Feature Design + Validation

At Doc.ai, one of the company's primary goals was to make data collection as simple and fun as possible so that customers could quickly build a health profile and receive insights. The company used computer vision to predict metrics such as gender, age, height, and weight from a selfie (picture taken of oneself).

To ensure this feature could provide value for all customers (and not just a specific demographic), we launched the Inclusive AI Project to improve our algorithm's ability to detect physical metrics from a selfie, and to inform how the feature should be designed in the product. 

CONTRIBUTIONS

• Interaction design

• Prototyping

• Visual design

• Usability testing

• Project management

• Marketing

Design Problem

One of Doc.ai's core value propositions to investors and customers, was the ability to use technology such as computer vision, to capture health data. In theory, computer vision could reduce friction in data collection (snapping a picture rather than typing in data points) and had the potential to be more accurate than traditional self-reporting. Given the excitement around this technology, Doc.ai wanted to incorporate it into the product, and I wanted to understand the following before doing so:

Given the limitations of the technology, how can we design a computer vision feature that provides people with a positive experience?

Process

01. DISCOVERY

At the beginning of this project, I needed to understand: 

• How computer vision algorithms are trained.

• The shortcomings of the current algorithm.

• The criteria needed to capture accurate data from a selfie.

• People's perception of and literacy in AI technology.

My primary findings from the discovery phase include:

1. Engineering and marketing had their own goals/needs for the project: Improve the computer vision algorithm and generate publicity for the company.

2. The current algorithm was trained primarily with Caucasian and African American faces against a plain background and therefore provided less accurate results for other ethnicities and when the picture was taken against a busy background.

3. The face and body positioning had to meet a number of specific criteria for the most accurate results.

4. People generally trust robots or computers and expect these types of features to be accurate more often than not.

02. DESIGN

Based on my findings, the design needed to accomplish the following goals:

1. Set realistic expectations for the accuracy of the feature.

2. Provide guidance to help people get the most accurate results from the feature without increasing cognitive load.

3. Encourage people to submit their data and share the project to improve the algorithm accuracy.

My design process involved:

• User flows

• UI research

  • Best practices/standards (looking at competitors or similar features, in this case credit cards)

  • UI patterns (both for navigation and visual)

• Sketching

• Multiple design explorations

• "Mini" itereations (quick internal testing)

03. FEATURE TESTING

Given the more experimental nature of this feature, usability testing and gathering initial feedback with representative users was a critical part of the process. 

We tested the feature with people of different genders, ages, and ethnicities. However, the findings are limited as we did not have participants of non-binary gender, under age 18 or over the age of 50. 

The project was originally called Selfie2BMI, which focused on the technology rather than the purpose of the project, which was to develop a more inclusive algorithm.

Study Design

The study I designed helped our team understand usability and perceptions in the following categories:

• Project participation

  • Does the narrative clearly explain the project purpose? Why/why not?

  • Are people comfortable sharing the project? Why/why not?

• Taking a selfie

  • Do people follow the instructions for taking their picture? Why/why not?

  • Are people able to successfully take their picture?

  • Is it clear how to retake the picture?

• Results/data

  • How do people react to their results?

  • Do people understand what their results mean?

  • Do people understand how their results were generated?

  • Are people able to successfully submit or edit their results?

• Privacy

  • Are people comfortable using this technology? Why/why not?

  • Are people comfortable submitting their results? Why/why not?

04. FINAL ITERATION

Design Recommendations

Most findings from our feature testing involved copywriting, discoverability, and guidance.

The changes made in the final design come from the recommendations based on our findings. A short list of the recommendations include:

1. Remove the concept of BMI. It complicates understanding and is not necessary information based on the project goals.

2. Reframe project title and information to better communicate project purpose.

3. Simplify narrative for selfie results and how to edit that information.

4. Clarify how the data is being used and who will have access to it.

5. Use both visuals and text to instruct people on how to take the picture.

6. Give people more warning before the picture is taken.

7. Improve discoverability of the "retake picture", "submit results" and "edit results" buttons.

8. Improve discoverability of privacy information.

9. Make it easier for people to retake their picture from any point in the flow.

10. Better articulate the value of diversity in AI for health to encourage more people to share the project.

Here you can view the final version of the design implemented on the doc.ai website.  

Key Challenges

1. Inaccurate predictions. Because Doc.ai was in the process of training the selfie algorithm while we launched the project, the predictions could be completely inaccurate. Although the goal of the project was to improve the algorithm it was predicting information that could be very sensitive in nature for users (height, weight, age, gender). If inaccurate or inappropriately suggestive, the experience could evoke a very negative response. It was important that the copy and guidance through the experience set the right expectations for users.

2. Friction. Although the feature was proposed as a way to reduce friction from manual data input, it was not accurate enough to do so entirely. While gender was often correctly assessed, people still had to edit age, height, and weight. An within-subject A/B test could more objectively measure how well the feature achieved this purpose.

3. Predicting gender. This particular metric was a very controversial addition to the feature. It is worth noting that the label "gender" is treated as biological sex. Although one's sex is a very important data point for health insights, the design team did not feel it was appropriate to predict this metric. However, it was included in the feature for marketing purposes. I call this a "we can, but should we?" moment.

4. Guidance for body positioning. I ensured that people were properly positioned in the frame by adding visual guides to the selfie camera setup. To get the most accurate results (and have the best experience), it was important for users to hold a particular body position. The optimal implementation would have involved automated body-position detection (as used for credit cards), however due to time and technical constraints we were unable to implement this design.

5. Encouraging data submission. Because we were collecting sensitive data, we included information in the experience that helped users understand why we were collecting the data and what we would do with it.

6. Data validation. With the current project design, we were unable to verify that the data submitted by users were accurate. To balance these effects, we set up a smaller-scale internal study using verified data.

7. Mobile devices. Given the amount of traffic on mobile, the design had to accommodate both laptop and phone screen sizes. However, constraints for body positioning and the aspect ratio led to suboptimal layouts across screen sizes. With additional time, this could have been improved with more built-in guidance to reduce the amount of instructions in text format. The feature was also limited to desktop given that the algorithm was trained using desktop cameras.