Seeing the unseen - Helping physicians to leverage AI in detecting Prognosis

app.whiteboard.microsoft.com

Cobb, Jessica

Cobb, Jessica

Co-pilot

Notes

Prognosis Report

The patient is a 58-year-old male with a history of type 2 diabetes and hypertension. He presented with chest pain and was diagnosed with a non-ST elevation myocardial infarction (NSTEMI). A stent was placed successfully. Given his stable vitals post-procedure and absence of complications, the prognosis is favorable. He is expected to recover well with adherence to medication, diet, and cardiac

EHR Consistency Check

Co-Pilot Control

Risk Factor Scan

Additional checkpoints

Comorbidities

Accounted for diabetes and hypertension in recovery outlook?

Long-Term Risks

Considered heart failure or recurrent MI?

Rehab Impact

Rehab compliance flagged as a recovery risk?

BMI Factor

Have you considered how the patient’s weight (BMI 32) might influence recovery?

Adherence History

Is “expected to recover well” supported by this patient’s past adherence in the EHR?

Create Note

Ambient 1

Prog 2

Prog 3

My Note

ED Provider Name - 4/22/2025 11:25AM

ROS

History of Present Illness

Review of Systems

Objective

Physical exam

Response to interventions

Overview

Most physicians provide standardized prognoses to all patients, despite each having a unique medical history and lifestyle. This can lead to serious consequences, including the loss of patients. The goal is to design a system that can predict critical issues based on a patient’s history before the situation becomes dire.

Timeline

4 months

(Jan-Apr 2025)

Role

UX Researcher

UX/UI Designer

Platform

Web application

Status

Developed

Contribution

UX Research

UX Design

Vibe coding

Team

It was a research-informed design project, and I worked on it independently with guidance from my mentors as part of my capstone.

Some context on prognosis

prog

.

no

.

sis

the likely course of a disease or ailment.

How it’s done?

Day 1: Patient presents symptoms and Doctor makes a diagnosis (based on tests, history)

Then comes prognosis:

Doctor discusses what to expect going forward, how the condition may progress, chances of recovery, possible complications and outcomes

Some Background based on true story

From the story below, you’ll understand the unpredictability and how hard it is for physicians to predict outcomes

Two middle-aged men are both wincing in chest pain


Doctors reviewed identical issues and performed the same life-saving procedure on both patients

Despite the same care, their journeys end very differently


But why is it unpredictable?

Based on my interviews with 2 physicians, I found that several factors contribute to this unpredictability.

Limited information about the patient

Patient’s health uniqueness


Complex symptoms post surgical

I found a way to solve this problem

Based on secondary research, AI has been doing wonders in healthcare industry

AI can aid prognosis by identifying risk factors

Contextual understanding

Limited knowledge

Can work 24/7 accurately

Relies solely on data

Both have their own strengths and limitations

So, how might we enable seamless collaboration between Physicians and AI?

but there was a problem!!

While I started the research I realized that there’s already a research on AI assisted decision making systems in healthcare and it didn’t work because

Expectation

Physician

AI

Physician + AI

Accuracy

Reality

Physician

AI

Physician + AI

Accuracy

Over - reliance and

Less cognitive engagement

So I pivoted my problem statement to

How might we make AI and Physicians work together to improve prognosis without over-reliance or dismissal?

So the first thing I did is to understand why over-reliance is happening to understand the overlying problem

When conversational agents(AI) become more human-like

When users have limited expertise in a subject area

When AI provides quick answers, leading users to choose efficiency over critical evaluation

Priming beliefs about AI motives as empathetic can lead to trust but in reality it’s not the actual behavior

So changed my solution to be

seamless

integration

reduce
over-reliance

during

interaction

My solution

Solution exploration

Solution 1.1 - To address the over-reliance

Concept 1.1

Initial Approach

I introduced reasoning checkpoints - questions the AI asks during interactions to prompt physicians to verify decisions.

What I Learned

I realized the same type of questions might bore some users or challenge others, depending on their background.

Pivot

I shifted to context-aware reasoning checkpoints that adapt based on the physician’s experience and clinical context.

Outcome

This ensures conversations stayed relevant and neither overwhelming nor too simplistic.

Reasoning checkpoints to the physicians

Limited knowledge in the subject

Expertise in the subject


Eg: Before accepting this suggestion consider:

Check differential diagnosis

Check atypical symptoms

Eg: Does this match with your assessment?

Does it align with the latest research?

Solution 1.2 - Positioning of the app and seamless integration

Concept 1.2.1

Initial Approach

I initially planned to integrate the system with scribing, using voice recognition during physician–patient interactions.

What I scribing

Scribing refers to real-time documentation of clinical encounters either by a human or voice-enabled software.

Why I pivoted?

Scribing aims to minimize screen time for physicians during interaction, so adding extra steps could defeat its purpose. Also, many doctors prefer using their phone's microphone for convenience, being on a phone during patient interaction may not be ideal.

Key insight I received from my primary research

Unlike diagnosis, prognosis isn’t about real-time interaction. It focuses on post-diagnosis planning and communication, making it ideal for asynchronous AI support.

Concept 1.2.2

I then pivoted to chatbot integration, allowing physicians to use it for generating prognoses. However, I encountered significant friction the chatbot wasn’t seamlessly integrated into the systems physicians already use. This approach also risked doubling their workload.

I developed a chatbot using the Perplexity API and prompted it to act as an assistant

Concept 1.2.3

To improve workflow integration, I designed and developed a co-pilot system where AI suggestions appear as physicians write their notes.

How it works

As the physician takes notes, the system analyzes their input in real time. A personalized pop-up appears if any important detail is missing, based on both what’s being written and the patient’s medical history.

The co-pilot nudges physicians to reflect and consider alternative prognoses, factoring in variables like age, blood pressure, gender, and more.

User Testing Insights

I conducted usability testing with my peers to gather feedback on the user experience

The sudden pop-up could disrupt the physician’s train of thought.

It may also serve as a distraction during note-taking.

Final design

In my final design, I replicated an EHR interface and integrated prognosis features to ensure a seamless experience. Based on earlier feedback, I combined two concepts: constant context-aware checkpoints (marked as 1 in the image) to guide without overwhelming, and a risk factor scan at submission for final review.

app.whiteboard.microsoft.com

Cobb, Jessica

Cobb, Jessica

Co-pilot

Notes

Prognosis Report

The patient is a 58-year-old male with a history of type 2 diabetes and hypertension. He presented with chest pain and was diagnosed with a non-ST elevation myocardial infarction (NSTEMI). A stent was placed successfully. Given his stable vitals post-procedure and absence of complications, the prognosis is favorable. He is expected to recover well with adherence to medication, diet, and cardiac

EHR Consistency Check

Co-Pilot Control

Risk Factor Scan

Additional checkpoints

Comorbidities

Accounted for diabetes and hypertension in recovery outlook?

Long-Term Risks

Considered heart failure or recurrent MI?

Rehab Impact

Rehab compliance flagged as a recovery risk?

BMI Factor

Have you considered how the patient’s weight (BMI 32) might influence recovery?

Adherence History

Is “expected to recover well” supported by this patient’s past adherence in the EHR?

Create Note

Ambient 1

Prog 2

Prog 3

My Note

ED Provider Name - 4/22/2025 11:25AM

ROS

History of Present Illness

Review of Systems

Objective

Physical exam

Response to interventions

Set how active your co-pilot should be from Light Touch → Moderate Help → Full Insight

Standardize notes for clinical documentation - concise and structured

Context aware reasoning checkpoints

Junior physicians need guidance whereas senior physicians benefit from being challenged to validate their knowledge

Risk Factor Scan - for additional check

Acts as an intelligent assistant after prognosis writing, offering an additional layer of checks by analyzing the patient’s medical history

Trade offs

I’m not trying to build trust in AI among physicians who are resistant to it. My focus is specifically on those who are already open to using AI in their workflow.

I acknowledge that for physicians who resist AI, this solution might not be the right fit, they may simply ignore it altogether. Additionally, the AI model will take time to adapt to each physician’s preferences and needs.

Impact

My co-pilot system enhances clinical decision-making by providing context-aware suggestions as physicians draft patient notes.

My co-pilot system enhances clinical decision-making by providing context-aware suggestions as physicians draft patient notes.

While minimizing disruption to physicians’ workflow, it fosters a co-learning environment where both the AI and the physician improve over time. This can potentially lower prognostic errors and improve patient outcomes.

Reflection

Designing for AI–Human Collaboration

As a designer, I’ve come to realize how important our role is in bridging the gap between AI and humans. AI can feel like a maze, and it’s our job to guide users through it to show them what’s possible in a clear, meaningful way.

Growth Through Feedback

Throughout this process, I’ve learned to be more open to feedback and less defensive. This project helped me grow as a designer who is curious and collaborative.

Returning to My Engineering Roots

By the end of this project, I realized I can’t move away from my background as a software engineer. I began developing the solution to ensure it was technically feasible, especially since designers are often dismissed with comments like “this isn’t possible” and AI is not magic. I wanted to avoid that situation.

Other works

www.iaspire.accenture.com

About

My Squad

My Missions

League Board

Winners

Explore all the actions you and your squad can take to enhance your squad’s strategy

Weekly Bonus Cards

Submit your idea for the iAspire feature of the future


WEEKLY QUEST

300

All members of the league have saved their industry Aspiration

WEEKLY QUEST

300

Take action

All members of the league have saved their industry

Coming soon

300

Career Actions To Accomplish

Here all the actions you can take over the course of the league to help you and your squad to get ahead!

Define your career path and set and update your aspiration

Take action

90

Add an industry and delivery skill

Take action

80

Add professional learning on the career path page

Take action

30

www.iaspire.accenture.com

My Projects (8)

Select the projects to view the people details and team analysis respectively

Reset

Dove

|

Unilever

39 People

BMS

|

Unilever

24 People

BPP

|

Verizon

20 People

IOM

|

Verizon

19 People

Levels

All

SE

SSE

TL

AM

Roles

7 roles | 26 people based on selected levels

15

Application/Cloud Support

12

Packaged Application Developement

11

Packaged Application Developement

08

Application/Cloud Support

07

Application/Cloud Support

06

Application/Cloud Support

05

Application/Cloud Support

Skills

43 skills based on selected levels

IT Financial Management

Information Communication
Technology(ICT)

Microsoft Azure Architecture

Microsoft Azure Devops

Microsoft Azure Iaas

Microsoft Azure Paas

Microsoft Azure Paas

2

4

6

8

10

Number of people

No Proficiency

P1

P2

P3

P4

P5

Let’s chat!

Always happy to connect, brainstorm, or dive into something creative together.

This website is best viewed on desktop

Madhumitha Manimaran © 2023.