Interpreting EEG Waves
Reading an EEG is hard. This isn’t a groundbreaking statement, and any new researcher or neurology resident would whole heartedly agree. Even after delving into EEG primers, online educational courses, years of residency and any additional available materials, detecting abnormalities within long-term EEGs is akin to discriminating between different types of white paint. Is this more of a “Swiss coffee” or a “Chantilly lace”? As a researcher, misidentifying a spike or sharp wave does not have a huge impact on any individual patient, but for practicing neurologists, the cost for an inaccurate assessment is high1, 2. Failing to detect epileptiform activity can prevent timely treatment and over-reading an EEG can expose patients to potentially negative side effects of antiseizure medications unnecessarily. Thus, accuracy in EEG diagnostics is key.
EEG Reading is More Art Than Science
EEG is a vital tool for neurologists as it is a relatively non-invasive, real-time assessment of neural activity. However, determining whether an individual EEG is “normal” or “abnormal” comes down to the physician(s) that reviews, or reads, the raw data.
Research has demonstrated that expert readers, those with many years of training specifically detecting epileptiform abnormalities within EEG, don’t always agree. Overall, expert epileptologists generally agree whether a full EEG contains any abnormalities, but the reliability goes down when assessing whether an individual wave is abnormal.3 This finding seems to stem from two linked issues:
1) Forced binary choices
2) Subjective thresholds for these decisions
Diagnostic tools tend to support the idea of a binary diagnostic decision, i.e., either “normal” or “abnormal” instead of some gradient of a possible abnormality. When having to decide between these two options, experts may apply different thresholds of what they consider “abnormal” when assessing specific spikes.3
These subjective thresholds may explain why there is poor agreement even within a single rater – a single rater is not guaranteed to provide the same diagnosis when blindly reviewing the same exam (one study finds only moderate agreement with an aggregate Cohen’s Kappa = .59).4 Reliability is even poorer between multiple raters (aggregate Cohen’s Kappa decreases to .44).4
Complicating this issue is that despite these low reliability numbers, EEG readers are highly confident in their assessments. In the same study reporting low reliability, the median reported confidence was 99%, indicating that most of the diagnostic interpretations were provided with almost absolute certainty, even when expert raters disagreed with their own assessments4. The variety of potential EEG findings can be assumed to cause these disagreements, nevertheless disagreements between readers still occur around 12% of the time, even when simplifying interpretation categories to simple “normal,” “ictal,” or “non-ictal”.5
Teaching EEG Reads
In a 2017 survey, just over a third of AAN graduating residents reported they were confident in performing or interpreting an adult EEG independently.6 For those specializing in pediatrics, the numbers rose to two thirds of respondents expressing confidence,6 presumably due to more exposure to this tool during training.
- 96% of residents considered learning to read EEGs at least very important 7
- Yet 43% of residents reported being able to read EEGs even with supervision, and were unable to identify over half of normal and abnormal EEG findings on assessments 7
- Around 1/3 reported confidence in identifying common abnormalities, normal variants, and artifacts7
These findings highlight a discrepancy between outcomes and expectations, and reflect a complaint within the field that training for reliable EEG readers is lacking.8 Common barriers include insufficient exposure to raw EEG data, lack of responsibility in independently performing diagnoses, or lack of direct patient care.7
Ways to Improve the Reliability of EEG Reads
The chasm between overconfident experts and underconfident trainees is most likely due to experience. Experts have waded the waters of raw EEG for years and have tuned their subjective thresholds on what is, and is not, considered problematic. They have, most likely, committed the errors of over (or under) reading an exam and adjusted treatment based upon experience.
Online resources such as AAN NeuroLearn and AES EEG Learning Curriculum seek to provide experience in reading raw EEG with feedback from experts in the field. However, these resources exist behind “paywalls” and do not have the real-time guidance typically afforded by residency programs. Furthermore, residency milestones like “EEG competency” have started to be challenged to provide more specificity in the methods of training (e.g., having a minimum number of EEGs to read during training, or requiring an Epilepsy Monitoring Unit rotation) as well as the types of characteristics that a graduate should be able to identify independently (e.g., having a common standard across the field of what are “common EEG abnormalities”).9
The implementation of co-readers or consensus reads on EEGs can increase the reliability of any single diagnostic decision.10 However, another way to improve EEG reader reliability may lie within new technology as opposed to training upgrades. Seizure detection software can supplement the tedious task of identifying abnormal incidences within long-term EEG, and ever-improving machine learning can support readers making confident diagnostic decisions. Additionally, if deep learning models are trained on consensus data obtained from a large panel of experts, one could potentially see interpretations more consistently accurate than any single human reader. Given that deep learning models are not limited to finding patterns detectable with human visual analysis, detection of epileptic activity in what appears to be normal EEG may be possible. This has already been demonstrated using voltage maps of spike activity11 and is an ongoing project within Stratus’ R&D division, MERLN.
Finally, an alternate path has been proposed to embrace ambiguity in diagnoses,3, 12 providing the diagnosis of “possible epilepsy” or “unclassified paroxysmal event.” The idea here is that it may reduce misdiagnoses by eliminating the binary distinction of “abnormal” versus “normal,” allowing for treatment while seeking additional information. Inaccurate diagnoses may take years to discover and remove,13 and theoretically, accepting a degree of uncertainty raises subjective thresholds for confirmed epilepsy diagnoses and could potentially improve diagnostic reliability across the board. Ultimately improving reader reliability may involve a combination of these proposed interventions, but the necessary treatment for this issue can arise only when the field recognizes these issues. To improve proper diagnostic and treatment outcomes for patients, the field needs to identify and treat the challenges within practice and training.
About the Author
Hans Klein, Ph.D., is the Manager of Scientific Publications for Stratus. Dr. Klein is a social neuroscientist by training and received his doctorate at the University of Texas at Dallas, where his research focused on the neural underpinnings of social cognitive deficits within schizophrenia spectrum disorders, as well as methods for improving measurement and research design.
- Benbadis SR, Kaplan PW. The Dangers of Over-Reading an EEG. Journal of Clinical Neurophysiology. 2019;36(4). doi: 10.1097/WNP.0000000000000598
- Tatum WO. How not to read an EEG. Neurology. 2013;80(1 Supplement 1):S1. doi:10.1212/WNL.0b013e318279730e
- Jing J, Herlopian A, Karakis I, Ng M, Halford JJ, Lam A, et al. Interrater Reliability of Experts in Identifying Interictal Epileptiform Discharges in Electroencephalograms. JAMA Neurol. 2020;77(1):49-57. doi:10.1001/jamaneurol.2019.3531
- Grant AC, Abdel-Baki SG, Weedon J, Arnedo V, Chari G, Koziorynska E, et al. EEG interpretation reliability and interpreter confidence: a large single-center study. Epilepsy Behav. 2014;32:102-7. doi:10.1016/j.yebeh.2014.01.011
- Brenner JM, Kent P, Wojcik SM, Grant W. Rapid diagnosis of nonconvulsive status epilepticus using reduced-lead electroencephalography. West J Emerg Med. 2015;16(3):442-6. doi:10.5811/westjem.2015.3.24137
- Mahajan A, Cahill C, Scharf E, Gupta S, Ahrens S, Joe E, et al. Neurology residency training in 2017: A survey of preparation, perspectives, and plans. Neurology. 2019;92(2):76-83. doi:10.1212/WNL.0000000000006739
- Nascimento FA, Maheshwari A, Chu J, Gavvala JR. EEG education in neurology residency: background knowledge and focal challenges. Epileptic Disord. 2020;22(6):769-74. doi:10.1684/epd.2020.1231
- Noe K. Most Experts Agree … But What About Other EEG Readers? Epilepsy Currents. 2020;20(2):78-9. doi:10.1177/1535759720901511
- Nascimento FA, Gavvala JR. Education Research: Neurology Resident EEG Education: A Survey of US Neurology Residency Program Directors. Neurology. 2021;96(17):821-4. doi:10.1212/WNL.0000000000011354
- Halford JJ, Arain A, Kalamangalam GP, LaRoche SM, Leonardo B, Basha M, et al. Characteristics of EEG Interpreters Associated With Higher Interrater Agreement. J Clin Neurophysiol. 2017;34(2):168-73. doi:10.1097/WNP.0000000000000344
- Baldini S, Pittau F, Birot G, Rochas V, Tomescu MI, Vulliemoz S, et al. Detection of epileptic activity in presumably normal EEG. Brain Commun. 2020;2(2):fcaa104. doi:10.1093/braincomms/fcaa104
- Oto MM. The misdiagnosis of epilepsy: Appraising risks and managing uncertainty. Seizure. 2017;44:143-6. doi:10.1016/j.seizure.2016.11.029
- Asadi-Pooya AA. The true prevalence of psychogenic nonepileptic seizures is much higher than this. Epilepsia. 2021;62(11):2875-6. doi:10.1111/epi.17053