White Paper
AI presents incredible potential for digital health, but it also requires more work and risk management to unlock long term value. Generative AI (GenAI) is quickly supplementing classical AI in operational and clinical healthcare tasks, and we believe a successful AI strategy for digital health products must include both technologies.
For AI to be effective and see more adoption in digital health, it needs to deliver value in the short term, not just long term. We recommend focusing on deploying use cases to create engaging and efficient digital health products that improve the user experience, while cautiously investing in more powerful clinical use cases for the long run, such as solving for treatment and care adherence.
While the use of AI in healthcare is not new, the surge of Generative AI in late 2022 and its subsequent spread into the mainstream has injected a renewed sense of possibility, urgency and also reservation in the world of digital health. Pain points that before seemed too big to tackle, like the burden of documentation for clinicians, may now be solved with GenAI. At the same time, new problems are emerging, like what to do about the proliferation of misinformation or outright hallucinations.1,2
In this white paper, we will examine the possibilities for AI in digital health, and show how we can start unlocking value in the near term with specific use cases.
1 https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10552880/
2 AI hallucinations are incorrect or misleading results that AI models generate. See also https://cloud.google.com/discover/what-are-ai-hallucinations
To keep things simple, we can think of AI models in two broad classes: traditional (or predictive) AI and generative AI. Predictive AI uses techniques like machine learning to recognize patterns in data and predict an outcome; generative AI effectively learns from data to create new content. Both can be, and often are, intertwined with deterministic rule-based intelligence to produce a certain output (e.g., calculate a risk score and flag if it exceeds a certain value).
Until recently, much of the use of AI in healthcare was focused on leveraging predictive AI models in the screening and diagnosis stages of the patient journey. Some of these models are built into existing healthcare software (e.g., Epic EMR) and devices (e.g. GE), but in many cases, particularly in pathology and radiology, regulatory-grade predictive AI is the core product. The FDA now counts 600+ AI tools approved as SaMD3, with more on the way. Calculating risk, spotting disease markers in an X-ray or flagging patients for follow-ups are all tasks that these AI methods are well suited to perform.
The advent of GenAI is now also allowing us to support digital health users with more content and data-heavy tasks in a way that was not possible at scale before large language models (LLMs). This means tasks like processing swaths of notes for a clinician or answering queries from a patient no longer have to be manual or static. The application of GenAI here, if delivered correctly and safely, has the potential to relieve the burden of care and improve access. The risk of using GenAI is that the outputs can be unpredictable and are less explainable, so we want to exercise caution about what tasks it’s used for, and what data goes into it.
An effective AI strategy in digital health can combine all of the above by blending predictive and rule-based capabilities for decision-making with GenAI functionalities for content-intensive tasks.
If we want to think about where AI gets applied in digital health, consider two types of use cases or “jobs to be done” along the patient journey:
While by no means are these two categories exhaustive, we see an array of promising clinical and operational use cases for AI in patient- and clinician-facing digital health solutions.
Broadly speaking, traditional AI would be used for clinical use cases, and GenAI more immediately lends itself to more operational use cases. However, this isn’t a hard and fast rule.
Predictive AI tools like “next best action” style models can be used to direct a patient or clinician to the next step in the workflow, in the same way that many consumer applications personalize customer journeys for their users based on previous actions, such as Netflix’s advanced algorithms that recommend the next best show for you to watch based on your previous viewing content.
Conversely, recent advancements in GenAI have made it possible to support clinical decisions with multi-modal data (e.g., Google MedPalm4) or to support care delivery in conversation-heavy therapy (e.g., mental health chatbots). The use of GenAI for clinical and regulatory-grade applications is still in its infancy and the regulatory frameworks for how to handle these solutions as Software as a Medical Device (SaMD) is still evolving. This is why, as an industry, we need to be very judicious about their evidence and safety before broad rollout.
As we can see, however, there is potential to leverage AI at literally every step of the care journey, in ways big and small.
AI USE CASES IN DIGITAL HEALTH
With such a gamut of AI applications in digital health, the fundamental question is one of prioritization. Which use cases should we start with? Which should we invest in for the long term? Decision-makers must weigh the value a use case can deliver to its stakeholders against the ease and cost of implementing it.
The effort to integrate AI into a patient- or clinician-facing digital health solution can range from low to substantial, depending on how custom and complex the AI model is. Creating an accurate and reliable predictive AI model for a high-stakes and complex outcome, such as disease deterioration, requires more time and resources than integrating a GenAI chatbot to summarize medical content. In fact, one of the most appealing aspects of GenAI is that it’s made it possible to incorporate AI capabilities into existing products with minimal development efforts; adding a GenAI chatbot may be faster than building a ML model from scratch.
Given that our customers’ digital health solutions typically come into play after a patient has been diagnosed and treatment has been selected, here’s what an impact vs. effort assessment of use cases for AI post-prescription might look like:
USE CASE VALUE VS. EFFORT
“Lowest hanging fruit” use cases are those where GenAI chatbots can be leveraged with minimal workflow changes (e.g., as an interface for patient and clinician educational content). This is doable today with manageable effort and delivers good value, especially in spaces where patient awareness or education are low, such as rare diseases or treatments that are complex to self-administer.
“Middle of the road” use cases can utilize a combination of rules, predictive and GenAI to meaningfully optimize the experience of a digital health solution in the short or medium term. This could be a mix of “next best action” guidance, chatbot-based symptom logging or auto-generated messages to the care team. These entail more workflow customization than just adding a chatbot.
“Holy grail” use cases are the ones that solve the toughest problems, but require significant research and custom development. This includes predicting adherence or predicting and intervening on treatment response in real time, based on patient ePROs.
Based on our experience and conversations with industry leaders, we believe AI in digital health needs to start with use cases that are deployable now and can show ROI to life sciences customers and users sooner rather than later, even as we invest in longer-term capabilities in parallel. For that reason, the use cases we’re most excited about are the ones that help digital health solutions become more engaging, efficient and effective:
For the long term, particularly if we think about utilizing patient-facing digital health solutions to manage care in an ongoing way, adhering to treatment and a care regimen is the most interesting problem to solve with AI, and one that can unlock great benefits in both clinical and economic outcomes.
At BrightInsight we always have an eye on data privacy and security. Our BrightInsight Platform regularly undergoes rigorous independent verification to ensure conformance with compliance controls, achieving certifications against global standards to earn the trust of our clients and business partners.
BrightInsight applies this same exacting level of accountability with our deployment of AI-powered features for our Disease Management Solution. We have developed concrete, industry-specific principles that get us beyond the generic buzzwords, and guide us to ensure proper and effective use of AI in our product development.
Learn more about BrightInsight’s AI principles on our blog.
Learn how our Disease Management Solution can help you improve the patient and provider experience when dealing with REMS protocol drugs.
Request a Demo