Homni Stroke Detector

Improving Human-AI Diagnostic Experience

Project Overview

Problem

Homni is an AI-assist stroke detection tool aims to improve stroke diagnostic efficiency and accuracy for various medical professionals. Upon I join the team, the product is undergoing a pilot testing process for the UX of the product with wizard of oz on the AI part. While AI accuracy is not the concern at this stage, the efficiency of the tool, however, has been questioned by stroke experts.

Solution

My goal is to understand how various medical professionals naturally integrate this tool into their workflows. By gaining this insight, we aim to offer a redesign that assists in reducing the time it takes for medical professionals to assess stroke symptoms.

Impact

- Exam completion time decreased about 50%.

- Started a research project that is forming collaboration between cross-universities stroke centers.

How can we make medical professionals feel supported by Homni AI to improve stroke diagnosis efficiency?

Explore The Problems

Expert Interview: Mismatch between users’ natural workflow and current functions.

Methodology: I conducted 6 rounds of unmoderated interviews with six stroke-related experts, including neurologists from clinical settings and Emergency Medical Technicians (EMTs) who handle stroke emergencies. We initiated our outreach by sending cold mails through LinkedIn and leveraging our team members' personal networks.

Process: During these interviews, we asked about their perspectives on AI-assisted diagnostic tool and presented a brief demonstration of the Homni stroke assessment flow to gain their feedback. Additionally, we inquired about their experiences with assistive technology and their collaboration with other professionals in stroke diagnosis.

Three Insights

How they deal with discrepency?

  • the logic of AI algorithm

  • experts don’t trust AI that much

  • how they deal with discrepency?

Experts pointed out that in real world not only stroke specialists would deal with the patients, but a lot other medical professionals from other department, and suggest us considering the big hospital picture.

  • we need to kind of consider people before or after this exam part.

  • how do they access, how do they interpret

  • the wording for people in other floors to understand how to do the assessment

Different users have different workflow when it comes to assess stroke symptoms, and the current design did not fit naturally into the real-world workflow.

  • Expert might not be the best target user for this tool

  • we lost focus of the target user from academic research to product phase

  • can focus on the less expertise users

Persona&Lit Review: Refocus The Users

  • Based on the generative research outcome, I decide to refocus on the users. I made some lit search and deciding 3 groups and recruit them for a usability testing to observe their how the tool fit their workflow.

  • After the interview with the stroke experts helped us find the wicked problem, I choose to conduct secondary research and collect survey data to better understand the users. I presented my findings with expert interviews with my team and discussed how we want to shape the product. Although we have 3 types of target users, we need a main focus. I discuss this problem internally with the team and we decide to choose the mid-expertise people because the are mostly likely to be the most benefit from it.

  • 15 survey data from UCSD hospital and medical campus showed 80% of medical professionals’ stroke related workflow shifting back and forth from urgent and non-urgent. While the current design only provide a fixed workflow.

  • Notes: While I revisit this procedure, I think we should use field study to observe their natural workflows.

Stroke Specialists

  • Need flexibility to change between in-patient management and emergent patients.

  • Trust issue with AI

  • See the value to help them leave some burdens

Urgent Care Personnel/medium experts?

  • EMTs, Nurses, those who have knowledge with stroke but might no too much?

  • hope AI can help with efficiency

Other Medical Professionals/ lower expertise

  • those are in medical field but not specialize in stroke

  • might benefit most from it

  • could expand to the care center workers.

  • Literature shows that diagnostic decision aids technology better help the less experienced medical professionals (xxx). 

Journey Map:

  • Upon review the personas, I created 3 journey maps for three groups of users.

From research, I conclude:

How might we?

From Research Insights to Design Criteria

Insight 1:

xxxxx

Iteration 1:

we need to be very careful about showing the discrepency

Iteration 2:

xxxxx

Iteration 3:

xxxxx

Usability Testing

Task 1: File Patient Info

Task 2: Conduct Exam

Task 3: Leave info to other people

Experimental Study: Medical Human-AI collaboration pattern across expertise. (Ongoing and became a new project)

  • During the usability testing I found out xxxx, to answer this question, we would need to design an experimental study. I talked to the team and begin a new research projects.

Reflections/Takeaways

  1. It is easy to lost target user during design, especially have multiple user type.