Researchers are studying causal machine learning, a new advancement in AI in healthcare

Artificial intelligence is also advancing in the medical field. When it comes to imaging technology and health risk calculations, there are a number of AI methods in development and testing stages. Machines are expected to greatly benefit humanity when it comes to recognizing patterns in large amounts of data. Following the classical model, AI compares information to learned examples, draws conclusions, and makes inferences.

Currently, an international team led by Professor Stefan Voeriger, director of LMU's Institute for Artificial Intelligence (AI) Management, is exploring the potential of a relatively new field of AI for diagnosis and treatment. Can causal machine learning (ML) estimate treatment outcomes better than previously commonly used ML techniques? Yes, a study by a group says. natural medicine The title is “Causal ML can improve treatment efficacy and safety.''

In particular, new ML variants offer “rich opportunities to personalize treatment strategies and individually improve patient health,” the researchers from Munich, Cambridge (UK) and Boston (US) wrote. There is. Stefan Bauer, professor of computer science at the Technical University of Munich (TUM) and group leader of Helmholtz AI, and his Niki Kilbertus.

Regarding machine assistance in treatment decisions, the authors expect decisive advances in quality. They claim that classical ML recognizes patterns and discovers correlations. However, the causal principle of cause and effect remains, in principle, closed to machines. They cannot address the question of why. However, many of the questions that arise when making treatment decisions involve issues of causation.

The authors illustrate this with the example of diabetes. Classical ML aims to predict the likelihood of disease in a given patient with various risk factors. Causal ML ideally makes it possible to answer how the risk changes if a patient takes an antidiabetic drug. That is, evaluate the effect of the cause (medication prescription). For example, it is possible to estimate whether another treatment plan is better than the commonly prescribed drug metformin.

However, to be able to estimate a hypothetical treatment effect, “the AI ​​model needs to learn how to answer 'what if?' questions. Naturally,” says Feurigel, a doctoral candidate on Feurigel's team. says one Jonas Schweitar.

“We give the machine rules to recognize the causal structure and formalize the problem correctly,” Feurigel says. Next, the machine must learn to recognize the effects of the intervention and, so to speak, understand how real-world results are reflected in the data entered into the computer.

“The software required for causal ML methods in medicine does not exist out of the box,” Feuerriegel says. Rather, it requires “complex modeling” of each problem, with “close collaboration between AI experts and doctors.”

Like his TUM colleagues Stefan Bauer and Niki Kilbertus, Feuerriegel is also a researcher at the Munich Center for Machine Learning (MCML) and the Konrad Zuse School of Excellence in Reliable AI related to AI in healthcare, decision-making, and other topics. I am researching the problem.

In other application areas, such as marketing, causal ML research has already been in the testing phase for several years, Feurigel explains. “Our goal is to bring this technique one step closer to practice. This paper describes the direction in which things could move in the coming years.”