Strategic and Value-based Implementation of Generative AI in Healthcare
By Piyush Mathur, MD,FCCM,FASA,FAMIA, Co-Founder, BrainX; Associate Professor of Anesthesiology,Cleveland Clinic Lerner College of Medicine, Case Western Reserve University
The promise of generative artificial intelligence (AI) applications has sparked a wave of optimism. In the US, it’s projected that generative AI applications could automate 25% of all work tasks and boost productivity by 9% in the coming years. The proposed benefits in healthcare are particularly inspiring, aiming to tackle some of the most persistent issues, including electronic health record-related burnout, improving patient access, drug discovery, optimizing revenue cycle processes, and enhancing patient experience. This promise, coupled with substantial investments in the iterative improvement of generative AI models, implementation frameworks, and data centers, paints a hopeful picture for the future of healthcare.
Three approaches for successful AI implementation have been proposed in the recent past. A data-centric AI approach focuses on the availability of high-quality data rather than a high volume of data to build successful models. A model-centric AI approach is centered around selecting and optimizing an AI model for the highest performance. The human values-based AI approach, however, is perhaps the most reassuring. It’s based on aligning human values with the AI model performance, ensuring that ethical considerations are at the forefront of AI implementation. While none of these approaches are mutually exclusive, aligning them with the solution sought to maximize value is essential.
The process of successful implementation of generative AI solutions first begins with defining the problem and refining the targets for AI applications. This is likely to maximize the benefits of generative AI solutions rather than tackle every aspect of a problem with it. The next step is to identify and verify the need to pair a generative AI solution with the defined problem and evaluate alternative methods. Many of the commonly used AI solutions, such as deep learning models or machine learning models, have significant implementation experience. These might also provide higher performance at a lower cost for certain solutions. So, ask yourself, do you really need generative AI for every AI solution?
It has been said that AI is the new electricity. Especially for healthcare, we need to build AI utility companies to harness the power of Generative AI.
High-quality data is essential to the development of a successful model. The volume of data alone does not overcome the issue of model performance and can be a major stumbling block. In healthcare, labeled data comes at a significant cost. Various methods that help with automated or semi-automated data labeling methods are in use to overcome this issue. Another issue with healthcare data to consider is its bias. Multicenter and large datasets have been used to decrease bias. The value of having multidisciplinary teams that focus on viable solutions to evaluate for biases in the data and develop solutions to overcome bias is of immense importance.
With the proliferation of open-source high-parameter models over the last few years, there are plenty of generative AI models to choose from. Which model to pick is an important decision. To facilitate this decision, it is important to understand the architecture of the model itself, the data it was trained on, its performance on various tasks, and its ease of implementation. Having a team that has expertise in reviewing these aspects of the generative AI model is essential to aligning it with the proposed solution. There are many leaderboards from popular websites, such as Huggingface, that can be utilized for this decision-making. To develop one’s own model is not a trivial task. Beyond the data and expertise, access to expensive and in-demand high-performance computational resources involving many graphic processing units (GPU) is essential. These costs can run into millions of dollars to build a single model and take days to build. Can the existing models, especially the open-source ones, be then modified to fit one’s needs? The answer is possibly yes. There are many techniques to fine-tune or adapt the existing multi-billion-parameter generative AI models to fit one’s needs at a much lower cost and with fewer resources. All these are critical decisions that play an important strategic role in finding a cost-effective generative AI solution.
Most of the generative AI solutions need to be implemented via electronic health records or device interfaces in healthcare. The efficiency and effectiveness of the frameworks or application programming interfaces (APIs) that can deliver generative AI output to the right person, at the right time, and in the right format, maximizes the opportunity for it to be utilized. Many studies in healthcare have shown that suboptimal or ineffective human computer interfaces have limited clinician adoption of high-performing solutions. This last mile of human-values alignment to deliver the desired solution is probably the most under-resourced, though probably the most important factor in determining success.
ChatGPT ushered in an era of excitement and a race to develop generative AI solutions. However, the race is not won until the value proposition of these models is established for everyday use. This requires strategic thinking and development over the next few years. Healthcare has been the leading area of research and venture capital funding for AI solutions over the past decade. Can generative AI be the solution that leads to the successful implementation and integration of AI in healthcare? It has been said that AI is the new electricity. Especially for healthcare, we need to build AI utility companies to harness the power of generative AI. Your strategic thinking and development efforts are crucial in this journey.