AI in healthcare
An autopilot for physicians
Franz Pfister realized early that AI could improve healthcare. So he left his position as a medical resident to study Data Science at LMU. A student project led to a diagnostic algorithm, which is now undergoing its first practical test.
For most people, the study of a subject precedes the application of what one has learned in an appropriate setting. Franz Pfister’s career took a less conventional turn. On completing his medical degree at LMU six years ago at the age of 25, he worked in the neurology section of the Schön Clinic in Munich, specializing in Parkinson’s and stroke patients. He soon discovered that, in both clinical medicine and the healthcare system as a whole, standard procedures are far from being optimally structured. “At that time, as a resident physician, my hands were tied,” he recalls. The idea of developing feasible business models for process optimization and quality enhancement had taken root in his thinking. As luck would have it, a research project on machine learning had just been started in the hospital. “And that made me realize that the idea could potentially be implemented.” But there was a problem: He lacked the required training in the field of artificial intelligence.
A short time later, Pfister (31) began to take online further-education courses to gain a grounding in informatics. “I was immediately hooked,” he remembers. He then sought a way to extend his grasp of the subject, and came across LMU’s new Master’s course in Data Science. He applied for a place – in spite of his conviction that a young physician without a background in advanced mathematics wouldn’t even get as far as the interview. However, LMU not only invited him to an interview, he made such a good impression that he was admitted to the course. So Pfister gave up his job in the clinic and became a student again. “The course not only taught me the basics of data science, it provided me with a deep understanding of the subject,” he says. One of the projects in which he was involved at that time explored how machine learning could be utilized in medicine. This ultimately gave rise to deepc, a concept for which Pfister and his six colleagues won First Prize in the annual competition for the best idea to emerge from the Students’ Innovation Labs set up by the Bavarian Center for Digitalization (Zentrum Digitalisierung Bayern).
Medical tests are becoming increasingly complex
Over the past 10 years, the demand for radiologists has almost doubled, but the number available has hardly budged. This means that each of these specialists now has less time to devote to the individual patient. “And the greater the pressure, the higher the likelihood of diagnostic errors,” Pfister points out. Another factor which comes into play in this context is what is known as ‘satisfaction of search’. This refers to the phenomenon that leads a busy physician to overlook the indications of further clinical abnormalities in a scan once he has discovered one. – Pfister illustrates this type of lapse by remarking that “the patient could even have fleas and lice!” Furthermore, not only must each doctor carry out more diagnostic procedures, the tests themselves have also become much more complex. Modern magnetic resonance tomography generates 2000 high-resolution images per patient. The sheer volume of data collected makes errors in analysis unavoidable.
Some physicians regard any reference to such factors as an imputation of incompetence, and react by adopting a defensive strategy. However, Pfister is not at all interested in apportioning blame. Diagnostic errors that cost lives can never be entirely ruled out. “However, as a rule, physicians are highly skilled in the interpretation of scans and other test results,” he emphasizes. The goal of his AI-based work is to provide a kind of autopilot, which can serve as a reliable back-up in the event of lapses in the consultant’s concentration. He compares its function to that of the spelling checker in a text processing program. We all overlook misprints in our own documents, but the program clearly signals their presence with a wavy red line. “And deepc provides such an underlining feature for physicians. In this way, it helps to ensure that diagnostician will not fail to register any relevant clinical indicators.”
Improving diagnostics, streamlining procedures
In essence, deepc makes use of algorithms and models drawn from machine learning to detect anomalies in medical data. These are then flagged for the physician, who can then make a more objective and more reliable diagnosis. The software is designed to analyze data generated by the various types of imaging scanners routinely used in clinical settings. When deepc comes across an apparent anomaly, it alerts the consultant physician, who can then review the feature and come to an informed decision as to its clinical significance. If nothing abnormal is identified, after a quick check of the data, the doctor can send the patient home without further delay. In other words, deepc is not intended to replace the diagnostic skills of the trained professional, but to enable these skills to be applied more efficiently. The program is now being tested in the Neuroradiology Department of the Medical Center attached to the Technical University of Munich (TUM). “This marks the realization of the goal that I set for myself in my early years in clinical practice,” says Pfister. “We are increasing the quality of diagnostic procedures and making them more efficient.”
This is not to say that he believes that deepc will make the diagnostic skills of professional doctors obsolete. “I remain dedicated to medicine, and I have no intention of either making the medical profession superfluous or betraying its ideals.” The demand for diagnosticians will always exceed the available supply. However, AI will inevitably alter the everyday practice of radiologists and other medical specialists. – But that in itself is nothing new. “For example, 50 years ago, there were no MRI scanners, and cases of stroke were diagnosed by clinicians.” AI will find its niche in medicine in the same way, he avers, which does not mean that clinicians will become superfluous any time soon.
Pfister is prepared to make his own contribution to the integration of machine learning into medical practice. As a trained physician, data scientist and entrepreneur, he is involved in several other start-ups apart from deepc. He is developing wearable monitoring devices that can detect early signs of Parkinson’s disease. Another project is devoted to finding ways to use the blockchain concept as a basis for the secure digital exchange of patients’ files. This would permit these data to be used for research purposes without infringing data protection provisions, and make it possible to extend the range of application of future AI systems. Why do all of his business ideas have to do with the sphere of medicine? “The application of AI to medicine can save lives. That makes far more sense than using it to find ways of increasing click rates for marketing purposes!”, he replies.