Healthcare AI Solutions: Balancing Value, Privacy, Costs and Rapid Evolution
The Clinovera team has been building AI-driven solutions for clients for over 20 years – utilizing classical technologies of the past, such as machine learning, natural language processing, Bayesian models, rules-based systems, and now GenAI.
GenAI has revolutionized and dramatically accelerated solution development. Yet, there are a plethora of complexities with every solution we create today. Fortunately, these complexities are not blockers, and there are options and tradeoffs that we know how to address.
This document describes and classifies key concerns associated with the development of AI-driven solutions at a high level. The concerns can be broken into the following categories:
- Cost Control
- Performance and Quality
- Privacy and Security
- Rapid evolution of AI and Supporting Infrastructures
The diagram below represents and breaks down each category into more specificity, which we will later clarify with examples.
Commercial Models vs. Privately-Deployed Models
With respect to the usage of language models, all GenAI solutions utilize one of two strategies:
- Commercially-available AI services like Open AI services on Azure infrastructure or
- Privately-deployed AI models like LLAMA, which is often open sourced and deployed on privately-purchased VM infrastructure.
While the challenges and tradeoffs with these two approaches are largely the same, the options to address them may be somewhat different, which we detail below.
Cost Control
When you consider the expense of using AI models, it all depends on the strategy you choose. With the commercial AI services this is usually the cost per token (or 1000s of tokens), while invoking AI service API. For privately-deployed models, this is usually fractional infrastructure cost of a provisioned VM. The hardware to run an “intelligent” LLM with many billions of parameters usually requires an expensive graphics card and powerful CPUs, which could be expensive with continuous usage.
Excessive Usage Costs
After the solution has been utilized extensively by users or systems, significant costs will incur. For example, users may interact with a chatbot for a prolonged period or upload countless documents of a significant size, resulting in a major usage costs.
Runaway Costs
Runaway costs are different from the excessive usage costs in that they are associated with certain internal issues or inherent application workflow, such as potential glitches or unanticipated application behavior. For example, due to a bug in the application, the AI services were called in an infinite loop, resulting in a significant accumulation of costs.
Performance and Quality Challenges
There are instances when the AI-driven solution cannot achieve desired functional outcomes.
Slow, Delayed Processing
Slow, delayed processing may be caused by a number of individual or combined factors:
- Commercial AI service providers often throttle down API endpoint request processing when token limits are reached
- Large and complex documents or an overwhelming quantity of documents submitted for processing by the LLM
- Insufficient hardware with privately-deployed models
- Complex or poorly engineered prompts
- Too many users deploying the solution at the same time
Insufficiently “Intelligent” Solution
When the model generates poor quality, irrelevant or misleading responses, the contributing factors could be:
- Usage of a smaller, less intelligent language model
- Poorly-engineered prompts
- Poor quality data or documents in the context
Solutions That Are Too Generic
An AI solution is too generic when the response it generates doesn’t have sufficient relevance to the intended purpose of the application. This usually stems from lack of provided context or insufficient training of the model. For example, when an application designed to assist in making patient admission decisions is prompted to inform about patient medications, it simply returns a medication list, rather than indicating which medications are costly or psychotic.
Privacy and Security Concerns
Concerns related to privacy and security are usually raised when dealing with commercial AI services. The assumption is that privately-deployed models on a privately-provisioned hardware are more secure.
Protecting PHI/PII
Protected Health Information (PHI) and Personally Identifiable Information (PII)or content submitted to commercial AI services may be used for training of the model and consequently appear in responses to other users or systems. There are a number of available mechanisms and methodologies for protecting the PHI with AI services, and all of them have their own advantages and shortcomings. These rapid methodologies include deidentification of data, making appropriate arrangements with the commercial AI service providers, instantiating private LLMs and others.
Protecting Intellectual Property
Protecting IP is of high importance for many organizations. Often, it can be protected in similar ways as PHI, though the frameworks for this are much less developed.
Rapid Evolution of AI Technologies
Rapid advancements of AI is both an opportunity and a challenge. It forces organizations to constantly rethink and adapt their business and technology strategies, as well as their approaches to cost management and performance optimization. We loosely break this topic into three areas:
- Evolution of, and emergence of, new GenAI models
- Improvements in computing power, and
- Changing regulatory landscape
New and Improved GenAI Models and Technologies
GenAI models are evolving in multiple ways.
- Models are becoming more intelligent
- New versions of mainstream models are getting bigger, requiring more computing power
- Smaller but still powerful models emerge, providing opportunities to run them on less powerful hardware.
- New(er) models expose updated features, such as asynchronous functions, explainability, enhanced reasoning and deliberation, multimodal and real time interactions, specialized architectures and memory efficiency.
These advancements reflect a shift towards more intelligent, versatile, and interactive AI systems, offering new opportunities and challenges for solution developers.
Advancements in Computing Power
Advancements in computing power have a significant impact on AI solution development. Newer, more capable hardware can run bigger, smarter models with more parameters and more training data. The latest chips, like Apple’s Neural Engine or NVIDIA Jetson, let sophisticated AI run on phones, drones, and IoT devices. The challenges, however, include costs, energy pressure and sustainability concerns.
Changing Regulatory Landscape
Legal frameworks for AI in healthcare are evolving rapidly to address the unique challenges posed by integrating advanced technologies into medical settings. Most of these efforts revolve around data privacy and security, bias and fairness, and accountability and liability.
Here are examples of some key developments:
- The EU’s AI Act set to take effect in August of 2026 categorizes AI systems based on risk levels. In healthcare, many applications fall under the “high-risk” category, subjecting them to stringent requirements for transparency, accountability, and human oversight.
- In the US, the FDA regulations emphasize risk assessments, performance monitoring, and transparency in AI decision-making.
- There are US state and local laws dealing with AI risks, especially in healthcare.
Many of these regulatory developments have a profound impact on healthcare organizations and AI solutions developers, encouraging them to stay informed about regulatory changes and engaging in proactive compliance efforts.
Ready for GenAI to Transform Your Healthcare Operation?
Understanding how to balance these opportunities and challenges with your healthcare operation’s needs is exactly where the Clinovera team can help. Please fill out the contact us form to speak with our experts.