One of the most prevalent misconceptions in the AI world, and particularly among AI companies, is the false equivalence of AI models and AI products. However, in the same way that code alone does not constitute a product, an AI model does not represent an end-to-end solution, especially in a clinical environment. Several layers contribute to creating a viable healthcare product, of which the AI model is just one aspect. Let’s delve into some of the key differentiators when transforming a model to a viable product:
Embracing AI at the DNA Level
An interesting revelation from OpenAI’s acknowledgements of their team (comprising around 400 individuals) is the emergence of novel roles, such as the “experiment babysitter.” This mirrors our experience at Aidoc, indicating that operationalizing AI in a product requires a unique company DNA.
For example, at the heart of this transformation is the creation of an “AI Operations” (AIOps) team. Think about them, in very simple terms, as AI support. Their role is to monitor, validate and fix performance or data issues.
This is particularly critical in healthcare, where patient outcomes can hinge on the model’s accuracy and reliability.
This is one of many roles we have to enable the delivery of accurate AI at scale.
Mastering Data Normalization in Healthcare
A robust AI product necessitates a solid data acquisition and normalization layer. Given the unstructured, inconsistent and sometimes missing nature of healthcare data, it’s essential to unify and standardize the data into a usable form. This step, crucial to the product’s success, ensures the algorithm is fueled by high-quality and trustworthy data. From our experience, a new type of data layer, incorporating AI-based orchestration for ensuring data integrity and completeness, is needed. This includes components such as:
- Classifying image characteristics such as anatomies, contrast phases and reconstruction methods.
- Continuous data completeness, including automatic adaptation to newly added scanners, institutions and protocols.
- AI-aware data consolidation, for example in case multiple studies are performed for a single patient.
- Automatically identifying IT integration issues which make optimal protocols arrive late, and many more.
Without this layer, the risk of feeding inappropriate, redundant or sub-optimal data to the algorithms is high, such as providing a head-bleed algorithm with a neck series or a PE algorithm with a chest image lacking adequate contrast. This advanced data layer not only addresses the real-time evolution of data but also accommodates variations across different institutions within a health system, in a way that is transparent to the hospital’s IT team.
From Data to Action
When it comes to differentiating algorithms from products, generating insights is crucial – but it is only half the battle. An AI product must also have an activation layer to drive insight-based actions. This requires alerting the relevant stakeholders, prompting necessary next steps and integrating smoothly into the workflows of healthcare professionals. This often means building dozens of new integrations in an otherwise fragmented healthcare IT landscape.
This poses far more than an integration challenge. Adequately tackling it requires making it a central aspect of the product design from the first stages of development. For example, with certain medical conditions, simply alerting on the existence of the finding is not actionable and therefore irrelevant for the clinical workflow, without additional information such as the diameter of the finding or how its shape developed over time. Acquiring these characteristics for each finding requires an AI of its own, and thus should be taken into account from the first stages of development.
Strategy and Change Management
Even with all these components in place, operationalizing AI in healthcare demands a thoughtful strategy, robust governance, and effective change management. In the complex healthcare system, delivering the right insight at the right place is not enough; change management and governance are key to ensuring the insights drive the desired actions. Today’s fragmented approach, where each service line independently defines their AI workflows, needs to evolve towards a more coordinated strategy. We envision a new breed of human-AI clinicians building AI-driven workflows in the future. At our organization, we have dozens of people dedicated to change management and governance, working closely with our partners.
The Continued Healthcare AI Product Journey
An AI model may be the engine of the system, but an AI product encompasses the entire vehicle, complete with navigation, user interface and maintenance protocols. Building an AI product, therefore, requires wrapping the core model with these vital components to ensure robustness, practical utility and seamless integration into the healthcare ecosystem. From our experience, successfully developing these components is at least as challenging as developing the AI itself. The creation of an AI product thus represents a far more extensive and multifaceted endeavor than just the development of an AI model.
