To build or to buy? With healthcare AI, both may be the answer

When evaluating artificial intelligence strategies, hospitals and health systems face an age-old IT decision: build or buy? Keck Medicine of USC chooses a third path, blending both approaches.

Yesha Patel is associate director, data science and AI, at Keck Medicine of USC in Los Angeles, California. She holds a PhD in data science. Patel is scheduled to speak in an educational session titled “Case Study: A Blended Approach to AI Acquisition and Implementation” next month at the HIMSS AI in Healthcare Forum in Brooklyn.

Patel’s aim is for attendees to learn how her organization strategically combined custom development with select vendor systems to support effective and secure AI integration within the existing health IT infrastructure.

Finding the optimal balance

“In choosing a blended approach between a build-or-buy strategy, Keck Medicine of USC has found an optimal balance between innovation, speed to market and long-term sustainability when deploying AI systems,” she explained in a preview of her session. “Healthcare organizations face unique pressures, where patient safety cannot be compromised, data must be protected and clinicians must trust the tools they use.

“Relying solely on a build-or-buy strategy can limit an organization’s flexibility and effectiveness,” she continued. “By strategically blending internal development with external systems, Keck has been able to tailor AI applications to our unique workflows while still leveraging mature, proven technologies that reduce the time and cost associated with development.”

This approach acknowledges that no one-size-fits-all system exists in healthcare AI, she added.

“Certain use cases – such as predictive modeling for patient deterioration or administrative workflow optimization – may benefit from out of the box commercial systems with existing electronic medical records or third-party platforms,” she said. “Other areas, especially those requiring deep domain knowledge, contextual understanding, or integration with proprietary data systems and niche platforms, may call for customized systems developed in-house.

“The key is being deliberate about these decisions,” she noted. “Leaders should evaluate each use case to determine the right mix. We ask ourselves: Do we have the internal expertise? How critical is customization? What are the ongoing costs? Can we maintain this long term? These are not just technical questions – they are strategic decisions that shape how we operate in the years to come.”

How blended can be the better approach

So how is this blended approach more effective than other approaches when it comes to AI integration into an existing health technology ecosystem? Patel offers a peek at the complete answer she will provide in her session.

“A blended approach is more effective than committing fully to either a build-only or buy-only model because it provides agility and scalability while aligning with the unique clinical and operational realities of a healthcare organization,” she explained. “Building an AI system from scratch sounds appealing, until practical constraints like limited budgets, hiring challenges, and overstretched IT and data science teams come into play.

“On the flipside, relying solely on external systems can lead to limited customization, integration challenges with EHR systems not configured with modern AI in mind, and workflows that fail to reflect frontline clinical needs, all of which hinder user adoption and trust,” she continued.

A blended approach allows institutions to selectively extend capabilities through vendor work while maintaining control over core models and data pipelines, she added.

Reducing development time

“For example, a health system might purchase a third-party natural language processing tool to extract insights from radiology notes but retain control over how those findings are used by building its own clinical rules engine or prioritization layer – ensuring alignment with internal triage protocols and decision-making workflows,” she said. “This hybridization dramatically reduces development time without compromising specificity, quality or compliance.”

Patel went on to explain a challenge she and her colleagues faced when considering the buying AI approach.

“One significant challenge was the accumulation of technical debt from deploying disparate, one-off AI systems,” she recalled. “When you purchase standalone AI tools to solve specific problems – say, a sepsis prediction model from one vendor, a radiology tool interpreting images from another, and another for processing clinical notes – you quickly find yourself managing and maintaining a collection of systems that don’t work all too well together.

“Each system comes with its own infrastructure requirements, different data formats, APIs and integration protocols,” she continued. “Over time, this creates a patchwork of systems that don’t communicate well with each other or with your existing EHR and clinical workflows. We find ourselves spending significant resources just maintaining these integrations, updating different vendor systems on different schedules, and troubleshooting compatibility issues when one system gets updated and likely breaks another.”

Long-term technical debt

The real issue wasn’t just the upfront integration work – it was the long-term complexity Keck Medicine of USC was building into its environment, the long-term technical debt.

“Each additional point system added complexity to the IT infrastructure, requiring specialized knowledge and support teams to maintain, and create dependencies on multiple vendors with different support models and lifecycles,” Patel said. “Eventually, you reach a point where the cost and complexity of maintaining these individual systems can outweigh their individual benefits.

“This is actually one of the key reasons we moved toward our blended approach – being more deliberate about when to buy versus build and making sure any external system we adopt fits seamlessly into a unified, scalable system aligned with our broader architecture and workflows,” she concluded.

The HIMSS AI in Healthcare Forum is scheduled to take place July 10-11 in Brooklyn. Learn more and register.