A codec designed to make music sound good to human ears could quietly erase information that doctors need from recordings of the brain and muscles, according to a new study. Researchers evaluating MPEG-4 Advanced Audio Coding, or AAC, found that the widely used audio-compression format introduced distortions into electroencephalography (EEG) and electromyography (EMG) signals—sometimes altering frequency bands that carry clinically meaningful information. The findings suggest that an audio file that sounds perfectly acceptable to a listener may be an unreliable record of neurophysiological activity, raising concerns about using conventional multimedia codecs to store or transmit medical waveforms.
The study, published in the Journal of Medical Systems, examined whether AAC could serve as a compression method for neurophysiology data intended for the Digital Imaging and Communications in Medicine (DICOM) ecosystem. EEG records electrical activity generated by synchronized populations of neurons through electrodes placed on the scalp or, in some cases, implanted inside the skull. Clinicians use it to investigate seizures, sleep disorders, brain injuries, encephalopathies and neurodegenerative disease. EMG records electrical impulses produced by muscles during activity and helps diagnose conditions involving muscle tissue, peripheral nerves and motor neurons, including amyotrophic lateral sclerosis and muscular dystrophy. Both techniques generate large, high-resolution datasets, particularly as recording systems gain more channels and higher sampling rates.
Compression is increasingly important because modern neurophysiology systems can produce enormous quantities of data that must be archived, retrieved and transferred between hospitals or research centers. But biomedical waveforms pose a fundamentally different challenge from music. Audio codecs are engineered around psychoacoustic models: mathematical descriptions of what human listeners are likely to notice. When a quiet, high-frequency component is masked by a louder sound, an audio encoder may reduce or remove it because the change is expected to be imperceptible. In EEG or EMG, however, a low-amplitude high-frequency feature can be diagnostically important even if it contributes little to the overall signal energy. A sharp epileptiform transient or a high-frequency oscillation does not become irrelevant merely because human hearing would not detect its loss.
The researchers tested de-identified EEG and EMG recordings with the Fraunhofer Institute for Integrated Circuits’ FDK AAC encoder. Their EEG collection contained approximately 10-minute excerpts from 41 recordings made using Natus equipment, including scalp recordings based on the international 10–20 electrode system and intracranial EEG captured from depth electrodes. From those data, the team selected 150 one-minute, single-channel segments containing seizures or interictal epileptiform discharges—abnormal electrical patterns that occur between seizures. These events were deliberately useful stress tests because they are nonstationary, meaning their statistical properties change over time, and can include abrupt, sharp transients that are vulnerable to lossy processing. The EMG sample consisted of 145 single-channel segments drawn from multichannel surface recordings of four forearm muscles in 40 subjects.
To judge whether the reconstructed signals still looked clinically trustworthy, the team displayed original and compressed-then-decompressed waveforms to eight experts through EEGnet, a web-based system for viewing and annotating biomedical signals. Each reviewer answered whether the two versions showed a clinically significant difference, defined as a distortion that could meaningfully affect interpretation. The resulting score ranged from zero, when no experts identified an unacceptable difference, to eight, when all reviewers did. The investigators also calculated Percentage Root Mean Square Difference, or PRD, a standard numerical measure of reconstruction error. PRD compares the squared difference between every original sample and its reconstructed counterpart with the energy of the original signal, then expresses the result as a percentage. A low PRD indicates closer numerical agreement, although the study showed that a single global error number does not always predict what experts will notice.
The visual assessments pointed to a striking difference between the two signal types. In the tested EEG data, reconstructed waveforms generally appeared acceptable to most reviewers when PRD remained below approximately 15 percent. Above that level, expert judgments more often indicated clinically meaningful artifacts. For surface EMG, possible quality loss appeared at a much lower PRD, around 1 percent. The authors emphasize that these values are exploratory observations from the datasets and scoring system, not universal clinical thresholds. In fact, none of the tested EMG distortion levels produced majority agreement among experts that the signals were clinically unacceptable, making the approximately 1 percent observation tentative. Some EEG segments with relatively low PRD were still judged distorted by individual reviewers, while certain segments with higher PRD attracted little concern, underscoring the limitations of relying on one aggregate metric.
The frequency-domain analysis provided stronger evidence that AAC’s errors were not distributed harmlessly across the signal. Researchers estimated power spectral density—the distribution of signal power across frequencies—using Welch’s method and compared the original and reconstructed recordings in the delta, theta, alpha, beta and gamma bands. For EMG, which contains substantial energy above 100 hertz, the gamma range was further divided into low gamma from 30 to 100 hertz and high gamma from 100 to 400 hertz. The largest changes in EEG were concentrated in gamma and beta activity, with additional effects depending on whether the signal came from the scalp or from intracranial electrodes. Surface EEG showed prominent error power in gamma, beta and theta bands, whereas intracranial EEG showed the strongest effects in gamma, beta and alpha. EMG differences were most apparent in gamma, beta and alpha, and high-gamma distortion rose slightly at higher compression levels.
Statistical testing supported these patterns. The researchers compared low- and high-PRD groups within the EEG and EMG datasets using Wilcoxon rank-sum tests, a nonparametric method suitable when distributions cannot safely be assumed to be normal. They corrected for the five canonical frequency-band comparisons using a Bonferroni-adjusted significance threshold of 0.01. Scalp EEG displayed small-to-moderate effects in a subset of bands, particularly delta and gamma. Intracranial EEG exhibited broader and larger effects across the frequency spectrum, while EMG showed large effects across all canonical bands. The greater sensitivity of intracranial recordings may reflect their higher sampling rates and distinct signal morphology, although the study included only 16 intracranial EEG recordings and was not designed to establish a separate acceptability limit for that modality. Some segments originated from the same parent recording, another reason the authors describe the analysis as exploratory rather than a definitive population-level assessment.
The underlying problem is that AAC optimizes perceptual fidelity, not scientific fidelity. Its psychoacoustic model can attenuate weak components when stronger neighboring frequencies are present, especially in portions of the spectrum that are less important to human hearing. For speech or music, this trade-off can produce dramatic reductions in file size with little audible penalty. For neurological recordings, the same operation may smooth a waveform, suppress a transient or reshape a frequency distribution. That could complicate the recognition of epileptiform spikes, seizure onset patterns or high-frequency oscillations. The researchers do not claim that every AAC-compressed recording would cause a missed diagnosis, but they argue that the possibility of selective, clinically relevant information loss makes the codec unsuitable for archival neurophysiology data where the original waveform may later be reanalyzed.
The study arrives as medical-data standards organizations seek a dedicated way to compress biomedical waveforms without sacrificing features that audio codecs are designed to discard. DICOM supports transfer syntaxes based on standardized codecs, yet no broadly adopted standard currently addresses EEG, EMG and related scientific waveforms with the same interoperability framework. Academic alternatives—including wavelet, predictive, entropy-based, compressed-sensing and machine-learning methods—can achieve low reconstruction error at selected operating points, but many are not standardized or interoperable within DICOM. The authors say their results contributed evidence to international discussions, including an ITU-T call for proposals on biomedical waveform coding. Future work will evaluate an emerging dedicated codec associated with ITU-T Recommendation H.BWC and the proposed MPEG T.261 standard, while expanding testing to needle EMG, microelectrode EEG and larger intracranial datasets. For now, the message is simple but consequential: a file can be small, playable and apparently clean while no longer preserving the biology it was meant to record.
