SAI CHAVA

AIR Module, Artificial Intelligence for Rare Diseases
Designing AI decision support clinicians can trust in the emergency department
Retrieval-based clinical decision support AI panel that integrates with existing clinical systems to help emergency clinicians recognise and manage rare diseases. 0→1 build

My role
Sole product designer: discovery, product definition, interaction design and prototyping.
Timeline
9 Months (2025)
Tools
Figma, Claude Code, Lovable, Anything AI, Kling AI, ChatGPT
Client
NHS / HealthTech Project
Interviews
Focus
Workshops
Hi-Fi Prototype
Wireframing
Usability Testing
AI
Remote Testing
Version control for keeping track of patient and patient status
05

Emergency patients flow is chaotic with each clinician tending multiple patients at once, and the same patients at different times. I added a version control to help clinician retrieve the status if the patient, such as their second visit after imaging.
Clinical softwares are often tacky and loaded with information, it was important to consider the clinicians next step in the work flow, which range from dismissal, treatment, testing, or imaging.
Interoperability with EPR functions such as Tests and Imaging.


Any suggested tests added in the plugin module, would feed it into the EPR, so no double effort is required.
Tabs to navigate between Tests and Imagining, with recommendations.
04
03
Presenting evidence for AI deductions - adding trust and agency


Suggested diagnosis and bullets explaining why, with symptom matching between patient and the condition.
Confidence levels represented in colour, commonality indicator and clinical match %.
While presenting a diagnosis, it was important to understand two things: what information (such as supporting symptoms, and degree of possibility, etc) counts as supporting facts, and what information would a clinician want to see as supporting facts for AI retrieved suggestions (such as confidence in diagnosis, and matching medical evidence)
Presenting potential diagnosis- understanding how, and what information to present.
02


Plain language summary, auto generated based on nurses notes, and vitals.
Match percentages, such as 72% for pulmonary tuberculosis. Based on history + clinicians input
What shifted: Naming a diagnosis to informing a treatment decision vs with the clinician having control over what's been suggested and why.
Slide to view v1/ v2
Predictive information based on auto-detected nurses notes and vitals changed into layered interactions with more clinician input. Results and confidence change based on input.
Factoring in patient information and linking to EPR in two iterations.
01
Opens with assessment history and symptom tags, as a fixed window


.png)

Opens with triage category, vitals, long-term conditions and allergies, as a flexible window
Plugin visibility and orientation on the EPR had two possibilities: fixed window and flexible window. Flexible window was more preferred by clinicians.
Iteration 3
Under progress.
Methodology
The first product decision was the AI interaction model. I compared three approaches:
Choosing the right kind of AI
How might we help emergency clinicians recognise rare diseases quickly and confidently, without adding to their workload or overriding their judgement?
The challenge
In most clinical scenarios, the person deciding is also the person who knows most. Rare disease inverts this: the patient often holds the condition-specific knowledge, while the clinician holds decision authority.

An emergency clinician has minutes to treat a patient whose condition they may never have seen before. For people with rare diseases, gaps in knowledge and bias mean care is often delayed or inadequate or denied: only 1 in 3 adults with sickle cell disease say emergency staff eased their pain quickly enough.
Problem
3.5 million people in the UK live with one of 7,000+ rare diseases, and too often receive delayed or inadequate emergency care.
Cause
Design decisions v1/v2
At a glance
Clinical softwares are often tacky and loaded with information, it was important to consider the clinicians next step in the work flow, which range from dismissal, treatment, testing, or imaging.
Functional prototype available on request
