How to Boost Digital Mental Health Therapy with AI: Step-by-Step Journey and Technologies
The impending ongoing global crisis promises lots of problems. At the same time, this situation opens up new opportunities. Specifically, the wellness and mental health market, like any other industry, strives for digitalization. The area demands that market players be science-proven, data-driven, customer-centered, and, more importantly, AI-based.
Recently, we discussed pain points faced by the mental health market and its consumers in our latest publication: AI for Mental Health and Greater Wellness: A Myth or the Upcoming Reality? We named increased levels of anxiety and depression, the negative influence of COVID-19, worsened sleep conditions, and minimized mindfulness and concentration levels as critical issues faced by the industry.
To take advantage of the available opportunities, you need to become a leader and set the tone for digital innovation. You need to offer an entirely new customer experience and increase product performance to stay ahead of the competition.
If you still need to figure out if you can avoid nurturing digitalization, look at the recent research by McKinsey & Company.
In this article, the Neurons Lab team continues its research on innovating with the help of artificial intelligence in behavioral and mental health care.
Digital mental health market: Key drivers & top players
The market size of global mental health digital solutions is expected to amount to $5.5B in 2022. The global mental health app market size is slated to reach $10 billion in 2026 at a CAGR of 16.62%.
In addition, according to the National Alliance for Mental Illness study, more than 21% of USA adults experienced mental illness and more than 26.3 million adults have received virtual mental health services.
There are three main drivers of the digital mental health market:
- The pandemic, war, and upcoming recession have significantly increased the risks of anxiety, depression, and post-traumatic stress disorders.
- Increasing awareness about mental health and its significance has allowed a broader population to seek professional mental help.
- Technological advancement is a crucial trend gaining popularity in the mental health app market, driving innovation forward.
These days, the primary mental health apps market players are Mindscape, Calm, MoodMission, Headspace, Flow Neurosciences, Youper, Happify, Sanvello, NOCD, Talkspace, Addicaid, Silvercloud Health, Moodfit, BetterHelp, and eMoods.
In addition, we list our own top companies that are now actively exploring the market and trying new approaches: Moodwork, Mantra Health, Meditopia, wellabe, Mindstrong, ONVY HealthTech.
Have you seen any signals that indicate that you need to update your product?
If you have just launched your mental health product or it has already been on the market for some time, then you are probably faced with the need to measure the product performance.
Product performance is directly related to your business economy. For example, standard metrics like the unit economics, churn, and customer retention are all signals of a product’s solid health and operation.
On the other hand, most often, these metrics depend on either customer engagement and/or market competition.
So how can you evaluate the existing product value? First, you can answer the following questions:
- Who is the target audience? (meaning your product-market fit)
- What value do you give to the client?
- How do you give this value? (to evaluate your consumer experience)
- How do you manage customer expectations?
The answers to these obvious questions can provide insight into product performance and act as a signal to work on improving your existing product.
Next, we want to showcase the solution to named problems and some ways to answer these questions depending on the mental health products.
Do you need a mental health product metrics explanation?
Many companies operating in the mental health market are focused on providing technologically-advanced solutions like artificial intelligence in therapy and AI mental health apps to meet end-users’ needs and product-market fit to strengthen market position.
However, in the pursuit of increasing technical superiority, many focus on new technical features instead of on customer. But in order to meet end-users’ needs as much as possible, you need to be customer-first, or customer-centered.
What does it mean to be “customer-centered” in a technological way?
Let’s figure it out.
Being customer-centered means bringing extra value, service, individual recommendations, interventions, processes, and whatever else may be possible to personalize your offering.
Value and personalization are powerful tools! However, they are nothing without customer engagement.
Patient engagement is the holy grail for digital mental health. Unfortunately, if you google “digital mental health patient engagement,” you’ll be lost for weeks.
To create this additional value, it is necessary to build data-driven approaches. Enhanced personalization requires a very flexible and adaptive recommendation system, preferably one that functions in real-time. This solution can be achieved with an AI-based engine or at least an ML-based one.
Recent research in the Journal of Clinical Psychology discovered that the attitudes to mental health therapy are the best predictors of patient engagement compared to metrics like demographics, subjective norms, perceived behavioral control, intention, and eHealth literacy. Additionally, attitudes and adherence were reported as significant predictors for other digital therapies.
Well, this is pretty much as expected…
But how do we define these target attitudes within digital mental therapy, and how do we measure them?
First, the process may require a more sophisticated and regular assessment than just questionnaires.
And, more importantly, how do we manage these attitudes?
Some ideas exist that treatment attitudes and adherence are formed through expectation management and therapy routine experience.
There are only two ways one can use to manage the expectation and experience:
- Deep research for a universal algorithm that is capable of considering many factors.
- Adaptive recommendation systems that can quickly change the proposed interventions depending on the user’s mood, as effectively done via Tiktok, Instagram, and YouTube.
We do not believe in the first option, but are more confident in adaptive recommendation systems to get increased consumer engagement.
Science, Data, AI/ML — new magic pills for mental healthcare?
As discussed above, extra value and personalized consumer experience are the two main drivers of enhanced product performance.
Here, we want to present how science, data, machine learning, and artificial intelligence in behavioral and mental health care can become a competitive advantage.
Value, personalization, individualized value — all of these concepts include the ability to collect as much data as possible and conduct complex data analysis.
Patient data implies not only what the system knows about, but also what we actually can recommend to a specific patient. Any kind of content ranging from stories and music, physical activity and meditation, to complex therapeutic exercises can serve as a basis for personalized recommendation. All of these interventions should be labeled under specific terms that are commonly used in the mental health field.
From the view of patient data collection, the most prominent sources of information can be:
- Web portals or mobile applications are used to organize therapy. Here, the source of information can be medical records, questionnaires, text or video messages, records, as well as the behavior in the system, such as the choice and execution of recommendations.
- Smartphones and other wearable devices such as smartwatches and fitness bands. The technological basis here is Remote Patient Monitoring (RPM). We recently wrote an article about using AI with RPM: Speeding up AI integration in remote patient monitoring products.
- Photos, messages, posts, and articles in open public sources such as social networks.
We have listed the data sources and can now talk about data analytics and understand what ML and AI for mental health have to do with the concept.
There are four main components in mental health data analytics:
- Descriptive, for working with a massive amount of health data
- Diagnostic, for enhanced medical treatment
- Predictive, to foresee and eliminate risks
- Preventive analytics for mental health improvement
Not so long ago, we wrote a detailed article about health analytics. More information can be found here: AI/ML-Empowered Data Analytics for Digital Health.
Descriptive analytics is not only statistics on the collected data but also the extraction of various pieces of information and features. Big data is a very fertile ground for extracting new knowledge. Here we are doing data mining.
Let’s see how and what data it is possible to extract from daily user data.
1. Are questionnaires a downside of digital mental health therapy?
There are many questionnaires for diagnosing and revealing mental health problems, such as SCL-90-R and its various alternatives.
Of course, the direct diagnostic value of these questionnaires is undeniable. However, a typical problem with questionnaires is the barrier and discomfort they create for the user, drastically reducing patient engagement and feedback.
On other hand, if we look at questionnaires from the point of view of data science and ML, we will see that this is a multi-parameter classifier (or regressor). This means that optimization methods are applicable to work with questionnaires.
The first method is PCA (Principal Component Analysis) — which allows you to select a limited set of questions for diagnosing different problems together or separately.
Another method is DT (Decision Tree) — which allows for the replacement of the entire volume of the questionnaire with shorter branching sequences of questions.
Of course, other more sophisticated approaches, like graphs, can be utilized.
The essence of all these methods is to simplify the questionnaires significantly and make them unobtrusive without losing quality.
What could be more interesting is the ability to search for anomalies using ML in the answers to the questionnaires, which allows you to evaluate the quality and reliability of the estimates.
2. Сould biosignals bring more value to mental health assessment?
Another significant indirect method for assessing mental state is by using biosignals.
Modern wearable devices are equipped with a bunch of biosensors, including PPG, ECG, temperature, accelerometer, and more. What is especially important about such data is that it is collected automatically without needing to push and notify the user. As a result, this process significantly improves the return.
In most cases, frequency, time-domain, or entropy characteristics are first extracted from periodic biosignals. These aspects are used as predictors to classify or determine the degree of individual physiological and mental health states. This includes conditions such as stress, recovery, sleep quality, physical activity, burned calories, and anxiety.
The main methods of working with biosignals are standard machine learning (ML) techniques such as support vector machine (SVM) or decision tree (DT). As well, a convolution neural network (CNN) can sometimes be applied. This method uses the original raw time-series signals without extracting the ML feature.
For more extended information, you can see our article on data mining from biosignals.
3. How to spot socio-behavioral signals
So, can we use digital socio-behavioral signals for mental health assessment?
Social and behavioral activity in interactive or digital behavior consists of calls, SMS, tweets, and posts on Instagram, Facebook, or other social networks. Both the quantity of these activities and the intensity at different points in time, for example, at night, can act as significant indicators of one’s psychological state.
More advanced features, such as the top ties score, describe the ratio of top-N friends’ messages to everyone else. In this case, the primary ML tool is a regression model that allows you to select combinations of simple and complex features (predictors) associated with specific types of disorders.
There are a couple reasonably simple examples of how socio-behavioral signals in digital communications are associated with mental health disorders based on phone calls and Twitter data.
4. Could NLP listen as your mental therapist?
Text is another obvious (or not so obvious) source for a mental assessment and can be used to enhance existing AI mental health apps.
It is no longer new to anyone that Natural Language Processing works great in terms of “understanding” separate words, meanings, sentiments, messages, and even entire texts. The diagnosis of individual symptoms like emotions, moods, or diseases themselves, is no exception and is also available within NLP tools.
In this case, standard classification or regression methods from ML and DL are suitable. Without going into detail, it is only worth noting that neural network embeddings are used to work with texts and are a kind of text vector representation trained on large libraries, such as medical ones.
These neural network embeddings allow you to find a link between what is written in raw text and a diagnosis. Of course, it is optional to teach your system such embeddings yourself since there are ready-made options available such as Meta MultiRay or AWS Comprehend.
To determine the symptoms or disorders, all that remains is to take the text written by the customer and/or take text that he has already written on Twitter, Facebook, Instagram, or LinkedIn. In any case, do not forget to ask permission from patients before analyzing their texts.
5. Are recommendation systems a bottleneck for mental health therapy?
Detailed assessment and accurate diagnosis could improve the personalization of recommended interventions and, as a result, provide better treatment. This is the value that we give to the user — quite obviously.
So what do we do with the customer experience?
How can we improve the customer journey and experience with advanced recommendations, affecting output, churn, and retention?
There are standard recommender approaches like content-based, collaborative filtering, and hybrid that can be utilized.
These days, one of the top recommended algorithms is DeepFM. This algorithm combines factorization machines’ power for recommendation and deep learning for feature learning in a new neural network architecture.
Another significant step forward in the recommendation algorithm is a combination of batch training with online training, allowing for real-time adaptation. For example, this is how top recommendations work in social media like TikTok, Instagram, YouTube shorts, etc.
All in all, according to some developers, using DeepFM and online learning could improve the conversion rate by 30%, merchandise value by 30%, and engagement time by 50%.
6. Is modern conversational AI capable of providing a good user experience?
There is growing excitement around the potential of ChatGPT, a conversational version of the popular GPT-3 language model, for use in digital mental health therapy. Many believe that ChatGPT could be a gamechanger in this field, by providing a new and innovative way to generate messages and content for therapeutic experiences. While it is still early, the use of ChatGPT in mental health therapy is a promising development that could revolutionize the way we approach mental health treatment. If successful, ChatGPT could provide a much-needed tool for mental health professionals to help their clients in a more personalized and effective way.
What to do if you finally decide to implement ML and AI in mental health products
The first step in your journey of utilizing ML and AI in mental health solutions is to decide whether you have enough strength or need to resort to the help of specialists.
You can set up your own R&D team but let’s be honest — the process is long and painful. The best option would be to consolidate the efforts of your VPs in data science with an external R&D service agency (boutique option like Neurons Lab) — this is the co-innovation model. It is worth mentioning that, these days, many tech leaders are actively getting rid of their DS/ML/AI teams.
Secondly, you need to spend money on a feasibility study; a mini-project within which you will formulate use cases, study the feasibility of AI/ML tools, develop product metrics by which you will evaluate the effectiveness of new AI features, and make a backlog of 6–12 months to introduce new features. This discovery phase will allow you to assess the rate of investments and prioritize features for implementation in the AI mental health app or service.
How to choose an R&D partner for joint development of AI/ML-based mental health solutions
To effectively translate product goals into industry problems and AI tasks, you need to have a partner with competencies at the intersection of AI, HealthTech, and product development.
You must understand that AI and any R&D project implies an iterative process going from research to development to implementation and testing. This process is very distinct from typical software development.
Therefore, R&D project development requires a contract model such as Managed Capacity. On the one hand, this model allows you to rely on your partner’s competence (unlike the outstaffing model), and on the other hand, to avoid overpaying for risks (as in a fixed-scope model).
Effective operationalization requires experience in deploying AI solutions in the cloud and on mobile and edge devices. For effective deployment in the cloud, it is better to choose an authorized partner of one of the cloud providers.
Are you looking for an R&D service for mental health product development?
At Neurons Lab, we can assist you with implementing AI/ML into an existing or a new project within the mental health area. We can help proceed with all the steps mentioned and provide a solid evaluation of how AI/Ml can be utilized, particularly in your product.