• github
  • medium.png

Copyright © 2019 Sayena Majlesein.
All rights reserved.

AI-POWERED VOICE ASSISTANTS

Analysis of AI-powered personal health assistant apps in order to determine ways to improve user satisfaction.

Organization: University of Washington MHCI+D Program 
Team: Sayena Majlesein, Angela Yung, Saransh Solanki

My Role:  Formative research and analysis

Duration: Sep 2018 - Dec 2018  |  HCID 530 A - Usability and User Research

 

INTRODUCTION

A chatbot, by definition, is a software system which can interact or “chat” with a human user in natural language. For our research study, we focused on AI-powered chatbots and their role as virtual health assistants. Ada, an AIpowered health assistant, claims to be an app that is able to bring personalized healthcare to its users. It is marketed as a “smart phone-doctor app”. 

 

The doctor-patient relationship is a vital component of the healthcare system, and its effectiveness directly impacts the quality of care, achievement of a successful treatment, and patient satisfaction. In order to replicate the real-life doctor experience, using these apps would need to feel as though the user is developing a similar rapport.

 

We aimed to look into the success of such health assistants to replicate the real-life doctor experience, the situations when usage of these apps are most effective, and the underlying reasons which make patients not trust them. In order to do this, we gathered information about users’ experiences with both their General Practitioner (GP) and the health assistant app.

OBJECTIVE OF THE RESEARCH STUDY

  1. Find out more about users of AI-powered personal health assistant apps and the situations these users find it most effective for.

  2. Identify the most impactful factors that contribute to user satisfaction when using an AI-powered personal health assistant app.

  3. Identify the current pain points of using an AI-powered personal health assistant app as a medical care companion

 

RECRUITMENT AND PARTICIPANT PROFILES

All participants were recruited via a survey sent to several social media platforms. The study was comprised of 5 participants, all of which matched the following criteria:

  • 18-30 years of age

  • Has at least a moderate knowledge of technology

  • Has access to Android or iOS devices

  • Recently (in the past 3 months) visited their primary care doctor to seek medical advice

  • Have used Ada in order to obtain additional medical information.

 

RESEARCH METHODS

Both the interview and usability study, were designed to be done in-person or remotely. Zoom application was used to conduct the interview and the usability study over video calls and to record participants’ phone screens.

Interview

The session began with the interview which allowed us to obtain a better understanding of our users’ experiences with real-life doctors. We used elements of directed storytelling to get a narrative about their experiences with finding a doctor and their most recent appointment with their GP. Participants were asked questions to about what made them use an AI personal health assistant and their experience and impressions of the app.

Information was gathered about both their experiences with real-life doctors as well as AI personal health assistants to pinpoint the major discrepancies between the two experiences.

Usability study

After the interview, participants were instructed to remember and input the symptom(s) of their most recent condition that prompted them to see a doctor. Participants were instructed to think out loud as they walked us through the application and explained how they used the features and services. The usability study was conducted to build an effective contextual communication between our problem space and participants’ experiences and interactions with AI-powered personal health assistant apps, and thus we modified the usability study to better understand participants’ conceptual models towards these applications.

 

The scenario was broken down into 8 sub-tasks and their success criteria.

After the interview and usability study users were asked following post-test questions:

  • What do you think is the purpose of this app?

  • How do you feel about this app?

  • How much do you trust or distrust this app?

  • What would make you trust this technology (in general) more?

The main objective of conducting a usability study was to understand how users were interacting with the app and whether there were any major usability issues that hindered users from having a successful interaction with the app.

DATA COLLECTION

An assistant was in charge of setting up the screen capture in order to record the user’s screen during the usability test and making sure the sound recorder set up and turned on for the duration of the session. The facilitator conducted the interview and administered the usability testing exercise. During the interview and usability testing, the note taker documented quotes and notes from the participant’s responses and actions.

 

ANALYSIS OF DATA

Interview

In the process of analyzing the interview data, the session notes were coded. After coding, we created the following affinity map for a visual representation of the results.

 

Ada’s AI service and logical flow are among the highest rated in the field, however, Ada is a misunderstood product. The recent tech-blog coverage refers to Ada as a “smartphone doctor app”, which users do not relate to. There are many facets and intricacies of a doctor-patient relationship which the AI-bot is not able to replicate. The experience felt mechanical and tech-driven rather than patient centered to the users. Ada’s role in the current medical journey of a user is very ambiguous. Although Ada is currently being used by more than 3 million users across the globe, most of them turn out to be technology-savvy and the app has not yet reached mainstream adoption.

USABILITY STUDY

Following the usability study, each sub-task was analyzed based on the following success criteria and usability metrics.

Success criteria for each sub-task

The sub-task was completed successfully when the participant indicated the task’s goal had been obtained. If the participant was not able to complete the task or requests and received sufficient guidance from the interviewer to move forward, the sub-task was marked as unsuccessful.

Usability metrics

We defined the following usability metrics and classification to better address usability issues and errors that we found in our study.

 

Sub-tasks analysis

The scenario was broken down into 8 sub-tasks.

INSIGHTS AND RECOMMENDATIONS

Recommendation Summary

  • The service should support natural language conversations using voice-based conversational interface

  • Enable photo and video input methods to impute symptoms and other vital signs easily and comprehensively

  • Allow users to edit their previously entered symptoms easily.

  • Generate easy and comprehensible demonstration of results and diagnosis using photo and video descriptions

  • Provide users with tips and guides on further actions

  • Store and contextually use patient’s information and medical history to provide a better diagnosis

  • Link the process of assessment with a real-life doctor (could be done remotely as well) to build trust in the system and the diagnosis

  • The service should work alongside the current healthcare ecosystem. Link it to insurance providers, healthcare providers, and pharmacies to make the process of finding care more streamlined.

 

PRIORITY RATING KEY

Priority ratings were assigned to our recommendations in response to the insights that we found from the interviews and usability studies. This rating scale provides information on the urgency of our recommendations. The ratings are based on the following criteria:

  • Impact on the user’s satisfaction
    Does the solution tackle a problem that left them feeling unsatisfied with the app?

  • Frequency of users mentioning the issue
    Is this a solution to a problem for many participants?

Users do not feel that symptom input on Ada is a comprehensive process.

Ada’s AI logical structure does not match users’ mental models. Often times it asks questions which feel unnatural to the user. Participants also mentioned that Ada’s symptom selection was not comprehensive and only allowed users to input closest matching ones. When this was the case, participants became doubtful of the accuracy of the results. Participants believed that trusting only what patients say without analyzing the other plethora of vital signs and medical information could lead to an incorrect diagnosis. As a result, participants were more likely to trust their doctor over the app when there is a discrepancy between the two opinions.

The information that Ada provides about the diagnosis is too technical.

The language used in the app was very technical which many users felt was unpalatable. They found it to be full of jargon, and as a result, it was not effective in providing useful information. In general, participants said they wanted resources and education about their medical conditions. While many doctors do not provide sufficient information post-diagnosis, participants felt that Ada provided less in the way of resources and feedback. Due to these limitations, participants believed the app to be insufficient even as a supplement to their doctor’s opinion. Participants would feel more comfortable with the app if they were provided with photos or videos to aid diagnosis and condition explanations.

The interaction with Ada is not natural and does not feel like they are talking to a real doctor.

Participants mentioned that the aspects of their interactions with real doctors that foster trust do not occur when they are using Ada. For example, typing to a bot felt like an interaction with a machine. At a real doctor’s appointment, you would able to ask the doctor further questions or request additional feedback. Ada currently does not use a holistic approach when gathering information to suggest possible diagnoses or to recreate the doctor-patient relationship. Relying on what patients say without analyzing the other plethora of vital signs and medical history could potentially lead to an incorrect diagnosis. In addition, Ada currently lists several possible diagnoses. In a written format, it was very noticeable to users and negatively impacted their trust.

Users do not trust Ada because they are unsure of its role in their healthcare.

Many of our participants used Ada as a second opinion. During the usability tests, a few users got a diagnosis that conflicted with their doctor’s opinion. They were quick to judge the app and were not likely to use it again due to the mistake. Another notable complaint was the app was not able to provide “next steps” or any additional feedback. Users simply did not know what to do with the information obtained from the app. At the same time, Ada does not keep a log of your assessments and diagnoses. It does not correlate with 2 adjacent diagnoses to find possible patterns. While Ada is marketed as a smart-phone doctor, users are finding that this is not an effective use of the current technology. Due to this major limitation, many participants were confused on what Ada’s purpose is and how it fits into their already established medical system. Participants mentioned that if the app was somehow linked to their doctor, they would be more likely to trust it.

USABILITY ISSUES

Participants pointed out that you are unable to go back to an old question without closing out of the app and you are unable to edit old symptoms that you had previously input. There were no other major usability issues that hindered participants from being able to successfully use the app. Most complains were about the output of the app rather than the usability.

 

STUDY LIMITATIONS

Only Ada users

We initially planned to test multiple health assistants but all of our participants happened to use Ada. We were able to gain many insights from just Ada so we focused our research on this particular app.

Remote Testing Issues

Given the time constraints of the project and the low usage rate of Ada in the US market, a few of our users had to be tested remotely. Remote testing required additional commitment from the participants to download applications and complete steps online. This proved to be a hassle and time consuming for our participants. We also had some issues viewing the screens while they were completing user testings. Zoom user had a lag so the screens were not in real time.

Lack of diversity in the participation group

All 5 participants identified themselves as early adopters of technology. Some of them tried Ada because of it getting featured in various tech blogs like Wired, which got their attention. The issues identified in the study are likely not exclusive to this set of users, therefore a more diverse group (different nationalities, cultures, and demographic) could have given us some more insight. The economy impacts the problems and needs around healthcare so future research could explore how Ada’s purpose changes according to these factors.

Voice interface requires additional research

While switching to a voice interface can help Ada improve its interactions with users, we were not able to gain information in this study about what form the voice interface should take. Future research should include both secondary research and user’s preferences in relation to voice interfaces.

 

CONCLUSION

In attempting to determine ways to improve user satisfaction with AI-powered personal health assistant apps, we studied the quality of the user’s experience when interacting with health assistant applications.

Users unanimously mentioned the gap between the experience offered by Ada and a real-life doctor experience. Ada is not patient-centric and does very little to replicate the intricacies of the current healthcare journey. Participants have yet to adapt it into their daily lives because it adds additional effort with little benefit. In addition, Ada still feels like speaking to a machine rather than a doctor. It lacks the personal connection with users and as a result fails to fosters the doctor-patient rapport. Due to the impersonal process, users are hesitant to trust Ada. The usability study results showed that it was not usability, but rather the perception of the app that led to low adoption rate.

These issues can be mended by designing a process which keeps the users in mind. To address this issue, we recommended providing a more effective and robust means to communicate with the app. We propose moving Ada towards a voice-based conversational user interface to establish a more natural interaction between users and the app. In addition, we believe that Ada can explicitly be bridged to the current healthcare system by assisting the user with the steps to do next. By linking the app to a doctor would further increase credibility and trust towards Ada and its diagnoses.

In conclusion, we hope that by proposing these key recommendations, we can reduce the gap between the current healthcare system and Ada. Furthermore, resolving the above-mentioned issues and addressing these problems can lead to a more trusted experience with AI-powered personal health assistant apps and can also effectively enhance user satisfaction.