Can artificial intelligence recognize gender based on brain waves? Physicist Thomas Jochmann investigated this exciting question and came up with some surprising findings, which he shared with Anna Hansen, spokesperson for the AI Grid.
His research results show why it is necessary to address this topic: How can AI reliably improve medical diagnostics? And how can we ensure that algorithms make comprehensible decisions?
Billions of nerve cells work together in the brain to coordinate our sensory impressions, thoughts and feelings. 100 years ago, on July 6, 1924, the physician Hans Berger from Jena succeeded for the first time in using electroencephalography (EEG) to measure and visualize the electrical activity of nerve cell clusters as potential changes on the surface of a person's head.
Since then, the EEG has become firmly established in modern medicine and brain research and is used to diagnose diseases such as epilepsy, sleep disorders and other neurological disorders.
The medical evaluation of EEG recordings requires extensive work by specialists. They try to recognize patterns in the data records based on the recorded zigzag waves. It is assumed that much of the information remains hidden from the human eye and is not available for diagnosis. For some years now, scientists have therefore been trying to evaluate the huge amounts of data using deep learning methods (artificial intelligence) and uncover previously undiscovered information.
Analysis of brain activity can determine gender
There has been evidence of gender differences in the EEG for more than 50 years. Under certain conditions and depending on the composition of the study groups, statistical differences have been found in the EEG between men and women. Just as there are statistical differences in body size or other biological characteristics.
In 2018, a Dutch team also discovered that the gender of test subjects can be determined by analyzing brain activity. But can the difference between women's and men's brains really be read from the zigzag waves?
Scientist Thomas Jochmann has investigated the question of what type of brain activity differs between men and women. The graduate physicist is a member of the AI Grid, an initiative of the Federal Ministry of Education and Research that connects talented young scientists with AI experts from science and industry. In a research paper, the scientist recreated the Dutch experiment and discovered that the AI identifies women and men with an accuracy of 81 percent based on the EEG data. At the same time, however, he also discovered that the AI was cheating.
AI does not interpret brain waves, but heartbeats
Thomas Jochmann: "I had the aha effect as a visiting scientist at Harvard Medical School. There, together with colleagues, I investigated which parts of the EEG curves are crucial for determining gender. To do this, we developed a method that recognizes the zigzag patterns that the AI classifies as important or gender-specific. We then visualized these patterns and found that they always occur at exactly the same time as the heartbeat. The results suggest that the AI was not detecting differences in brain function, but rather exploiting differences in body build. Although the EEG is measured at the scalp, it not only records the electrical activity of the brain, but is also overlaid by the activity of the heart muscle. Obviously, there are gender differences in the propagation of electrical potentials from the heart to the scalp."
When distortions in black boxes falsify results
Neural networks are a kind of "black box" and it is often difficult to understand which calculations have led to a decision, explains the scientist and compares his study results with the "wolf-husky experiment". In this experiment, an image recognition system was trained to distinguish between huskies and wolves. The training data always showed huskies in a green environment and wolves with snow in the background. The classifier thus learned to associate snow with wolves and grass with huskies instead of recognizing the animals themselves. During the evaluation, the system was only able to distinguish between snow and grass, but not between the animals, which illustrates the problem of unbalanced data sets.
Checking AI predictions for gender bias
"Our results have also shown the importance of careful preparation and selection of data to ensure that the algorithms perform their calculations with relevant and meaningful features. It is not only through our research project that we know that gender can be a hidden influencing factor in machine learning models for the diagnosis of diseases. Therefore, AI-based predictions should always be checked for gender bias to ensure they are fair and accurate. When using large data sets, researchers should take care to thoroughly remove artifacts and use methods to detect invalid patterns early or exclude them from the outset," says Thomas Jochmann, who is researching new methods to better understand brain activity at the Institute of Biomedical Engineering and Computer Science at Ilmenau University of Technology.
Thomas Jochmann is a doctoral student in biomedical engineering at the Technische Universität Ilmenau and AI Grid member .
More information:
Here you can find the research article by Thomas Jochmann: Gender-related patterns in the electroencephalogram and their relevance for machine learning classifiers:
