The Hidden Enemy of AI Innovation in Healthcare: Technical Debt
Most healthcare AI pilots stall not because the model fails, but because of technical debt in the underlying data. The hidden costs, and how to assess whether your data is ready.
First published in The AI Health Pulse. Also on LinkedIn.
Healthcare is heading towards the widespread use of AI. If you’ve come across the difficulty of deploying AI in clinical practice, you appreciate that something is off. Despite immense effort, top-tier talent, and substantial funding, many projects do not progress to the implementation phase.
What is the reason? We are talking too little about technical debt.
I have over 20 years of experience designing data strategies in some of the largest and most complex healthcare systems in the US. From EHR modernization to the implementation of AI Analytics, I have observed an undeniable fact:
Yesterday’s infrastructure cannot support the innovations of the future.
Why AI Pilots Fail in the Healthcare Space
Healthcare is one of the industries where the AI boom is easy to recognize. There are applications in:
- Drafting clinical notes via ambient listening
- Accelerated and clearer billing via autonomous code generation
- AI assistance to derive insights from years of clinical documentation
While each of these applications is innovative and, at some level, proven, most implementations do not progress past the “demo” phase.
This is not the result of failed AI. This is due to the lack of organizational readiness of the underlying data, much like building a skyscraper on a swamp.
Understanding Technical Debt in Healthcare
It is common to hear the term 'technical debt' mentioned by departments that are IT-centric. The issue is that most healthcare executives view technical debt as an issue solely concerning IT. This is far from the truth.
Technical debt is a strategic issue and if you do not acknowledge it, you are throwing away millions.
In healthcare, persistent technical debt appear through:
Incorporated healthcare systems with misaligned EHRs
Clinicians who document the same concept in disparate ways
Outdated systems lacking metadata and/or documentation
Unstructured documentation making natural language processing (NLP) effectively useless
You can have the best AI model out there, but if the underlying data is disorganized, the model's outputs will be equally disorganized and unreliable.
Three Concealed Costs of Technical Debt
The following are the three most common ways technical debt destructively and quietly erodes the potential for Artificial Intelligence (AI) implementation.
- Impaired Inputs = Ineffective AI
If the data is inconsistent and incomplete, you cannot expect clinical decision support systems (AI) to work.
- Stagnant ROI with Diminishing Returns
Long lead times for AI implementation with little to no results eventually causes loss of budgets.
- Erosion of Clinician Trust
If AI cannot provide the support clinicians need, they will operate as if it is not available, and they will distrust all future innovations.
How Can You Assess If Your Data Is Actually Prepared?
Prior to launching any new Generative AI (GenAI) initiatives, consider the following:
Is the location of our crucial data identified?
Is our data consistent/integrated across systems and providers?
Have clinicians validated that data represents real-world workflows?
Is the absence of structured documentation balanced with unstructured strategies for NLP?
If your answer is no to most of these, you are likely not ready to implement GenAI at scale.
Best Practices of Leading Health Systems
Leading health systems treat AI like a long-term game, not a short-term bet. AI applications and systems are developing rapidly, and health systems are trying to stay up to date. Here are some ways leading health systems leverage the long-term investment of AI:
- Building the structure and system for data governance.
- Embedding AI specialists in clinical operations.
- Creating and capitalizing on scalable wins like ambient documentation and autonomous coding.
- Using GenAI to smartly and quickly normalize data.
Creating infrastructure for data governance and embedding AI practitioners breaks down some of the barriers that might impede the development of AI in a health system. Though challenging, the complexity of health systems is not a barrier for the implementation of AI.
Focus Where it Hurts (Predominantly, it is AI)
Avoid the FOMO (fear of missing out) for AI use cases. To gain the trust of health clinicians and system operators, it is critical to consider their pain points. Most clinicians are not interested in more technology or systems. They want less to do, and systems that foster AI to do less documentation and charting to improve health outcomes is a major win for health systems.
One signal a week. No noise.
Join healthcare leaders reading The AI Health Pulse every Monday.
Facing a challenge like this in your own system?
See how we approach healthcare AI consulting and data and analytics strategy, or book a call.