Chatbots in Healthcare
A quick browser search will tell you that, between 1964 and 1966, Joseph Weizenbaum at the MIT Artificial Intelligence Laboratory created ELIZA, an early natural language processing (NLP) computer program. ELIZA, running the DOCTOR script, then became the platform for the first patient facing therapeutic chatbot.
In my current role in healthcare marketing, we have a number of pharmaceutical clients exploring the opportunities presented by ever evolving digital channels and understanding the history of the buzzwords can be critical in identifying which implementation strategy will be a winner in the market. Following on from a recent piece on digital therapeutics, we’re looking here at the evolution of chatbots in healthcare over time, some incredible people behind the technology, and strategies pharmaceutical companies could take now to get on top of the wave.
Want the answers now?
Talk to your target audience, develop any services you build in collaboration with users and *deep breath* spend money, smartly. Research, insight and a great, cross-functional team are vital.
A deeper dive down the wikihole of chatbots reveals that the standard program development software, on which ELIZA ran, was designed by Mary Allen Wilkes, who joined the Digital Computer Group at Lincoln Laboratory just as work was beginning on the Laboratory INstrument Computer (LINC) project (1961), which produced the world’s first microcomputer. Wilkes was one of the first computer programmers, as well as one of the first people to have a computer in her home when one was shipped to her parent’s house in Baltimore so that she could continue her work.
This strong foundation of work enabled Joseph Weizenbaum, widely considered to be one of the founders of modern artificial intelligence, to design ELIZA and the program’s most famous script; DOCTOR. The script was modeled on the work of Carl Rogers, the founder of the humanistic school of person-centred therapy which introduced the notion of asking open ended questions to facilitate a client’s desire to self-actualize (act on their own innate ability to determine what is best for themselves) through acceptance, authenticity and understanding.
Reflection in this fashion meant that the program didn’t need a real world basis to act from, it only needed to apply pattern matching rules to statements to trigger responses to key words. Question responses take open forms like ‘how did this make you feel?’ and ‘how do you feel now?’, whilst statement responses are empathic like ‘I am sorry to hear that’.
Build better voice apps. Get more articles & interviews from voice technology experts at voicetechpodcast.com
The interaction was designed to be a parody, but the results of tests with patients surprised (and perturbed) those running the experiment. Individuals had ascribed human characteristics to ELIZA and had formed emotional attachments to the ghost they imagined in the machine. They occasionally forgot they were talking to a computer, and famously Weizenbaum’s secretary once asked him to leave the room so she and ELIZA could have a real conversation. Weizenbaum was concerned by this, later writing;
“I had not realized … that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people.”
ELIZA — a computer program for the study of natural language communication between man and machine. (1966)
Despite Weizenbaum’s reservations regarding the human tendency to anthropomorphise artificially intelligent machines, the field has continued to grow as natural language processing capabilities improve.
In 2017, Alison Darcey, Instructor at the Stanford University School of Medicine, Department of Psychiatry and Behavioural Sciences, spear-headed the team that built Woebot. Woebot is a chatbot interface that is leaps and bounds ahead of ELIZA, with the scientific creds to back it up.
In an unblinded trial, 70 individuals age 18–28 years were recruited online from a university community social media site and were randomized to receive 2 weeks(up to 20 sessions) of:
- Self-help content derived from Cognitive Behavioural Therapy (CBT) principles in a conversational format with a text-based conversational agent (Woebot).
- Or directions to the National Institute of Mental Health ebook, “Depression in College Students,” as an information-only control group.
Those in the Woebot group, who engaged an average of 12 times over the two weeks, significantly reduced their symptoms of depression over the study period, while those in the information control group did not.
This work is part of the case that interacting with artificially intelligent algorithms may, in the case of alleviating the burden of mental health challenges, be good for us. There are a proliferation of apps in this space, from Youper through to the NHS Ieso offering, so, what makes a great interaction?
Building your own: Psychology and Interface
In the evaluation of Woebot, participants were asked what the best thing about their experience had been, the results are shown below.
The divide between content and process makes it clear that a cross-functional team with a range of skills is required to build an app that delivers to the user need. The research highlights areas of success, and a similar piece of work would identify opportunities to tailor the service to the user base and improve the offering.
From a content perspective, success in therapy has a number of variants, including willingness of the participant side. A service provider has control over only a few of these elements. A study by the Stanford Natural Language Processing Group conducted on 15’000 conversations within a crisis text service (with real counsellors) indicated that successful interactions displayed common characteristics:
- Adaptability to the situation
- Dealing with ambiguity by reflecting back to check understanding
- Responding with creativity rather than using generic phrases
- Making progress by quickly identifying the core issue and making a collaborative plan to move forward
- Creating perspective change by framing the conversation positively.
The closer an algorithm can get to mimicking these characteristics, the more likely the outcome is to be successful. You can see in the Woebot research that these aspects are present in what participants enjoyed the most.
So, once you’ve hired your cognitive behavioural therapy experts to support the process, you need to build a usability layer on top of all that smarts. Looking again at the Woebot chart, you need coders capable of building the algorithms that deliver the knowledge, UI and UX experts to save the user from being overwhelmed,video content creators and game designers to make the content interactive and data scientists and visualisers to serve up information enabling users to reflect on their progress and share that information with health care providers.
😲$$$ 😲$$$ 😲$$$ 😲$$$ 😲$$$ 😲
That looks and sounds expensive, right? That’s why it is critical to use deep audience insights to determine whether or not this is the right channel for your target audience. By ‘deep’ we mean getting to know them via whatever medium is suitable. Qualitative and quantitatve approaches, including online surveys and phone interviews, are complemented by data analytics including behaviour mapping, user profiling and A/B testing.
It might be that you don’t need to go as far as offering therapy, and a simple web based bot sitting on your site, raising awareness of a disease and helping prospective patients determine whether they should see a doctor, is enough. Or it might be that you need to go all out and create a therapy service for people with a specific illness that aims to increase patient adherence to a prescription and reduce relapse rates.
Answers on the back of a postcard, please.
