Saathealth’s Journey in Unlocking AI and Data To Drive Health Behaviour Change

Consumer-centered digital ecosystems are emerging across the world, designed to seamlessly deliver the right care in the right setting at the right time. This is made possible only by integrating three critical aspects of the healthcare journey:

1. A focus on specific outcomes that define success for an intervention.

2. A system of intelligence that leverages behavioural, social, and health data to analyze patients’ needs.

3. A technology backbone that enables data and insights to flow between providers, patients, and the digital intervention.

So how does Saathealth fit into such digital ecosystems? We aim to integrate these crucial aspects by creating a modular digital health solution that helps healthcare organizations drive positive health outcomes. Our work focuses on evidence-based communication strategies and intelligent technology to help bridge the gaps in patient and health consumer care pathways.

Here is how our approach to digital health solutions has evolved over the last two years.

“How can we engage people long enough to make a difference?” — Predicting user churn and optimizing user retention

Imagine this — you’ve got a digital health intervention with great content, a sleek user interface, and cutting-edge features. You even managed to get users on board. But the inevitable roadblock creeps up — users uninstall the app. And if you want to influence specific behaviours, you need sustained engagement.

We realized the challenges of user churn and retention in digital health interventions. To deepen our understanding, we investigated the use of machine learning algorithms to predict, model, and prevent user churn on the Saathealth app. But what exactly is user churn? It’s the number of users who uninstall an app in a pre-specified time period.

The methods we used to predict when users would uninstall the app. Adapted from Ganju et al. (2021).

We found that users had the app for an average of 38 days before uninstalling. We then thought: what can we do to prevent users from uninstalling the app? We explored these messaging strategies to lower user churn:

• Targeted messages to minimally engaged users
• Incentivization through points and rewards
• Informed users of positive behaviour change trends
• Deploying notifications with engaging images and text
• Increasing frequency of notifications from 5/week to 40–50/week

We observed that the positive correlation between notifications opened and days on app increased from 37% to 48%. This means that our messaging strategies led to users opening more notifications, which in turn prevented them from uninstalling the app. Thus, effective and precise messaging strategies, backed by AI-powered algorithms, allowed us to lengthen users’ engagement with the app.

You can read our entire research article, AI for Improving Children’s Health: A Community Case Study, published in the journal Frontiers in AI.

“How do we get users to do more on a digital health intervention?” — Using machine learning algorithms to improve engagement

Now that we had users keeping the Saathealth app on their phones for longer, the next challenge to overcome was getting them to consume more behaviour change content to drive positive health outcomes. This is when we turned to recommendation systems. A recommendation system surfaces content that a user is more likely to like and consume. It does this by using data on what a user has liked and consumed previously, as well as what similar users have liked and consumed previously.

We built two recommendation systems for the Saathealth app: content filtering and collaborative filtering systems. To help understand these concepts better, let’s take a look at two online streaming platform users who have similar demographics. User 1 is a Korean drama (K-drama) lover who has watched several Korean-language shows. A content filtering model would surface more K-dramas in their recommendation list because of their watch history.

In the content filtering recommendation engine, the algorithm predicts what a user would like based on what they have watched and liked previously, and then provides recommendations accordingly. Taken from Ganju et al. (2021)

Given that Users 1 and 2 share similar demographics, a collaborative filtering model assumes that both the users would like and watch the same content. In addition, a collaborative filtering makes recommendations based on what the user has liked and consumed previously. So, if User 2 watched one or two K-dramas and User 1 has watched a lot of K-dramas, the model would also surface more K-dramas in User 2’s recommendation list.

The collaborative recommendation engine takes into account what the user has previously liked and/or watched and what similar users have liked and/or watched to generate recommendations. Taken from Ganju et al. (2021)

We ran two experiments to investigate the use of machine learning algorithms in improving content consumption and engagement in the Saathealth app.

Compared to users in the ‘no recommendations’ arm, those in the content filtering arm

• had a higher median number of videos watched,
• had 53.80% higher complete video watches,
• had a 13.96% higher proportion of correct quiz responses, and
• spent 9.94% less time on the app.

When we pitted content filtering against collaborative filtering, we found that users in the collaborative filtering arm

• spent 8.45% more time on the app at 45 days and 15.01% more time on the app at 90 days,
• completed a 50.00% lower median number of sessions at 45 days but the same median number of sessions at 90 days,
• had a 14.55% and 15.84% lower proportion of correct quiz responses at 45 days and 90 days, respectively, and
• had 66.67% higher complete video watches at both 45 days and 90 days.

Here’s the gist of what we learned — health recommender systems do not need to choose between content and collaborative filtering mechanisms.

Collaborative filtering may influence the early journey and discovery by matching content with similar user profiles, whereas content filtering may support depth of engagement by surfacing content types that users are most likely to engage with. So it’s a win–win situation as we achieved what we set out to do — improve engagement with behaviour change content on the app.

You can find out more about our experiments in our publication, Machine learning-driven recommender systems to improve engagement with health content in a low-resource setting: Poster, which was presented at ACM COMPASS 2021.

“Can we segment users based on their online interactions?” — Leveraging digital phenotyping to optimize engagement with digital health interventions

Now we know recommender systems can help drive content consumption, but we still had some unanswered questions. How can we make a user’s experience on a digital health intervention more meaningful? How can we optimize a user’s learning journey on a digital platform? We don’t want users to spend hours using a digital health intervention, but we want them to extract the most out of the time they spend on it.

Understanding users’ traits, needs, and behaviours is the first step in personalizing digital experiences, and this can be done by segmentation. Businesses use customer segmentations to maximize their marketing efforts, improve budget efficiency, and build meaningful relationships. Digital health interventions can also benefit from such segmentation exercises to personalize user experiences based on indications, risk levels, and user journeys, which could help deepen user engagement.

The way people use their phones and the mobile internet is about as unique as their fingerprint, and there’s a word for it — digital phenotypes. Digital phenotypes are the features that describe a user’s digital behaviour based on millions of data points — both active and passive — gathered from how they interact with their device. In the health context, this segmentation process can help in detecting a disease early, identifying worsening of symptoms, and designing a more targeted early intervention.

How did Saathealth start exploring digital phenotypes?

We used 15 million data points and four app metrics to segment our user’s digital behaviours:

• Videos watched
• Number of quiz questions attempted
• Number of app shares
• Days spent on the app

Our analyses resulted in six unique digital phenotypes.

1. Pathfinders:

• Pathfinders are high consumers of app content.
• They are willing to champion the app, as indicated by the high app shares.
• They are also the most likely to stay on the app for long.

2. Restless enthusiasts:

• When restless enthusiasts install an app, they do a lot — they consume a lot of content and champion the app.
• They seem to seek instant gratification, as once they have explored the app, they uninstall it.

3. Idle spectators:

• Although idle spectators would not go out of their way to uninstall the app, they are less likely to engage with content on the app and share the app.

4. Modest explorers:

• Modest explorers are willing to explore what the app offers.
• They will also stay on the app to do so.

5. Episodic seekers:

• Episodic seekers only look for information if they really need it, or to merely test the waters.
• They would then uninstall the app once their content needs are met.
• Think of someone installing an online shopping app to just get a Diwali outfit, and then deleting the app after.

6. Unengaged deserters:

• Think of these users as the polar opposite of Pathfinders.
• They’re the least likely to do anything on the app, share the app, and stay on the app.

So where are we going next with digital phenotyping and AI for health consumers?

1. We are replicating the phenotype distribution on several digital health interventions.
2. Our 2.0 digital phenotyping models will account for more than 30 unique health consumer variables, including biometric data integrated from wearable devices like smartwatches.
3. We will be able to report the natural longitudinal journey of various digital phenotypes.
4. Finally, our 2.0 models will demonstrate adaptive, precision health o experiences tailored to each digital phenotype, with the aim of optimizing engagement levels. For example, can proactive, reward-based interactions help convert Idle Spectators, who otherwise demonstrate low engagement, into engaged Pathfinders? We foresee immense potential for personalizing and adapting user experiences using digital phenotypes, and our future work will continue to build on our early discoveries.

You can read more about our digital phenotyping work in our white paper, Digital Phenotyping To Improve User Engagement With Digital Health Interventions.

Is there a secret tech formula to guarantee the success of a digital health intervention?

The short answer is no. And if there is one, we haven’t discovered it yet.

The best part about being in the tech industry is that everyone’s learning — including the big tech giants and humble start-ups. What works for one context and population segment may not work for another in this ever-evolving digital health landscape.

At Saathealth, we are leveraging AI-powered tools to keep learning — learning about how to predict and nudge users’ behaviours, keep users engaged, augment their app experiences, and drive positive health outcomes. And through our learnings, we aim to advance machine learning algorithms around recommendations by incorporating digital phenotyping data.

The ultimate goal with all our digital health interventions is to improve health outcomes. Our cutting-edge work in machine learning and digital phenotyping, coupled with our behaviour change expertise, can help bring us closer to this goal.

Schenelle Dlima is a Scientific Content Writer at Saathealth, which delivers positive outcomes for health workers and consumers with personalized digital tools.