The advertising technology – or ad-tech – industry has made great strides in developing tools and techniques to leverage data in optimizing ad placements and delivering the right content to the right people at the right time. All these learnings can be valuable in the digital healthcare space. Digital healthcare is a rapidly growing, hot sector which aims to help answer the everyday needs of individuals across the world, improving the health and wellbeing of millions and making healthcare more accessible and sophisticated than ever before.

Digital transformation in healthcare means working with data at scale, sometimes to produce real-time insights on multiple clinical conditions for many populations –doing so in an appropriate, equitable, safe, and secure manner, with deidentified data where possible. Meanwhile, and across the road from traditional healthcare organizations, ad-tech, fintech, gaming, and other tech start-ups have benefited tremendously from data, analytics, and AI in the online tech sector. While they are not directly in the business of saving lives and improving people’s health, there’s plenty to learn from their best practices and tools.

In the ad-tech industry, data and analytical tools must be kept sharp due to the need to serve just the right ad to the right person, within milliseconds from the time that person entered a specific internet page/ mobile application somewhere around the web. It also needs to be based on that person’s preferences and history. This can be a highly complex challenge, especially when these ads make it to their destination by the billions daily – a tiny improvement (or mistake) can make or break the bank.

For this entire operation to function properly, appropriate data needs to flow rapidly across the network so that it can be prepared, cleaned, analyzed, used in training AI models, generate predictions and insights, and serve ads. All of this must be properly validated and monitored to ensure appropriate and responsible use, prevent errors, and reduce potential losses.

After a long career in advertising technology, I joined the healthcare sector. An important conclusion finally surfaced - why not try to apply some of these highly impactful methods within my current digital healthcare company? Such a multidisciplinary approach could lead to some very interesting and unconventional results.

Addressing a digital healthcare challenge with ad-tech-inspired methods

Ad-tech practitioners seek the ability to locate larger audiences than ever before, in a targeted, cost-effective manner. Lookalike models utilize and unlock the ability to build larger audiences from smaller customer segments, which have already proven successful for advertisers. Under the hood, elaborate AI models are trained to understand the characteristics of that group of successful clients based on relevant datasets.

Lookalike modelling allows advertisers to address audiences they might have never thought to target otherwise. It has indeed proven itself as a key AI-based technique that brings great value to an entire industry.

An illustration of how look-alike modelling works. Credit: Cedricc

Let's get back to digital healthcare.  The first step in the implementation of a digital healthcare solution is to reach out to the right population of consumers. This population will be most able to benefit from our solution. Assume we have developed an application that helps individuals with Crohn's disease better manage their chronic condition. Then, we can reach out to these individuals to get them to install our app and interact with it.

For some, our app could be super beneficial. They will enjoy engaging with the variety of features, could manage their condition properly, and avoid deterioration. Others may stop using the app shortly after the installation. It might be due to the content we offer, a UX issue, or simply because they don't have the time to invest. Can we figure out a way to target only those patients most likely to engage and benefit from our app?

With the help of lookalike modelling, we can create many characteristics (or features) that will describe this subgroup of people who might benefit from the app. We can do this by responsibly analyzing their medical history (previous diagnoses, procedures, prescription medications, lab results, and more), app engagement metrics, demographics, and behavioural profiles – again, using deidentified data when appropriate for the security of health plan members. Then, we use various natural language processing (NLP)-based, Deep Learning techniques to convert all these characteristics into dense vectors per person. This vector represents a person’s unique data fingerprint. Using this data, we can look for people with the most similar fingerprint to our "successful" person with Crohn's.

An illustration of an embeddings cloud for individuals with Crohn’s. A dot represents each patient. People with similar clinical histories are closer together. All individuals with be severe comorbidities are in red. Credit: Leonid Tsytrinbaum

An illustration of an embedding cloud for individuals with Crohn’s. A dot represents each patient. People with similar clinical histories are closer together. All individuals with severe comorbidities are in red.

When our model is in place, we can do more than just target new audiences. Now that we have our consumer "map" in place, we could potentially unlock a whole new level of analytical capabilities.

With clustering algorithms, we can create subgroups of individuals with Crohn's based on their traits, such as severity and demographics. Our app could then be tailored to meet the needs of each subgroup, improving our chances of making a difference for our users.

What does the future hold?

This is only the tip of the iceberg. There are so many challenges yet to be solved in healthcare, and solutions may be within reach using the lens of another industry. It will also take innovative individuals and leaders from many industries to help healthcare make this transition.

The future of digital healthcare will likely be defined by our ability to move away from one-off, ad-hoc analysis performed and presented in spreadsheet format to developing automated Extract/Transform/Load (ETL) processes, creating cloud-based feature stores and data warehouses, which could then serve as data sources for models and analytics presented via interactive visualization tools.

These could later be translated into insights and actions which can be automatically shipped to the relevant stakeholder/person for program management or intervention purposes. Such initiatives are already materializing in the healthcare industry; the future is bright, but the idea shopping from adjacent tech industries can be a worthwhile approach to help close gaps and reach such a future faster.

Written by Leonid Tsytrinbaum, Director of Artificial Intelligence at Carelon Digital Platforms, Israel