The Double-Edged Sword of AI Double-Edged Sword
( continued from page 7)
Since the first recording of human EEG by Hans Berger in 1924, there has been steady progress in every aspect of the techniques of recording and the interpretation of brain electric activity. During my early career, only analog, ink-pen recording was possible with eight( and later 16) channels on fan-folded paper, usually about 200-300 pages long per patient for a routine study. The art and science of paper EEG interpretation involved combining the skill of visual pattern recognition, frequency analysis and speed reading, coupled with an abundant knowledge base gathered over time. The neurology residents did rotations in EEG lab, but an additional year of fellowship was necessary to assimilate the necessary skill for EEG interpretation. Traditionally the new trainee had to be ready with a handwritten report of each EEG done that day, to present to the attending. The EEG attendings were usually tough teachers and I still recall the unique one who carried in his coat pockets two rubber stamps to affix on the fellow’ s EEG interpretation report based on the accuracy: one said“ OK” and the other“ B … S.”
The arrival of digital EEG recording and the ability to store and retrieve voluminous data revolutionized the practice of electroencephalography. Finally, there was no more manually turning pages after pages of EEG stacks. Our new ability to study the wave forms using different filters and montages led to more precise localization of abnormal waveforms! The sheer volume of data generated after continuous monitoring of EEG for several days in a seizure monitoring facility is challenging and often tiresome( usually a weekend chore) even to the experienced electroencephalographer. We all dreamed of the day when the analysis of such mammoth data could be automated and interpreted as soon as the EEG data was collected, with abnormal patterns highlighted and kept ready for verification by the electroencephalographer.
Our attempt to imitate the neural networks in the brain and thereby develop algorithms that simulate human intelligence( AI) has been in the works for many decades. The use of deep learning algorithms that permit self-training through multilayered neural networks led to models suitable for large volume data and the need for longer training time using supervised, unsupervised as well as reinforcement components. 1, 2 During the last two decades, significant progress has been made in EEG analysis using AI, leading to FDA clearance for several systems; many are already commercially available. One of the first such systems was trained on > 30,000 EEGs and was found to be capable of human expert-level performance for routine clinical EEG. Another system of AI-enabled EEG analysis apparently can detect abnormalities in data from routine studies as well as long term monitoring and ambulatory EEG, again at the expert level. Detection of electrographic status epilepticus is the goal of AI powered EEG in several platforms and has been found to be of great value in neonatal seizure monitoring. Considerable research is being done with some success to unravel EEG patterns that may eventually prove to be biomarkers of early cognitive decline. There is much hype about AI-powered ambulatory EEG analysis with easy to use“ dry” electrodes and wireless headsets opening a new era of brain-computer interphase. AI-assisted sleep monitoring is another area which has seen significant progress and practical application.
The American Medical Association has suggested the term“ augmented intelligence” to focus upon artificial intelligence’ s assistive role, emphasizing the fact that AI design enhances human intelligence rather than replace it. 3. This happens to be quite true regarding EEG interpretation by neurologists, considering the extreme complexity of patterns in different age groups, from premature neonate to the centenarian with a vast array of potential brain disorders. Appropriate use of AI can reduce human error and make the process quicker and more efficient and above all more accurate.
While there is healthy and much needed debate regarding the“ good, bad and ugly” aspects of AI, the medical community must learn to utilize and apply the new technology when and where it is appropriate. Let me end with an apt quote from one of my favorite authors, Isaac Asimov:“ If knowledge can create problems, it is not through ignorance that we can solve them.”
References:
1
Kaur T. et. al. Artificial intelligence in epilepsy. Neurol India 2021; 69:560-566
2
Poothrikovil et. al. Evolution of artificial intelligence-assisted EEG technology. The Neurodiagnostic journal 2025; 65( 3): 169-172
3 https:// www. ama-assn. org / practice-management / digital-health / augmented-intelligence-medicine
Dr. Iyer practices at the Neurodiagnostic Center of Louisville and is a retired professor of neurology at the University of Louisville School of Medicine.
8 LOUISVILLE MEDICINE