Healthcare Implementation: Where AI Programs Go to Die

The healthcare industry is in the middle of a transformative moment. Every hospital system is integrating AI/ML and NLP technologies—let’s collectively call them "AI" for the sake of simplicity—across their operations and products.

These advancements enable enhanced diagnostic accuracy, personalized medicine, and improved patient engagement. Earlier versions of AI in healthcare date back to project Dendral, in the late 60s, designed to analyze mass spectrometry data and determine the molecular structure of organic compounds, which had applications in drug discovery and chemistry. Other early use cases included medical imaging, robotics-assisted surgery, and AI-supported medical diagnosis.

These days, faster and better algorithms can analyze vast amounts of medical data, like electronic health records, medical images, and genomic information, to more accurately identify patterns, make predictions, and process information at scale. This is particularly useful for clinical trial optimization and streamlining administrative processing.

Essentially, every aspect of healthcare—from diagnostics to delivery—can still be transformed. Even with the long history of adoption, AI in healthcare is in its infancy. Putting aside the technical challenges around poorly structured data, lack of combined expertise in data science and clinical work, limited validation and generalizability, some basic human factors have slowed adoption, such as ethical and privacy concerns, cost, and poor management.

Healthcare is a data- and investment-rich industry. But the speed of achieving impact with AI programs has been slower than you might expect. Ultimately, this has led decision makers and budget managers to view AI as experimental—something to be dealt with in the future—and not essential to embed within all aspects of day to day operations.

Healthcare, a sector accounting for 20% of the US GDP, remains largely untapped by the AI industry.

Other industries like retail, commerce, and finance have made considerable progress on implementing AI in their day-to-day operations with customer service chatbots, fraud detection, and virtual try-ons, which have led to reduced costs and improved customer engagement. Meanwhile, healthcare, a consumer industry worth $4.1 trillion with a 7.1% CAGR accounting for 20% of the US GDP, remains largely untapped by the AI industry.

Why Is My AI Not Working?!

Looking past this moment's generative AI fever dream, it's important to remember that putting AI into traditional IT is like putting a square peg in a round hole. We can't just apply the same old thinking. The existing healthcare system infrastructure is extremely fragmented and integration efforts are bogged down by middleware companies’ slow pace and internal “complex” IT workflows, which makes adding new technology into legacy systems very complicated.

On a national scale, traditional healthcare incentives are tied to a payer reimbursement model that rewards quantity of care rather than quality of care. In recent years, a shift towards value-based care models has aimed to improve patient outcomes and incentivize better care quality. Nevertheless, healthcare entities have responded by mostly managing the metrics to maximize payment rather than building systems that improve quality of care.

Another challenge is low levels of adoption by clinicians, partly due to resistance to change or fear of adding more work to an already busy schedule, and partly due to lack of the necessary foundation to implement new tools. When the economy is robust, AI projects tend to resurface. When it's contracting, AI and innovation take a back seat. This results in intermittent progress rather than consistent change management efforts to effectively launch these programs into the clinical practice.

Ultimately, the result is an incomplete ecosystem-wide scale-up, leaving the executives to ask, “Why is my AI not working?!”.

Implementing AI in healthcare requires reimagining how healthcare is developed and delivered while simultaneously figuring out how to implement it within the existing machine. It starts with digitizing the current systems and processes and most importantly developing a plan around managing development talent and optimizing implementation.

How to Leverage AI Without Getting Stuck During Implementation

Successful transformation programs begin by creating a strategy that goes beyond short-term gain and positions the organization to create long-term differentiation. This is achieved by aligning the company's strengths with the customer's needs and the market opportunity, taking advantage of data and technology to streamline and automate as much as possible.

On the healthcare delivery side, organizations should identify sources of data, ideally unique and customizable, and create a data management and MLOps strategy that can automate and monitor the development process. Centralizing program management via a "hub and spoke" model can reduce team silos, but it requires high quality talent and experienced change management to reduce single points of failure and the inevitable slow down in operations resulting from the limited capacity of the hub.

AI's speed of processing makes traditional modes of problem-solving obsolete.

With platforms like OpenAI's ChatGPT and Google's Bard, smaller organizations can begin to create faster POC models and MVPs and draw support from the community of AI experts and healthcare innovators to navigate development challenges. They can also combine their internal, open-source health population data or leverage synthetic data with powerful LLMs to create their own fine-tuned models or partner with third-party providers.

AI's speed of processing makes traditional modes of problem-solving obsolete. It demands a rethinking of existing workflows. This is critical in the R&D phase to ensure the new solutions are truly focused on delivering results and not merely a band-aid to the old problem. For example, traditionally radiation oncologists have relied on manual examination of mammogram images to detect and diagnose breast cancer. Now, AI models trained on large datasets of medical images can learn patterns and features that are indicative of specific conditions or diseases. They can also be used to automatically analyze and interpret new images to identify subtle patterns potentially missed by human interpretation.

Upstream of care delivery, in drug discovery, companies like Owkin are serving larger organizations like Sanofi to quickly transfer critical knowledge and access to technology to enable faster innovation.

AI programs can quickly lose steam during implementation if they're cut off from the broader ecosystem before the program is fully integrated into the company's operations. To avoid this, AI programs must go beyond the development and technology departments and become a company-wide initiative with clear areas of responsibility and measurable goals and objectives in every area of the business to transition from a domain-only to an ecosystem-integrated program.

Another challenge is to upgrade existing workflows and connect the many disjointed health technology systems. Major shifts to any single or connected systems can be quite expensive and disruptive to daily care. While that's a real issue, the hidden cost of playing the same old game will be much more over time. Meaningful digital transformation is critical in driving long-term impact.

Achieving Impact: From Domain to Ecosystem Integration

To win in the AI race, companies need to navigate through the three stage gates that stand between wanting to leverage AI and driving impact from it.

For any company-wide initiative to be successful, the organization needs to be aligned, motivated to implement change, and stay consistent over time. Integrating AI will require the same level of early investment and long-term execution to fully embrace its impact in the ecosystem and not just one domain or area.

Here are a few things to keep in mind when building and implementing AI programs:

1. Take an inventory of all assets, talent, and internal data sources to develop a strong understanding of the areas of strengths and weaknesses. Assess the quality and structure of the data (e.g. synthetic, alternative, open-source, etc.) and identify additional data, technology, and talent needed to drive development, testing, and implementation.
2. Define and rank the problems you intend to solve, their cost, and RoI.
3. Develop a long-term talent, advisory, partnership, and vendor management plan. Compliance risk associated with unvetted AI companies and poor infrastructure planning can be detrimental to the program and result in data and privacy breaches and useless solutions, so they should be investigated early and monitored continuously.
4. Curate a list of existing systems that may be impacted by AI and what other systems and workflows are needed for implementation. For example, integrate AI into existing healthcare systems, such as electronic health records or diagnostic workflows.
5. Develop a strategy and tactical plan for an ecosystem-wide implementation and change-management. Foster collaboration between data scientists, healthcare providers, and IT professionals to bridge the gap between technology and clinical expertise.
6. Develop adoption success metrics and triage workflows.

The road to AI transformation may be rough and bumpy but its potential impact to significantly improve patient outcomes and reduce provider burnout makes it worth the journey.

Mida Pezeshkian, PhD, is an member of the A.Team healthcare guild and the founder of the consultancy, STEMA_cg.