Foundation Models in Healthcare: What You Need to Know

Picture a typical Monday morning in a busy hospital: Dr. Johnson navigates patient data scattered across multiple systems, making it difficult to obtain a complete, accurate view for decision-making. This scenario is not unique; healthcare providers everywhere are dealing with fragmented information, increasing pressure to boost efficiency, reduce costs, and improve patient outcomes. As these challenges become more urgent, executives are searching for smarter, more integrated solutions. Foundation models, or advanced AI systems, have emerged as a promising solution. They are rapidly gaining traction in healthcare for their ability to streamline clinical, operational, and patient-facing tasks.

What Is a Foundation Model?

A foundation model is a large-scale AI system trained on broad datasets to perform a variety of tasks without task-specific training. Unlike traditional models designed for a single purpose, foundation models learn patterns across multiple domains, enabling them to generalize knowledge. In practice, they are the engines behind advanced chatbots, medical imaging analysis, predictive analytics, and more.

Why Are They Gaining Popularity Now?

• Increased computational power: Cloud and on-premises resources now enable not only training but also the rapid deployment of large models. Just a decade ago, limitations restricted AI developments. For instance, training that once took weeks and was costly can now be completed in just days, making the broad application of AI in healthcare more feasible and transformative.
• Availability of massive datasets: More structured and unstructured health data allow models to learn nuanced patterns.
• Versatility: Foundation models can perform multiple tasks, from summarizing patient records to predicting disease risk.

Because of these factors, foundation models are an appealing choice for health systems looking to deploy AI across different parts of their technology. For instance, one high-impact use case is readmission prediction; AI models analyze patient data to identify patients at risk of readmission. This can significantly reduce readmission rates and improve patient care. Another practical application is workflow automation, where foundation models streamline administrative tasks, freeing up healthcare professionals to focus more on patient interaction.

How a Foundation Model Fits into Your Health System’s Tech Stack

Foundation models are not ready to use out of the box. Instead, they are tools that add to your current systems. You can connect them to electronic health records (EHRs), clinical decision support, or patient engagement platforms. When you do this, they offer: 

• Automated insights from large volumes of data
• Natural language interfaces for clinician and patient interaction
• Predictive modeling to support population health initiatives

The main point is: Use foundation models carefully. They should work alongside your current technology, not replace it. To ensure a smooth integration, start by assessing your current tech and identifying where foundation models can add the most value.

1. Collaborate with IT teams to understand existing systems and determine where integration is possible.

2. Discuss data flow and make sure data lifecycle management aligns with the use of foundation models.

3. Evaluate data storage to ensure compliance with regulations like HIPAA.

4. Develop clear use cases with measurable outcomes to guide the deployment of foundation models.

Risks with Foundation Models

Even though foundation models have significant potential, they also pose risks: sharing sensitive health data with third-party servers can create compliance issues. For example, in 2022, a major healthcare provider suffered a breach where PHI was exposed to unauthorized parties due to insufficient data encryption protocols. This incident not only compromised patient privacy but also led to legal and financial repercussions.

To mitigate these risks, health systems should: Conduct thorough vendor vetting to ensure that any third-party partners meet strict security and compliance standards. Implement advanced encryption standards to protect data both at rest and in transit. Establish a clear data management policy that regularly audits data usage and access to prevent unauthorized exposure.

• Accuracy and bias: Models trained on general data may produce unreliable or biased outputs in specialized clinical contexts.
• Lack of transparency: Complex models can behave unpredictably, making it hard to understand how decisions are made.

How to Govern a Foundation Model

Effective governance starts with establishing policies around usage, monitoring, and accountability:

• Define approved use cases and limitations, with the Chief Medical Information Officer (CMIO) ensuring clinical appropriateness and ethical standards.
• Monitor outputs continuously for bias, drift, and accuracy, under the purview of the Data Science Team, with the Chief Data Officer (CDO) overseeing adherence to performance benchmarks.
• Track where data is sent and stored to ensure compliance with HIPAA and other regulations, a responsibility of the IT Security Officer.
• Maintain an audit trail for model decisions that influence patient care, with reviews conducted by the Compliance Department, led by the Chief Compliance Officer (CCO).

Best Practices When Implementing a Foundation Model

When incorporating a foundation model into a health system:

• Start small: Implementing a pilot program can minimize risks and facilitate learning before wider deployment. Consider a pilot project like auto-drafting discharge summaries. Set a 3-month timeline to evaluate the pilot, aiming to reduce time spent on manual summary drafting by 25%. Track metrics like time saved, error reduction, and clinician satisfaction.
• Choose the right hosting: Decide whether the model will run on-premises, in a private cloud, or via a trusted vendor.
• Fine-tune for healthcare: General models like ChatGPT can help with basic tasks, but they often need to be tuned to ensure they are accurate for clinical use.
• Secure data: Ensure that any prompts or inputs remain within secure systems when patient privacy could be affected.

Specialized Use Cases: Can ChatGPT Cut It?

General-purpose models like ChatGPT can help with administrative or patient engagement tasks, but they are not always reliable for clinical work. For specialized uses like diagnosis support, imaging analysis, or predictive modeling, models need to be trained or adjusted on your historical patient data, then carefully validated. Establishing safety nets like human review and ensemble checks can reinforce trust. Incorporating validation layers ensures that models perform accurately and safely before their outputs are used in critical clinical decisions. If not, their results could be wrong, incomplete, or unsafe.

Know Where Your Data Goes

One important but often missed question is who hosts the model and where your data goes. Many SaaS solutions send data to outside servers. In healthcare, this can lead to privacy and compliance issues. Health systems should look at:

• Hosting location: On-premises vs. cloud, and cloud provider security standards.
• Data usage: Are prompts logged, stored, or used to further train the vendor’s model?
• Regulatory compliance: Does the setup meet HIPAA, GDPR, and other relevant frameworks?

Conclusion

Foundation models give healthcare new ways to improve operations and patient care, but they are not a cure-all. Health systems should set rules, customize them for clinical needs, and keep data safe. Foundation models can become a key part of a modern, AI-powered health system.

Our advice? Here’s where to start:

1. Create a task force with IT, compliance, and clinical leaders to oversee the integration of foundation models. This group can provide strategic direction and ensure all perspectives are aligned.

2. Start a readiness assessment to evaluate current capabilities, identify gaps, and outline the necessary steps for seamless integration.

3. Develop a roadmap with clear timelines, objectives, and metrics for success to guide the deployment process effectively.

‍