Artifact Removal Quality Assessment (ARQA): A Recording-Level Audit of EEG Artifact Removal Using Internal Preservation Controls

The human brain is a constant generator of electrical signals, a faint but rhythmic hum that researchers can capture by placing small sensors on the scalp. This recording, known as an electroencephalogram or EEG, offers a direct window into the mind's activity, revealing patterns that change with thought, sleep, or disease. However, this delicate signal is easily drowned out by the body's own movements. A blink, a grimace, a shift in posture, or even the subtle tension of a jaw muscle can create electrical spikes that look like brain activity but are actually noise. For decades, scientists have developed computer programs to scrub these errors away, hoping to leave behind a clean, pure record of the brain. The central challenge has never been whether these programs can remove the noise, but whether they accidentally throw out the brain's own voice along with it. A tool might smooth the line on a graph, making it look perfect, while simultaneously erasing the very biological rhythms doctors need to study.

In a new study, a team of researchers from institutions across the globe introduces a fresh way to check if these cleaning tools are doing their job correctly. They call their method an audit, a systematic review that treats the original recording like a control group. Instead of trusting the computer's internal report on how well it cleaned the data, the researchers first ask an expert to mark the parts of the original recording that are already clean and free of movement. They then run the recording through the cleaning software and compare the "clean" parts of the original against the "clean" parts of the processed version. If the software has done its job right, those untouched sections should look almost identical. If the software has been too aggressive, those same sections will show signs of damage, revealing that the tool is altering the brain's true signal even when it doesn't need to.

The researchers tested this audit on two real patient recordings, each presenting a different kind of challenge. The first recording was heavily contaminated with movement, filled with large, chaotic spikes that obscured the brain's activity. The second was remarkably quiet, containing only a few brief moments of disturbance. Both were processed using the same advanced cleaning software, a system designed to separate brain signals from noise without needing a human to manually pick out the errors. When the team applied their audit, the results were starkly different for the two cases. In the noisy recording, the software successfully reduced the large movement spikes, but it also significantly altered the parts of the signal that were supposed to be clean. The waves in the "clean" sections changed shape and lost much of their strength, particularly in the frequency bands associated with relaxed wakefulness. The software had smoothed the noise, but in doing so, it had also flattened the brain's natural rhythm, reducing the amplitude of these important signals by nearly ninety percent in some areas.

In contrast, the quiet recording told a different story. Because there was so little noise to begin with, the software had very little to remove. The audit showed that the "clean" sections of this recording remained almost perfectly preserved, with the processed signal matching the original to a degree of near perfection. The software did not distort the brain's activity here because there was no heavy contamination to force it to make drastic changes. However, this success came with a caveat: because the recording was so clean, the audit could not fully test how well the software handles difficult noise. It proved the tool could be gentle, but it could not prove it could be tough without causing harm.

The study also compared two different versions of the same cleaning software. The newer version showed a distinct improvement in handling the noisy recording. It managed to reduce the movement artifacts while causing far less damage to the clean brain signals, preserving the natural rhythms much better than the older version. Yet, when applied to the quiet recording, the newer version introduced a tiny, almost imperceptible change in the front of the brain that the older version had avoided. This subtle trade-off—better protection in messy situations, slightly more alteration in clean ones—was invisible to the software's own internal scoring system, which gave both versions the same high rating. The audit revealed what the internal score missed: that a tool's performance depends entirely on the specific recording it is cleaning, and that a single score cannot capture these nuances.

The researchers emphasize that their goal is not to declare one software better than another, but to provide a transparent report card for every single recording. They argue that a cleaning method cannot be judged as universally good or bad; its impact changes based on the amount and type of noise present. A tool that works perfectly on a quiet recording might be destructive on a noisy one, and vice versa. By separating the measurement of how well noise was removed from how well the original signal was kept, the audit offers a clearer picture of what actually happened to the data. This approach ensures that when scientists or doctors look at a cleaned EEG, they know exactly how much of the original brain activity remains and how much of the noise was truly eliminated, rather than relying on a computer's promise that the data is clean.

Ultimately, the study suggests that the future of brain signal analysis lies in this kind of detailed, recording-level scrutiny. The team plans to refine their method so that it can automatically identify the clean and noisy parts of a recording without needing a human expert to mark them first. Until then, their work serves as a reminder that in the quest for clean data, the most important step is verifying that the cleaning process has not erased the very thing we are trying to see. The audit does not replace the need for human judgment, but it provides a rigorous, evidence-based framework to support it, ensuring that the story told by the brain is not lost in the effort to silence the noise.