Interview: Clinical data are essential

Artificial intelligence (AI) offers a great opportunity in establishing causality. "Although a clear assignment of cause and effect is not always possible using AI, it can also help today to explore connections between symptoms and disease," says Ralph Steidl, CEO of Portabiles Healthcare Technologies, which was founded in 2016. In this way, he says, it can fertilize the development of "conventional" mathematics-based algorithms that enable greater traceability of results. Co-founders of the young company are Professor Björn Eskofier, Chair of Machine Learning and Data Analytics at the University of Erlangen-Nuremberg, and Professor Jochen Klucken, neurologist and Parkinson's expert at Erlangen University Hospital and medical director of the Medical Valley Digital Health Application Center. In an interdisciplinary collaboration, they developed a gait analysis for Parkinson's patients, the Mobile GaitLab. "We are developing a system to objectively measure the effect of a therapy on patients with Parkinson's syndrome or other movement disorders using motion sensors integrated into shoes," Steidl explains. Since patients can wear the shoes all day in their daily lives, he says, the attending physician has the opportunity to check the patient's ability to move on a daily basis. "This allows him to adapt his therapy to the patient individually in a timely and targeted manner." The development is to be supported by the use of AI. It is already possible today to calculate highly precise gait parameters using deep learning algorithms. However, AI can do even more: "By using artificial intelligence, we want to predict the course of the disease to some extent from the gait pattern and calculate, for example, how high the patient's individual risk of falling is," says Steidl.

Marc Batschkus also sees the added value of AI in prognostics using sensors. The physician has specialized in medical information processing and is also active as a consultant. "Small devices will be able to give chronically ill patients in particular better indications of their progress and any interventions that may be necessary," Batschkus says. The focus here, however, is on sensors, rather than AI, he said. "However, this data will be used for big data collection and predictions will be derived from it." Numerous suppliers to software and sensor technology will also showcase this interplay of hardware and software as the basis of future implementations of AI at Medtec Live in Nuremberg, Germany, May 21-23, 2019. "AI is one of the top topics both at the exhibition and in the supporting program, as well as at the Medtech Summit taking place in parallel," says Alexander Stein, Director Medtec Live at Messe Nürnberg.

Siemens Healthineers is also focusing on the possibilities of AI: already more than 40 AI-based applications are integrated in its products. One example is the Fast3D Camera of the Somatom go computed tomography scanners. "Supported by AI and deep learning technologies, the camera enables automatic, precise isocentric positioning of patients," explains Jörg Aumüller, Head of Digitizing Healthcare Marketing at Siemens Healthineers. According to studies, there is potential for improvement in over 90 percent of images with regard to patient positioning. But even small differences in patient positioning can be significant. "Even a few centimeters deviation from the ideal position can have a negative impact on the X-ray dose as well as the image quality," says Aumüller. For this purpose, the Somatom go computed tomographs are equipped with a 3D infrared camera based on AI and deep learning technologies. Already at the beginning of the diagnostic chain, unwanted deviations could thus be reduced and potential repeat scans avoided.

AIs also have to learn

For the systems to be able to perform such tasks at all, they must first be taught. "In machine learning, a human provides data to an algorithm," explains Dr. Matthias Weidler of Astrum IT, a Medical Valley-based service provider that offers both IT consulting and software engineering. In the training phase, for example, images of healthy and broken bones are shown. "The goal is to use the data to teach the system what belongs to which class," Weidler said. To do this, it is important that the data represent the entire diversity and are of high quality.

However, his colleague Dr. Jan Paulus points out that there are still certain limits: "AI is still a long way from being able to process rules itself. It has no consciousness, no creativity and is not yet capable of recognizing why something is the way it is." Weidler adds, "Anything the AI hasn't been trained on, it can't match." Therefore, a large amount of training data is essential. "To get the required amount of training data, clinical collaborations are essential," Aumüller cautions.

Healthineers is therefore focusing on international collaboration. In this way, the AI can also be trained across continents and cover different populations using merged data.  Portabiles has also put a lot of work into training its AI, which is still in the process of being developed: "We have performed more than 2,000 clinically stratified gait analyses, primarily with Parkinson's patients, and thus have a large data pool," reports CEO Ralph Steidl. This is growing additionally due to three clinical studies that are currently underway.

Errors are preprogrammed

In Europe in particular, the extraction of data involves comparatively more effort than in other countries due to the strict data protection regulation. China in particular is currently considered a pioneer in the field of AI. Medical expert Batschkus sees a further problem in the fact that a lot of data alone does not yet lead to the desired result. "The concept and methodology of AI are inextricably linked to Big Data, as large amounts of data are needed to create or train AI systems."

With Big Data, however, there are numerous methodological weaknesses and a hasty euphoria along the lines of "more data means more insight." In some cases, work is done here without any theoretical approach, which means that errors and wrong decisions are already pre-programmed during development. In fact, a certain know-how is a mandatory prerequisite for learning AI: "Often, the mistakes are not made maliciously, but because required knowledge is simply not available," notes Paul. To remedy such cases, Astrum, for example, offers supportive consulting. "This goes so far as to accompany the entire development process," concludes Weidler.