AI Lectures: LMU medical expert expects AI to lead to better prognoses, but more validation work is needed

Enlarge

Artificial intelligence has become hugely more important in medicine. This is mainly due to the availability of ever increasing amounts of patient data. “Genetics and imaging are allowing us to gain more and more detailed insights into diseases,” explains Professor Nikolaos Koutsouleris, who practices as a specialist in Psychiatry and Psychotherapy, speaking in the LMU’s second virtual AI Lecture. Despite this, he says, humans are often unable to properly process this much information because of limiting factors such as personnel or time availability, cognitive capacity and financial factors. The result of that is misdiagnosis.

AI can help clinical staff make the right diagnostic decisions and identify the optimal treatment options for patients. By way of example, Koutsouleris cites a study of AI-based breast cancer screening that looked at clinicians and patients from the UK and the United States. The study demonstrated that AI was able to provide almost better prognoses than when two clinicians screened these patients. “When only one clinician screened the patients, AI was actually superior,” states Koutsouleris. “Furthermore, the technology can serve as a kind of early warning system for affected individuals to seek medical treatment in a timely manner. That can save lives.”

Numerous studies on AI in medicine lack validation

The use of AI in medicine still has a long way to go, though. While there is no doubt that regulators in the US are approving more and more AI products, “the quality of the studies that are developing AI models has sadly not increased,” Koutsouleris points out. “Added to that, some 95 percent of AI models in medicine have not been externally validated, so the therapeutic benefit for patients has not been confirmed to the highest scientific standards.” As a result, currently only a fraction of all solutions for disease prognosis are actually usable in practice.

Importantly, the trained AI models must be able to be generalized and interpreted to reduce the risks for patients. AI currently manages to classify patients into the correct risk category in around three quarters of cases. When human intelligence is added in, the right decision can be made in as many as 86 percent of cases. But that is not yet enough for AI to be used in practical applications, because for time, financial and ethical reasons, not every person can be examined in advance as comprehensively as they are in some of the studies being done.

Researchers are therefore working hard to make the data and models more generally applicable so that they can make better disease prognoses. “If algorithm development continues at the present rate, I can see AI being heavily involved in medical decision-making processes in ten to 15 years,” says Koutsouleris.

AI Lectures: For more information, see the program