Abstracts

Rationale:
Long-term EEG monitoring (LTM) is increasingly used in hospitalized patients (Hill CE et al, Neurology, 2019). However, the massive data volume of recordings challenges their permanent archival. Traditionally, over 95% of raw data is discarded after the patient encounter, hindering post-hoc analysis, retrospective research, and clinical audit. We hypothesized that suitable numerical processing would substantially reduce size of raw LTM datasets without compromising diagnostic yield. In this pilot study, we applied two classical data processing techniques—singular value decomposition (SVD; Stewart GW, SIAM Review, 1993) and discrete cosine transform (DCT; Ahmed et al, IEEE Trans Comp, 1974)—to reduce raw LTM data into compact forms, subsequently visually comparing the reconstructed and original data.

Methods:
We analyzed one-hour segments of artifact-free scalp EEG data (1-70 Hz passband at 256 Hz) from 10 patients with typical diagnoses encountered in LTM (e.g., slowing, seizures, periodic patterns). Using custom scripts in MATLAB^®, we subjected successive 10-second epochs to SVD, retaining robust components (typically comprising 70-80% of variance). These retained temporal components (singular vectors) were then transformed with DCT, utilizing 16-bit quantization and run-length encoding. We achieved a target data reduction (compression ratio of 20 (CR = uncompressed size / compressed size) in two ways: (i) with CR of SVD and DCT 1.7 and 12 respectively (1.7 × 12 ≈ 20); and (ii) with CR of SVD and DCT 3.7 and 5.7, respectively (3.7 × 5.7 ≈ 20). The DCT and SVD were then inverted and concatenated serially into one-hour long reconstructions, COMP1 and COMP2. In preparation for visual review, COMP1 and COMP2, along with two copies of the original EEG (ORIG1 and ORIG2) for each patient, were randomized into 40 records. A board-certified electroencephalographer (M-JB), blinded to the patient selection process, scored these recordings across 35 diagnostic features in five main categories (Hirsch et al, JCNP, 2021). These scores were used to compare the performance of the two approaches (COMP1 and COMP2) against intra-rater variation (ORIG1 and ORIG2) through the Wilcoxon signed-rank test between the five quantities O12 = ORIG1 – ORIG2, D11 = ORIG1 - COMP1, D12 = ORIG1 – COMP2, D21 = ORIG2 – COMP1, and D22 = ORIG2 - COMP2.

Results:
O12 was not significantly smaller (p > 0.68) than D11, D12, D21, or D22.

Conclusions:
The current era of data science promises transformative change in clinical neurophysiology (Tveit et al., JAMA Neurology, 2023). Our study focused on dimension reduction of data, exploring redundancies in raw LTM EEG that nevertheless preserved visual diagnostic information. We show that LTM-EEG retains diagnostic fidelity even under extreme compression sufficient to reduce a 24-hour record to 5% of its value (72 minutes). Our techniques illustrate a way of archiving all LTM raw data, directly impacting patient care and future research. Validation of this work with a larger cohort will explore the biological significance of our data representations as potential 'core' brain signals.

Funding: None