In the last decade, there has been an increasing global necessity to effectively tackle mental health problems. From Internet support groups to online cognitive-behavioral therapy (CBT), more and more struggling individuals are gravitating towards online systems as new outlets for expression.
Recent advances in AI applications have made accessibility to mental health treatment easier than ever, where individuals have better, consistent service accessibility, as well as a much-needed layer of privacy protection that eases users in unpacking feelings of fear, guilt, and/or shame. Essentially, mental health AI-based applications are the key to unlocking deeper introspection within users, as it parallels real psychologists and clinical therapy sessions.
Still, most of the market for AI-based mental health applications are saturated with rudimentary technology that provides surface-level feedback to users. The current technology — which uses fixed, patterned algorithms — fails to adopt adaptive and empathetic qualities that are paramount to the client-therapist relationship.
The UCLA HCI department seeks to break the barrier in current AI-based technology using natural language processing (NLP), a branch of artificial intelligence that reads, deciphers, understands, and makes sense of the human languages in valuable ways.
Throughout the course of the past six months, I worked on Journey, an interactive web journaling application that uses NLP to emulate cognitive-behavioral therapy sessions. The only AI-based journaling application to date, Journey bridges the technological gap in mental health intervention by creating an emotional support platform that aims to minimize all roadblocks to expressive writing and to reduce feelings of depression, anxiety, and stress within users.
How can we make expressive writing as frictionless & comfortable as possible for the user?
I decided to look into the current market of the most popular AI-assisted mental health products to better understand how different companies utilize AI components to help maximize the self-reflection process.
Youper, Quirk, and Woebot are just a few of the AI applications that offered mental health intervention through introspection. All products emphasized the importance of personalization and positivity which greeted the user immediately. Youper and Woebot used automated computer-based feedback, while Quirk was more interactive and gave users greater freedom in determining how much intervention the AI provides.
Taken as a whole, the majority of AI applications have a very strict pattern of prompt-based chatbots to provide automated computer-based feedback. From this, I deduced that:
Despite chatbots attempting to prompt for personal messages, the Q&A-style interaction with limited answer options forces the user with heavy AI intervention to find their pain points.
With the exception of Quirk — which provides insight into cognitive distortions — the rest of the programs are not constructive and fail to provide users a framework towards healthier mindsets.
The universal feature that is found in all of these applications is how the AI attempts to discover the deeper rooted feelings within the user, where the end goal is to ensure that the user feels validated with their emotions.
Additionally, I did a lot of research into how specific cognitive psychological theories are able to provide constructive feedback in understanding the user. With every cognitive psychology theory, I found an existing NLP component in some other applications to brainstorm ideas for APIs developers on our team could use in Journey.
Cognitive distortions, for example, serve as a way to help users dispel the irrational thought process behind their false beliefs, which strengthens negative thinking and causes chronic mental illnesses like depression and anxiety.
IBM’s Watson NLU API analyzes cognitive distortions by compartmentalizing specific target phrases and keywords into emotions and sentiments (e.g. extremely negative phrases are flagged as all-or-nothing thinking, a type of cognitive distortion).
Self-disclosing personal and intimate information to others has demonstrated emotional, relational, and psychological benefits (if done in a safe and secure environment) as it emulates a sense of understanding.
Google’s Gmail Smart Compose helps generate real-time suggestions that not only reduces repetitive typing, but also predicts information using a neural language model to understand what the user will say next.
Part of my role in Journey was to determine how NLP components would be implemented in designing the user interface. We developed a chart to showcase a feature hierarchy to see how often the AI would interact with the user throughout the journaling process.
Four features tie in aspects of various NLP components: Reminders, Guiding Dialogues, Auto-Suggest, and Connection & Friendship. These features establish a framework to shift the user’s thought process to an environment similar to clinical therapy sessions.
Being the lead designer on this project, I wanted to integrate NLP in a meaningful way into the UI. I decided to walk through the logistics together with my project lead, Violynne, so we could validate each other on the various pain points for the four features we came up with before I got to designing.
The most fundamental part of designing for mental health applications is to reduce the cognitive load as much as possible for the user. This promotes expressive writing and ensures that users develop a personal relationship with the application.
Because of Journey's simple nature and past fieldwork research on features, I decided to skip low-fidelity prototyping and jump to mid-fidelity so our team’s developers would have a more effective framework to work on.
In this first version, I wanted to emulate skeuomorphic design with red lines (similar to a piece of lined paper) and stick to a minimalist style guide to reduce the cognitive load as much as possible. With minimalist iconography, one font family, and a couple of colors for cognitive distortions and types of days, these elements helped draw attention to the writing process more than anything else.
Our team consistently met with various professors and counselors to validate our prototypes, which sought to improve the information architecture by adding necessary features while removing useless ones. I had the privilege to meet with Dr. Greg Miller, a clinical psychologist at UCLA who specializes in CBT, who gave me questions into specific problem areas he found within Journey’s initial prototypes:
Individuals do not have to be shown consistent problems within their writing. Rather, the AI should consider how to make judgments about when to give a lecture on cognitive distortions and when to respond similarly to what a therapist would do.
Because the identification of the five main cognitive distortions is color-coded, this may cause users to visualize the same grey for different distortions. Implementing a colorblind palette would be optimal for all users.
The current version of the prototype underlines every sentence with color to showcase a unique cognitive distortion, where it then presents a similar example in the box. Although this is necessary, adding more features in the box below could make Journey a powerful tool in CBT.
Taking into account Greg’s feedback, I decided to undergo rapid prototyping to make sure that I addressed all of the pain points.
With rapid prototyping, I was able to come up with reflective mode, which essentially was the holy grail in solving all of the questions Greg had. In reflective mode, individuals can see everything — the unique cognitive distortions, auto-suggest replacements, guiding dialogs, and other features that may be implemented in Journey later on.
This reduced the cognitive load on the user during the expressive writing mode, as it was similar to regular journaling (only with minimal intervention with guiding dialogs). Additionally, this mode was able to make better use of functionality, using the visual stimuli as a call-to-action to reflect on writing.
Mr. Sun (the AI avatar similar to Mr. Clippy from Microsoft’s Office Assistant) would prompt the user after a while of being idle to reflect.
Following WCAG 2.0 standards for color contrast accessibility, I developed a universal colorblind palette and a dark palette to account for a larger demographic of users. In all modes, I scrapped the initial color palette and stuck to three uniform colors that associated all distortions with the type of day the user is having — all to further reduce cognitive load.
Dark mode sought to address nighttime use for night owls. Similar to f.lux, this would work by automatically changing color based on the timezone the individual was in, and if users wanted to change the mode indefinitely, they can use the toggle to switch from day to night.
I reduced the initial design a lot by omitting the cognitive distortions until the visual stimuli (Mr. Sun) prompts the user. Remembering the goal of designing this product, I aimed towards creating a frictionless experience in expressive writing, allowing individuals to write without obstruction.
In accounting for a larger demographic of nighttime users and colorblind individuals, the three modes (light mode, dark mode, colorblind mode) were created so people can focus on writing. Underlines were changed to highlights, but the original visual identity remained intact across all three modes.
Similar to Google's Smart Compose, the Journey AI would emulate “thinking” through ellipses. Additionally, being idle for too long would prompt the user to reflect on their feelings by clicking on Mr. Sun. After a short amount of time, the prompting message would disappear but reappears at the next idle moment to remind the user to reflect later.
It was really important to our team that we got the algorithm for auto-suggesting right. Upon being in reflective mode, users can click into negative phrases and replace them with more positive ones — we wanted to leave it up to the user to make these changes themselves, like a typical successful therapy session would.
Journey has been a journey. As mental health is not linear, our team is constantly trying to break into various ways we can improve features for AI intervention and evaluation.
Being the lead UX designer on this project for 9 months, I plan on doing more usability testing with a larger demographic to discover new features that would make expressive writing easier.
I am extremely grateful for the opportunities that Violynne Wang, my director, and Professor Xiang Chen have given me on this project. As we wrap this project up later this summer, I hope to improve user experiences for more AI-based projects with the UCLA HCI department, but stay tuned for more updates with Journey.
You can view my full-fledged research paper on Journey here.