AI in Medicine: Emerging Potential
Computer assistance in medicine has a long history, and with the impressive breakthroughs in artificial intelligence in recent years, the technology has certainly moved beyond purely academic use. Expectations for innovations are now enormous: Deep Learning applications show impressive accuracy in image recognition, AlphaFold revolutionized protein structure prediction, and recently, generative AI approaches like Large Language Models (LLMs) such as ChatGPT have gained widespread adoption in consumer applications. LLMs now truly demonstrate impressive results in research studies, initially in language-based tasks and increasingly in other fields such as the combination of image and text content.
What does this mean for real-world medical applications? These technological developments have not yet reached everyday use on a broad scale, and the gap between possibilities and actual value creation is growing. Instead, public perception is shaped by extremes: utopian visions of a technological revolution on one hand, and fear scenarios particularly in the areas of labor market, data protection, and ethics on the other.
Concrete Added Value Comes into Focus
The discussion is now shifting towards the actual value AI technologies can generate for medicine. Which clinical processes can benefit from it, and in which application fields can genuinely disruptive changes be achieved? In this article, I want to focus on the present and highlight the particular added value AI-based medical products provide in practice today. Based on exemplary success stories, I provide insights for the field that are also relevant for future developments and in the broader area.
Of course, the question of success factors has multiple dimensions. This article will not cover the equally critical aspects of development, including data collection and quality, clinical implementation, or market monitoring, but will focus on product positioning, market orientation, and especially clinical value generation.
Practical Applications in Imaging
I will exemplify the field of diagnostic imaging here as it holds a pioneering position. Naturally, there are also very successful applications in other areas such as pharmaceutical research, patient selection for clinical trials, or operational process optimization. However, the focus here will be on AI-based medical products. The technical foundations of the products that have been successful in the market so far (e.g., deep convolutional neural networks) differ from the new wave of generative AI applications (especially LLMs). Nevertheless, the insights from these can also be applied to new generative AI applications.
The vast majority of AI-based medical products approved for the American market come from the field of radiology (671 out of 882 FDA-listed approvals as of March 2024, FDA AI Devices). According to billing data, the most used applications are attributed to cardiology, ophthalmology, and radiology (Wu et al. 2023). In Europe, radiological applications for AI are likely to be even more prominent, although corresponding figures are not yet available.
Meanwhile, there is an overwhelming number of AI-based imaging products on the market, as each imaging modality and diagnostic question requires its own model and thus its own approval. For example, the Health AI Register lists over 200 CE-marked AI products for radiology at the time of publication of this article. These applications primarily show efficiency or quality improvements in radiological workflows, typically through anomaly detection (e.g., fractures in X-rays or lung nodules in CT), quantification (e.g., brain atrophy), and image enhancement (e.g., for accelerated MRI protocols). The extent to which these products are implemented or used is not publicly known, but overall, there is a rather slow adoption in the market, as was evident in presentations and discussions at the recent European Congress of Radiology (ECR).
Examples of Particularly Outstanding Applications
Some applications stand out as particularly effective from the large field. We will look at three positive examples below and discuss their peculiarities, which are also relevant for other areas of medicine:
Replacement of Invasive Vascular Function Tests
The most common AI application in the USA based on billing data involves the quantitative analysis of coronary arteries. Instead of an invasive pressure wire measurement performed during a coronary angiography, the so-called FFRCT (Fractional Flow Reserve Computed Tomography) allows a non-invasive assessment. By modeling the vessels and simulating blood flow, the effect of vascular constrictions can be determined, and the necessity of therapeutic interventions such as angioplasty or bypass surgery can be evaluated. The quantitative analysis of the CT images thus significantly enhances their diagnostic value, potentially eliminating the need for invasive diagnostics.
Screening
As mentioned earlier, there are specific AI products for most imaging modalities and many diagnostic questions. Many applications provide support for parts of the diagnosis, for example, by detecting some of the most important abnormalities in lung X-rays. However, there are already applications in the field of screening that can (partially) autonomously solve specific diagnostic questions. For example, in the USA, a system has been approved that autonomously detects diabetic retinopathy in diabetes patients, enabling early further treatment by specialists. Successful automation is also evident in Europe: In initial results from a randomized study in Sweden (Lång et al. 2023), a 44% reduction in workload was shown compared to standard double-reading, with at least equal quality. Such a replacement of a diagnosing person has already arrived in routine practice (Karolinska News).
Triage in Stroke Diagnosis
A particularly widespread AI application is the prioritization of critical cases in the emergency room. Especially in the field of stroke diagnosis, systems are used to detect ischemia (reduced blood flow) and thereby accelerate the possible minimally invasive treatment process. Such applications have recently been preliminarily recommended for use by the British National Institute for Health and Care Excellence (NICE) of the NHS. Speeding up the treatment process is particularly important in this application case because promptly initiated thrombolysis can reduce brain damage and increase the likelihood of subsequent symptom-free recovery.
Success Factors
What sets these examples apart from the market of AI applications? What insights can be gained for one’s own innovation projects? Three aspects, in particular, stand out that pose specific requirements for the selection of applications, development processes, and marketing.
Focus on Clinical Effect Beyond Imaging
First, all applications stand out due to a very clear clinical orientation. This may seem self-evident. Unfortunately, in the research context, it is more common to choose a field of application based on technical possibilities and data availability. However, when development and validation are driven by the larger treatment context and a clear focus on the overall clinical benefit, the achievable added value is generally significantly greater.
This particularly enables greater value creation beyond the financial limitations of diagnostics and allows for other reimbursement models than pure efficiency considerations within diagnostic workflows. For example, direct effects on the further course of treatment are possible in the mentioned examples. Of course, this also requires consideration of the respective clinical and reimbursement guidelines.
Holistic View of the User Experience
Second, all affected persons must be taken into account. Unlike traditional product design, AI applications often require a broader context. Instead of the specific user experience of the software for radiological personnel, a holistic view must now be considered. For example, in the application case of stroke prioritization, communication with the treatment team via a mobile app with a chat function is a critical factor for the overall benefit. Specific challenges of AI systems, such as “automation bias,” must also be considered. Reliable technologies can lead to people preferring or not sufficiently critically adopting decisions from automated systems. This must be realistically included in the concept of application practice and profitability considerations, as in the case of mammography screening.
Value-Based Marketing
Third, the mentioned successful applications are characterized by the fact that clinical validation is at the center of product marketing. In fact, it is not technical metrics like sensitivity and specificity that lead to high trust in the value of innovations but clinical results and testimonials. This often poses a particular challenge for product teams, and clinical collaborations are often considered only after product development. However, organizations that have strong clinical connections and achieve thought leadership in the medical field are much better positioned here. It is possible, even in early development phases, to generate interest and demand for new technologies through suitable publications and events. Since building visibility and trust, especially in medicine, cannot be achieved with short-term activities, strategic planning and early initiative are crucial.
Conclusion
Taken together, these examples of particularly outstanding AI-based medical products are characterized by a holistic view of the clinical context. For new developments, the following questions should be asked:
- What benefits can the product achieve in the entire treatment process?
- How should, in addition to the users of the system, all affected persons be considered in the design?
- How can the value be credibly communicated to stand out from the mass of product innovations?
These aspects should be considered at all phases of the innovation process, especially at the beginning, to prepare for market success. This requires a strategic approach and an open and realistic view.
What other aspects do you see as critical success factors? I look forward to engaging with the community and to hearing your exciting opinions, additions, and questions.
0 Comments