Coronary artery disease (CAD) often presents with narrowed arteries that restrict adequate oxygen supply to the heart, a condition referred to as stenoses. Stents can provide relief here. These are optimized medical implants that serve to dilate the narrowed segment in the artery, thereby restoring optimal blood flow. Stents, resembling diminutive tubular wire meshes, have proven effective in most cases. However, there are some cases in which the body’s response deviates from the expected course. The part of the artery that was opened using a stent becomes narrow again. The question that physicians and the medical community are trying to answer is: What causes these unusual cases? Finding partial answers could already provide medical knowledge that might help guide physicians in making better treatment decisions. In collaboration with a medical technology company researchers at Fraunhofer IKS are using artificial intelligence (AI) to investigate and understand the complex factors behind theses exceptional cases.
Stents under the magnifying glass
Basis of the investigation was a modern stent designed with a unique attribute: Not only does it serve to keep constricted vessels open, thus ensuring their proper functionality, but it also self-dissolves after a certain amount of time. A lot of data has already been collected for such stents in medical studies. It shows that here too, sporadic complications occurred, which are summarized by the medical term "post-interventional target lesion failure (TLF)."
The challenge at hand was to unravel the reasons behind how and why these complications occur. What can be learned from them? And how can such complications be avoided in future products? "Study data from about 2,000 patients with relevant measurements were made available," explains Narges Ahmidi, department head of Reasoned AI Decisions at the Fraunhofer Institute for Cognitive Systems IKS. "On this data, we used artificial intelligence to play Sherlock Holmes and do some detective work, if you will." That's because the power of AI extends not only to supporting and augmenting tasks within human capabilities — such as autonomous driving— but also, as in this instance, to addressing challenges and problems that push human limitations. "Our first step was to understand the medical background, so we could evaluate appropriate AI models. For example, the necessity of predilating the vessel with a balloon before the actual stent could be inserted."
With the help of AI, the researchers could ideally help physicians find the optimal stent for patients with CAD, minimizing the complication rate. This challenge is without a solution up to date. To achieve this, comprehensive patient data on various stents would be required. "However, we were able to achieve something significant. Specifically, we investigated the theoretical compatibility of the novel stent and analyzed ways to improve the surgical guideline. We intend to publish the information we have gathered to contribute to the discussion within the medical community" Ahmidi explains.
Predicting whether patients will experience complications
Ahmidi's research team separated the task into two project parts. The first part covered attempting to predict, at the time a patient is discharged from the hospital with an implanted stent, whether complications will arise or not? If so, physicians could identify high-risk patients early on and follow up on them more closely. To answer this question, the researchers fed the AI with data.
The datasets included both the patients' health backgrounds and information about the surgery and treatment. "While general risk factors for complications are known, such as diabetes, obesity, and multiple heart problems, these only indicate general trends. But what does that mean for the individual patient? How the group of patients who remained healthy differs from the group that developed complications is not necessarily obvious to a human brain," Ahmidi says. However, AI has the advantage of being able to read data faster than any human and efficiently search for high-dimensional correlations. The algorithm provides physicians with a "yes-or-no" answer, such as: This patient is likely to develop complications. The evaluation found that the AI's prediction was reliable. The algorithm successfully classified 1635 patients as low risk, with a 90% concurrence with reality. Additionally, over half of the patients with complications were correctly identified. The team published detailed results in the journal Applied Sciences by the publisher MDPI.
Optimization: Searching for an enhanced surgical guideline
In the second subproject, the researchers addressed the question: Are there parameters that could be adjusted during surgery to reduce the risk of later complications? While physicians do have many hypotheses on how to enhance the well-being of their patients, it is not feasible for them to test all of them. The only way available is to subject a hypothesis to a clinical trial — an expensive and time-consuming process that starts anew for each hypothesis. "What we offer at Fraunhofer IKS is this: we take the current operational data that is already available" says Ahmidi, "and reject non-applicable medical hypotheses with the help of AI. This includes aspects like reducing the inflation pressure in the balloon used to widen the narrowed vessel. In this way, the team can contribute to making medical research more efficient." The results of the second subproject are currently being published.
Thus, the researchers at Fraunhofer IKS were able to demonstrate that AI can be used to tackle complex medical problems — such as early detection of future adverse events in various medical scenarios and the search for hidden underlying mechanisms.