaction-data review.

Patient + knowledge

"Given her age, comorbidities, and medications, what's the recommended treatment approach?"

Patient history

"Has this patient ever been diagnosed with atrial fibrillation?"

Medications

"Is this medication safe for her given her kidney function?"

Labs & trends

"What's her A1c trend been over the last year?"

Patient + knowledge

"Does she meet criteria for starting a statin?"

Diagnosis & reasoning

"What is the most likely cause of her worsening shortness of breath?"

Guidelines & evidence

"What does UpToDate recommend for managing diabetes in a patient with CKD and cardiovascular disease?"

Patient + knowledge

"What evidence supports treatment option A versus B for someone like this patient?"

Patient history

"What changed since the last visit?"

Medications

"Is she taking any medications that increase fall risk?"

Patient + knowledge

"What care gaps should I address during today's visit?"

Fig 01 – A sample of clinician questions
The direction

Combined Reasoning

A physician doesn't think "let me check the patient's eGFR" and then separately "let me check if this medication is safe at that level." They think the combined question: is this medication safe for this patient?

A system that answers questions requiring both patient-specific data and general medical knowledge together.

Suki's existing EHR integrations already gave us live access to a patient's problems, medications, labs, vitals, and visit history, data. Combining that with general medical knowledge, in one surface, was something no competitor was doing.

Jump to solution
Use cases

Where and when Q&A shows up

01 · View schedule
Highlights since last opened, idle prompts
02 · Pre-visit review
Prep stage: patient history and context summary, vitals, meds, allergies etc.
03 · On-the-go prep
Compressed patient review, quick follow ups
04 · During visit
Orders, drug interactions, one-off patient history questions
05 · In-note editing
Post visit: Drug interactions, diagnoses, decision making support, clinical guidelines, billing and coding
Fig 02 – Q&A across different parts of the clinical workflow
Design guidelines

First principles

To ensure consistency and clarity, I first established the system's expected behaviours and the foundational design principles that would guide its evolution.

Present, not distracting
Information retrieval blends into the workflow. It never competes with the clinician's primary task.
Respect cognitive load
Answer, then step back. If a question has no role in the clinician's current process, don't surface it.
Build and retain trust
Medical software has no tolerance for ambiguity. Show sources, show reasoning, respect the physician's judgement.
Layout decision

Exploring layout directions

Before committing to a panel structure, I explored four layout directions for how a Q&A response could sit relative to the clinician's primary task on the existing product.

Option A
Bottom sheet
A response surfaces as a sheet that slides up over the main content, mobile-native.
Bottom sheet layout variant
Option B
Floating chat box
A floating window, resizable and repositionable. Familiar from consumer chat interfaces. Competes visually with the primary task.
Floating chat box layout variant
Option C
Overlay right panel
A panel that slides in from the right edge over the main content. The primary content remains beneath it but is partially obscured.
Overlay right panel layout variant
Chosen
Option D
Side-by-side
A response surfaces in a panel beside the main content, resizing the existing content view and visible at the same time as the chart or note.
Side-by-side layout variant

Side-by-side won out, for a few reasons:

Context stays visible
Keeps the main chart or note visible while reading the answer. A clinician still needs the chart in view to act on what they learn.
Respects cognitive load
A panel that covers the primary task conflicts directly with "be present, not distracting." Covering the chart to answer a question about the chart undermines the principle.
Scales across surfaces
Scales across surfaces where horizontal space exists, web and the EHR extension, while presenting as a full-screen experience on mobile.
Question and response types

Response categories

To understand the required response formats, I built a taxonomy of question types. Any question and response could be a combination of these distinct types. The surface itself would be dynamic, expanding, shrinking, or minimising based on context, importance, and screen size.

TypeExampleResponse format
Binary (Y/N)"Does this patient have an active POLST on file?"Text: Yes/no + context
Individual"Who is this patient's nephrologist?"Text: Short, specific answer
Cumulative"Show me the patient's current medication list."List, visual, or trend graphs
Computed"What is the patient's average blood pressure?"Interactive calculator, calculated answer
Time-specific"What was the patient's last A1c, and when?"Date-anchored answer
Trend"Graph the patient's hemoglobin over the last year."Charts and interactive visualisations
External"Show me management recommendations for peripheral artery disease."Longform answer, cited paragraphs (General Medical Knowledge)
Table 02 – Types of questions and their responses that the Q&A system would have to accommodate
Panel design nuances diagram
Fig 03 – Panel design nuances: zero-state information, response form factor, question suggestions, and panel/surface behaviour
Entry point

The badge

Suki is a voice-first product. While the 'Blue Dot' was central to its identity, clinicians rarely understood its purpose, and voice wasn't practical in shared charting spaces or patient rooms. We evolved the 'Blue Dot' from a voice trigger into a functional entry point for Q&A and broader assistant workflows.

The updated badge entry point

This mirrored where the broader industry was heading too: combined chat-and-voice interfaces were becoming the norm. The text alone told a clinician what the system could do. Typing and voice both worked, and a clinician could switch mid-task.

Fig 04 – Voice input via the badge
Fig 05 – Text input via the badge
Approval

Getting to yes

I built a Q&A vision demo grounded in day-to-day clinical workflows. It showcased realistic questions across different points in the clinician journey, multiple input modalities, and both patient-specific and general medical knowledge, while demonstrating how responses could adapt to different information needs.

Secured leadership approval for the unified direction: one Q&A surface, combining patient data and general medical knowledge, with a shared interface.

01

Pre-charting

Before a visit, clinicians need a rapid understanding of the patient's history. The experience begins with AI-generated summaries and suggested follow-up questions, allowing them to drill into abnormal labs, vitals, medications and prior visits. Optimised for heavy patient loads, the focus is on minimal clicks, low latency and quick retrieval over chart navigation.

Fig 06 – Pre-charting: understand the patient in seconds
02

Longitudinal review

Many clinical questions are about trends rather than single values. I explored longitudinal views of labs, vitals and medications, presenting changes over time through clear visualisations. This helps clinicians quickly identify meaningful patterns during the limited time available before entering the patient room.

Fig 07 – Longitudinal review: understand what changed
03

Post-visit reasoning

After the visit, clinicians shift from information retrieval to clinical reasoning while completing documentation. Integrated alongside the note editor, Q&A combines patient-specific context with general medical knowledge to answer questions about guidelines, drug interactions and care decisions. With the patient no longer present, voice also becomes a more practical input modality.

Fig 08 – Post-visit reasoning: combine patient and medical knowledge
Execution

Contract constraints meant Patient Data Q&A and General Medical Knowledge Q&A had to be built as separate tracks, on different timelines, with different vendor dependencies (UpToDate for General Medical Knowledge, our own EHR integrations for patient data).

Design separately to generalise across both.

Design details

Follow-up questions

Clinicians didn't start out knowing what Suki could answer, so suggestions in the empty state taught them what to ask. Suggested questions are based on prior questions, the patient's data and the clinician's specialty context. After a response, a good follow-up made it easy to keep rabbit-holing, one tap deeper into a patient's history or a clinical question, without retyping.

Follow-up questions
Fig 09 – Suggested questions across both engines, the future combined state, and follow up questions

Sources, citations, and trust

Every answer carried a collapsible "Sources" section. "Powered by UpToDate" surfaced directly in the response, a clinical reference physicians already trust. Trust itself came from the sources backing an answer, linked citations and a visible explanation of how the answer was derived.

Sources and citations – GMK
Sources and citations – Patient data
Fig 10 – Sources and trust signals

Graph responses

Trend and cumulative questions, the two categories from the taxonomy that needed visual responses, were represented with graphs or tables, depending on the use case.

Graph response – GMK only
Graph response – GMK + patient data
Fig 11 – Graphical and tabular responses

Managing perceived latency

A response that takes too long to appear, or appears all at once with no warning, breaks the sense that Suki is keeping pace with the clinician.

Suggested and follow-up questions were built with cached answers so they'd feel near-instant when tapped. Loader states told the user what was happening while answers loaded, building confidence in the retrieval process.

Latency loader states
Fig 12 – Managing perceived latency: preloaded suggestions and loader states

Mobile

The Q&A surface adapts to mobile as a full-screen experience – clinicians on the go or at point of care could access the same functionality from their phone.

Mobile screens
Fig 13 – Mobile screens
Future scope

What this opens up

A few directions came out of this work as designed concepts, ready to build on, rather than open questions.

Proactive nudges concept
Proactive nudges
The same data layer that powers Combined Reasoning, patient data plus general medical knowledge, doesn't have to wait for a question. A discrepancy like "BP abnormal but not addressed" or a suggestion like "consider switching this NSAID due to CKD" can surface unprompted, in-line, during note review.
Ambient-mode Q&A concept
Ambient-mode Q&A
Letting a clinician ask Suki a question while ambient recording is active, without disrupting the recording or the conversation with the patient.
Home-page redesign concept
Home-page redesign
Making Q&A a primary entry point on the home screen itself, with general medical knowledge suggestions ("recently trending in your specialty") alongside patient-specific prompts for the day's schedule, so Q&A is waiting for a clinician at login rather than something they have to remember to open.
Impact

Early outcomes

The first month of usage validated two distinct jobs-to-be-done.

Patient-specific Q&A
Patient data
325 questions  ·  33 clinicians  ·  12 healthcare organisations
  • 52% retrieved a single, most recent patient value – labs, medications, vitals, imaging
  • 28% summarised a previous visit or clinical note
  • 5% asked for clinical reasoning or recommendations – intentionally out of scope for the initial release, but the clearest signal of future demand

The Ask function is great. Accurate answers and so much easier than digging through the chart... essentially a more advanced form of the Find function.

– Physician, Accompany Health
General medical knowledge
Medical reference
221 questions  ·  18 clinicians  ·  5 healthcare organisations
General medical knowledge queries followed a noticeably different pattern. Clinicians primarily used the system as an in-workflow reference tool:
  • ~45% medication dosing, contraindications and drug interactions, most commonly during pre-charting. Given early users were primarily in primary care, clinicians typically entered visits with a treatment plan in mind and used Q&A as a quick reference to validate details.
  • ~30% clinical guidelines and screening recommendations, often immediately before or after patient visits. This highlighted the importance of mobile and point-of-care access, where clinicians were less likely to be using Suki on the web.
  • ~15% diagnostic criteria and disease management questions
  • ~10% medical calculations
Reflection

Learnings

01
Vision accelerates alignment
High-fidelity vision prototypes aligned leadership, engineering and product around the long-term direction far more effectively than specifications or roadmap documents.
02
Discoverability shapes expectations
Much of the early data suggested that users asked questions very similar to what was suggested in the UI empty states. Starter questions were important for setting user expectations about what the assistant could and couldn't do.
03
Design for real behaviour, not imagined behaviour
I expected conversational back-and-forth. Instead, clinicians averaged roughly one question per patient session, making speed and precision far more valuable than sophisticated multi-turn chat interactions.
04
Market timing is a design constraint
Conversational Q&A evolved rapidly during the project, making it essential to track assistant and retrieval experiences across industries – not just healthcare. The challenge wasn't simply matching emerging conventions, but deciding where to converge with the market and where to preserve opportunities for future differentiation.