top of page
BayesEM
A new mobile application to support medical diagnosis in the emergency room
Team: 3 Developers, 1 Product Manager, Designer (Me)
Duration: 6 months (Mar-Jun, Sep-Nov 2022)
Impact: Designed & launched a 0->1 diagnostic mobile app for a client at Dartmouth’s Applied Learning and Innovation Lab, now used by 1,000+ users at Dartmouth Hitchcock Medical Center via Testflight.
Role: As Lead Designer, I led the end-to-end design of core flows, facilitated design critiques across design teams, collaborated closely with PM/Eng.
PROBLEM
Background
Advanced practice practitioners (APPs) need to make quick medical diagnoses in the emergency room (ER).
This is a challenge as:
1) They need to obtain information & make decisions in under 10 minutes
2) There is high stress, pressure & distractions in the ER
3) It is cognitively taxing & unreliable to rely on memory alone to recall literature across diverse ER conditions
BUSINESS
CONTEXT
The ER is one of the biggest revenue drivers for hospitals.
Improving the speed & accuracy of diagnosis allows hospitals to accommodate more patients, save more lives and grow their biggest revenue stream.
OPPORTUNITY
How might we support APPs & medical residents in developing quick, accurate and data-driven critical decisions for the ER?
SOLUTION
OVERVIEW
BayesEM
A medical diagnosis tool introducing novel Bayesian inference to help APPs & medical residents make quicker & accurate decisions in the ER
PROCESS
RESEARCH
Competitive
Research
I analyzed key diagnositc tools to understand current APP workflows, identify strengths and gaps, and define how BayesEM can stand out. Here are the top 4 competitors:
OPPORTUNITY
Existing tools are NOT holistic.
BayesEM will create a patient-centric approach where APPs can factor in patient history, underlying conditions & clinical intuition through an initial confidence input.
RESEARCH
User
Interviews
With limited prior knowledge of the diagnostic process, I conducted in-depth user interviews to learn their clinical decision-making process, pain points, user journey & physician-patient interactions.
I interviewed:
- 3 First year residents
- 3 Fourth year residents
- 3 Lead APPs
RESEARCH
Affinity
Diagrams,
User Journeys,
Personas
I led collaborative synthesis workshops to help the team empathize with users and extract key insights through affinity mapping, journey mapping, and persona development.
RESEARCH
I found 3 key user insights:
#1: Medical residents follow standard protocols, APPs rely on intuition
-> Tailored audience & purpose to be for medical residents as a learning tool (while supporting APPs with a lighter confirmation). APPs rely on their own fluid thought processes. Medical residents still developing their diagnostic framework found greater value in structured support.
#2: APPs think rapidly, do NOT interrupt their flow.
-> Refined ideal use case for after the patient history is taken and during the critical 10 minute window when physicians decide which tests to order. During patient interactions, interruptions can disrupt this cognitive flow.
#3: Mobile tools are quicker & preferred by young medical residents
Especially among younger & tech-savvy medical residents. It enables quick access to guidance or confirmation on the go.
DEFINE
Design
Goals
Based on research, I defined clear design goals:
Design for trust
Build confidence in high-stakes decisions by providing clear, supportive interactions.
Fast & intuitive user flow
Reduce decision-making time with simple, fast & frictionless interactions.
Minimize diagnostic bias
Help users to consider broader possibilities and reduce anchoring or unconscious bias.
IDEATE
Feature
Storm
I began with a team-wide feature storm to generate a broad set of ideas through open, unconstrained ideation.
IDEATE
MVP
Features
Given our fast-paced 10-week timeline, I defined and prioritized core features for an MVP. I broke down the Bayesian diagnostic process into clear, actionable steps and mapped them into a streamlined MVP foundation - alongside a few critical features from our feature storm.
IDEATE
User Flow
I explored multiple user flows - evaluating each based on clarity, speed, and alignment with real diagnostic reasoning.
Final user flow
With user insights to guide our decision process, I created a final user flow that:
-
Starts with the patient: Begins by considering the patient through a patient creation screen. This prompts awareness of potential biases and emphasizes holistic care.
-
Mimics patient-physician interaction: Take patient history -> break -> then order tests
-
Mimics diagnostic thought process: Patient profile -> diagnosis -> symptoms -> tests
-
Mimics intuitive verification - Allowing users to select an initial condition & confidence level. Residents/APPs often have 1-2 conditions in mind early on, supporting their natural reasoning process.
IDEATE
Next, I translated our MVP user flow into initial greyscale wireframes, in collaboration with 2 designers, iterating through ongoing team discussions.
Low Fidelity
Wireframes
PROTOTYPE & TEST
User Testing &
High Fidelity Screens
After validating feasibility with our developers, I moved into high-fidelity design. I conducted multiple rounds of user testing, starting with low-fidelity wireframes, which allowed me to quickly iterate and refine key interactions.
From Condition Comparison to Confidence Calibration
Initial idea: To help physicians confidently narrow down a diagnosis, I designed a probability comparison feature that displays the likelihood of the selected condition alongside the top five adjacent conditions. Based on the user’s symptom and test inputs, this allows quick comparison and helps distinguish between clinically similar options.
User insight #1: APPs and residents typically enter the diagnostic process with one or two likely conditions already in mind. Comparing these against a broader list of alternatives felt unnecessary.
User insight #2: During the Bayesian flow, users focused more on how their initial probability inputs influenced the final result.
Refined idea: I displayed how final probabilities would change if different initial confidence levels were selected.
I learned that clinicians wanted to see how outcomes shifted if their initial confidence was halved or doubled. As a result, we displayed three key points: halved, selected, and doubled initial probabilities. I also added familiar terminology "pre-test" & "post-test" probability to align with physicians' mental models.
Designing a Bayesian Flow Aligned with Physician Reasoning
Initial designs: In early greyscale designs, I prioritized speed - fastest path from input to result to match the high-pressure, time-sensitive nature of emergency care. I also addressed a key pain point by allowing customizable symptom input, which physicians emphasized as lacking in many existing tools.
User insight: I learned that physicians mentally group information into distinct “buckets”: general symptoms, physical exams, and tests. Within each category, they organize symptoms by nature, type, and urgency.
Final design: Reflected this "Buckets" mental model - structuring inputs to match how clinicians naturally process and prioritize information.
Aligning Patient Profile Creation with Physician Thinking
Exploration: I explored how clinicians mentally construct a patient profile during interactions.
User insight + refinement: Physicians often identify and recall patients by their primary symptom (e.g., “chest pain”) rather than by name or ID.
Final design: I reflected this mental model by allowing users to label patients by chief complaint. I also refined the demographics and notes section to focus on the details clinicians prioritize most (such as age, risk factors, and relevant medical history).
Meeting archiving needs
User insight: Physicians rarely revisit a patient once a diagnosis is made. Since patients are identified by chief complaint (e.g. “chest pain”), an unfiltered log can quickly become cluttered and confusing, making it hard to distinguish between current and past cases.
Final design: To streamline the archiving process, I enabled users to archive multiple patients at once and removed the archived patient log, which testing showed was unnecessary. I also added subtle red prompts to help users quickly identify outdated cases for archiving -reducing clutter & improving focus.
PROTOTYPE
& TEST
Visual Design System
I designed a modern & high-contrast interface to support fast, focused interactions in high-pressure ER settings:​
​
-
Dark Mode for Visual Ease: Easier on the eyes for APPs and residents working long shifts, while increasing visual clarity and element separation.​
-
Modern & Up-to-Date Aesthetic: A sleek, minimal interface reinforces a sense of innovation and information being 'up-to-date'.​
-
Distraction-Free Layout: Minimalist design reduces cognitive load, allowing users to focus on key actions under time pressure.
-
Visual Hierarchy & Urgency:
-
Red highlights critical patient information or urgent conditions.
-
High-contrast icons draw attention to required actions without overwhelming the user.
-
PROTOTYPE
& TEST
High Fidelity Prototype
View the final prototype here!
I built an interactive prototype to support user testing, communicate flows and interactions to developers, and help our clients present the product to stakeholders and investors.
The app is fully deployed on TestFlight and currently being used by 1000+ residents and APPs at Dartmouth Hitchcock Medical Center as a first trial experience.
REFLECT
Next Steps
Areas for future development & exploration:
1) Contributing to Medical Research & ML Model Through User Data: Once archived, each diagnosis stores rich data (symptoms, tests, history, and outcomes). In the long term, this anonymized data could contribute to research and improve the accuracy of Bayesian inference models. While the current user base is too small for statistical reliability, this presents a promising opportunity as adoption scales.
2) Implementing Continuous User Feedback Loops: As we onboard more users, embedding a lightweight feedback mechanism will be essential for capturing real-world insights and guiding iterative improvements during early growth phases.
3) Exploring Advanced Functionality in a Settings Page: - With growing usage and stored data, a settings page could offer meaningful controls - such as toggling between using only verified medical data vs. incorporating community-collected app data for inference, supporting transparency and user preference.
4) Preparing for Public Deployment: Currently available via TestFlight for investor demos, the app remains invite-only. Once stable and validated, we aim to deploy BayesEM publicly via the App Store to broaden access and drive real-world impact
bottom of page