Classifying Sleep Slow Oscillations in Low Density EEG

This study relied on two datasets of sleep polysomnography nights acquired in the Mednick Lab at the University of California, Irvine. The first was a base dataset (used to train and test our ML model) comprising of EEG data from 22 healthy individuals (9 females and 13 males; ages 18–34, mean age 21.9) acquired with a 64-channel setup, including 58 head EEG channels sampled at 1000 Hz and six additional channels used for reference, grounding, and the capture of other biosignals such as electromyographic, electrooculographic and electrocardiographic (Seok et al., 2022; Alipour et al., 2024). In accordance with our goal to create a robustly generalizable model, the second dataset was set apart as a validation set of nighttime sleep from 34 healthy adults (18 females and 16 males; ages 18–30, mean age 20.7) acquired with a 32-channel EEG, including 24 head channels sampled at 1000 Hz and 10 channels supporting collection of other biosignals (mastoids, EMG, EOG, EKG) (Rechtschaffen & Kales, 1968; Malerba et al., 2019). In both cohorts, the data was originally referenced through FCz and then re-refenced to the contralateral mastoid; reference was not subsequently adjusted during analysis. The similar age distributions of the two cohorts addressed the well-attested effect of age on SOs (Landolt et al., 1996; Carrier et al., 2001; Mander et al., 2017; Rosinvil et al., 2021) and restricted the applicability of our classifier to persons falling in this age-range. The Mednick laboratory performed sleep staging to divide the data into N2, N3 and additional sleep-stages following the criteria in Rechtschaffen & Kales’ manual as applied in the Matlab package HUME, and identified epochs in which all channels were free of artifacts. Artifact marking was performed first by visual inspection (Mednick group), and supplemented by algorithmic detection in each channel (Malerba group). We used two computational algorithms for removing muscle movement artifacts. In the first, introduced by Brunner and colleagues, we filtered the signal to the range 26.25–32 Hz and then split into 4-second time bins; bins more than 4 times greater than the median of the 45 bins (3 min) surrounding the epoch were excluded as artifacts (Brunner et al., 1996). Second, we applied an analogous method from Wang et al. that used a similar filtration to 4–50 Hz, a splitting of the signal into 5-second bins, and a removal of bins exceeding 6 times the median value over all bins (Wang et al., 2020).

We detected SO events at each channel for all nights in both datasets, leveraging the algorithmic detection approach developed by our group (Niknazar et al., 2022; Alipour et al., 2025), which closely follows those set forth in Massimini, and in Dang-Vu (Massimini et al., 2004; Dang-Vu and others, 2008). In brief, the signal was band-filtered to the range 0.1–4 Hz, and an SO defined to be any segment satisfying the following conditions. First, the event had to be included in the time between three subsequent zero-crossings, with the signal being negative between the first two crossings and positive between the second and third crossings. Also, the first two crossings had to occur between 0.3 and 1s of each other and the whole event (from first to third zero crossing) could span at most 10s. The two time constraints ensured that the de-polarization, hyper-polarization and subsequent relaxation of voltage in the channel were all part of a single physiological event. Second, the trough voltage value was required to be at most − 80 uV (i.e., being a negative value that exceeded − 80), and the whole event range from peak-to-trough had to be at least 80 uV. These requirements ensure that the magnitude of the SO event would be large compared to ongoing dynamics. Among all the candidate events that satisfied the conditions, only events that began and ended within the same sleep stage (N2 or N3) were kept. Lastly, events with amplitudes larger than four standard deviations among the remaining candidate detections at each channel were discarded as possible artifacts.

Our SO sets were subsequently assigned to the categories of Global, Frontal, or Local by applying a k-means clustering to a binary matrix encoding the co-detections of trough events across multiple channels at a short delay (± 400 ms) (Malerba et al., 2019; Niknazar et al., 2022; Alipour et al., 2025; Snedden et al., 2026). Finally, we restricted our study to SOs detected in at least one of our 8 channels of interest (see below). This procedure yielded 118,308 SOs during N3 sleep in our base set (44,477 Global, 34,840 Frontal and 38,991 Local; and mean 5,378 SOs per individual), and 196,131 SOs during N3 in our validation set (86,278 Global, 48,021 Frontal and 61,832 Local; and mean 5,768 SOs per individual). Supplementary Table S1 in the Online Resource provides aggregate SO counts by type across sleep stages and datasets, and Supplementary Table S2 gives SO counts by individual. All these data processing steps were accomplished with in-house software in Matlab (MathWorks, Natick, MA).

Intuitively, these three SO types can be understood as follows: Global SOs as events large in amplitude and showing a large footprint on the scalp; Frontal SOs as moderate amplitude events restricted to mostly the frontal regions; and Local SOs as small amplitude events found in any scalp area without specific preference .

Having preprocessed the Global, Frontal, and Local SOs datasets, each SO event was uniquely associated with its channel and time of detection. We then set out to isolate a spatially down-sampled version of the EEG trace for each detected SO. We focused the work on at most 8 channels: F3, F4, C3, C4, P3, P4, O1 and O2, which track activity in respectively the frontal, central, parietal and occipital regions. For each SO event, we exported the data from these 8 channels for a 2s-long epoch centered on the time of the trough of the detected SO. Fig. 1A shows an example SO with the 8 traces stacked from frontal to occipital. We then imported these data into Python as a single large numpy array. For ease of computation, we down-sampled each channel 2s-long epoch to 200 Hz prior to calculating features.

The next step was the computation of ML features, which our classifier would use to infer the SO’s type. We computed 35 total features at each of 8 channels, for a total of 280 predictors. We initially defined a special class of “base features,” chosen for their easy interpretability and high predictive power (Fig. 1B). These base features comprised the amplitude and time offset from the SO detection at three key events: the channel trough, the last upstream peak before the SO’s occurrence (t = 0), and the first downstream peak after the SO’s occurrence. As shown in Fig. 1B, these quantities can be thought of as the x- and y-coordinates of the three events on the graph of the waveform within the channel. We chose these features to be fundamental, robust and accessible. We emphasize that the time-offsets were measured from the node (electrode) of detection and could therefore be very different from 0 in the trough. In addition to the six base features, at each channel we computed 29 additional features, which we refer to as “advanced features.” These values encompass an array of electrophysiological, information-theoretic measures and miscellaneous waveform properties, falling into five categories:

As an intuitive introduction to the different quantities that the advanced features encode, we plot the distribution of two features in Fig. 1C-D: mean power density in the delta band in electrode C3 and permutational entropy in F4. The mean power density captures the energy of the waveform in a particular frequency band, while permutational entropy quantifies the diversity of ordinal relationships – up, up, down or up, down, up – in the data. The differentiation of feature distributions across the three SO classes is what empowers our classifier. Specifics of the calculation of each feature are given in Supplementary Information, distributions of additional select features are depicted in Supplementary Figure S1, and a full catalog of features with mean values in each channel during N3 is provided in Supplementary Table S3.

After constructing our bank of feature values for each SO event in each dataset, our next step was to construct classifiers to predict SO type. We began by testing 27 different ML architectures on our train/test dataset using the python package lazypredict (Pandala, 2024) (Supplementary Figure S2), restricting to sleep stage N3 and utilizing all features in a multiclass framework. Importantly, no data from the validation set was used in the construction of the classifiers. The top performing architectures proved to be two tree-based paradigms: LGBMCLassifier and XGBClassifier. Next, because neural nets are often the most powerful architecture but may require meticulous preprocessing, we manually constructed several neural network designs through PyTorch, experimenting with different topologies and input normalization rules. This approach failed to produce a model equaling the performance of our naïve tree-based classifiers. Therefore, we adopted one of the tree-based models as our final architecture of choice, settling on XGBoost for its speed, familiarity and interpretability.

Next, we investigated whether multiple binary classifiers might yield better results than a single multiclassifier. However, multiclass proved the superior option. Finally, we explored whether the Z-normalization of feature-values, either globally or by subject, might confer any gain in performance. No gain from normalization was apparent, so we opted to use raw inputs, a choice that also enhanced the model’s portability.

We employed five random 70 − 30 train-test splits of our train/test dataset, splitting by human subject in order to avoid data leakage between SOs from the same individual. We then trained a model on these five splits, and took our final predicted probability to be the average across the five predictions.

Despite restrictions to a much smaller set of inputs than SO classification has historically utilized, our models achieved high efficacy in identifying the three SO types. Fig. 2 summarizes the performance of our full and base-feature models across our test and validation sets. In Fig. 2A-B, the performance of our full-feature model is shown in blue, while the results for the base-feature model are shown in red. Results on the held-out tests are shown in darker hues, while the lighter hues indicate results on the validation dataset. In both plots, we have included a reference line (in red) for performance at chance. Both our models (base- and full-feature) achieved accuracy (defined as correctly classified records divided by total records (Rainio et al., 2024) between 74% and 77%, with slightly lower performance on the validation set. To verify that accuracy was not achieved solely by trivially assigning every record to the dominant class, we also used the log-loss metric, shown in Fig. 2B. The models showed log-loss values between 0.55 and 0.6, with a slight increase for the validation set.

The confusion matrix in Fig. 2C, which depicts our full-feature model’s performance on the test set, provides a more granular view of the performance. Notably, the most problematic distinction to draw was between the Frontal and Local classes, with the off-diagonal cells, which signify misclassification, containing more elements in this case than do most of the off-diagonal cells involving Global SOs, despite Global being the largest class. We provide confusion-matrix data for all models in Supplementary Table S4.

We apply the SHAP feature-importance package to discern the features that are most informative for each SO type (Lundberg and Lee, 2017). These results, computed across each held-out test set, are summarized in Fig. 3. In the plots, we report features ranked by the total length of their SHAP bars. We interpret the top features as being those most characteristic of the three SO types.

In Fig. 3A and B, colors represent the three SO types, while bar length shows the magnitude of the contribution of a given feature to predicting that the SO would be (or not be) of a given type. Bar direction along the x-axis indicates the sign of the correlation between feature and classification. For example, in Fig. 3A, the long right-directed green bar for the first feature of ‘δ-power in C3’ signifies that a high-value of this feature contributes towards a prediction of Global label, and equally, that a low value discourages a Global classification. On the other hand, the long negative blue bar in the second feature, ‘Trough time offset in F3’, indicates that a low value of this feature pushes the model towards a classification of local, while a high value discourages it. Fig. 3A and B depict the top features for our full and base models, as measured by average absolute SHAP value across all classes. The plot shows that the full-feature classifier with 8 channels is likely to use delta power at C3, the offset of the trough at F3 and the amplitude of the trough at F4 when assigning a label to an SO (note that our trough amplitude feature values are kept as negative numbers when passed to the models, so a low value of amplitude means a large trough whereas a large value of amplitude maps to a shallow trough). A more detailed exposition of SHAP and its usage in Fig. 3 is given in the “SHAP Values” section in Supplementary Information, and in Supplementary Figure S3.

All of our features were computed independently for every active electrode. To establish which features were most informative across all electrodes in general, rather than within a particular scalp region, we computed the average absolute SHAP value of each feature across all channels. The results, shown in Fig. 3C and D, show that trough amplitude was consistently the most informative feature on SO type, followed by time offset.

These results demonstrated that the features of delta power spectral density and trough amplitude were most informative for classification of the full features and base models, respectively. We were interested in establishing whether these features had a differential impact on classification depending on the electrode at which they were measured. Thus, we compared the SHAP value contribution to the Global prediction at each of our eight electrodes on a topographical scalp plot, while using the absolute value for trough amplitude to give large troughs a positive sign and emphasize the similarity of the two features’ substantive contributions. Fig. 4 shows that the two features serve approximately equivalent functions in the two models, with (1) heightened values in central and parietal locations increasing likelihood of global labeling, and low likelihood of frontal or local labeling; (2) large frontal activity in the absence of central activity increases chance of frontal labeling; and (3) frontal activity of average size strongly reduces the likelihood of local labeling.

Having demonstrated that we could use feature-derived information from a low-density EEG montage with eight channels to identify the space-time profile of a given SO with high accuracy, we were interested in evaluating whether this approach could achieve comparable success when applied to montages with even lower electrode density. Thus, we re-trained our full- and base-feature models on different channel subsets, progressively removing channels at occipital, parietal and central locations. Our full list of models comprised 1) our original “FCPO” set (eight channels: F3, F4, C3, C4, P3, P4, O1, and O2); 2) an “ FCP” set (six channels, removing the occipital leads); 3) a clinically motivated “FCO” set (six channels, removing the parietal leads); 4) an “FC” set (four channels, removing occipital and parietal leads); and finally 5) a two-channel set “F” (only F3 and F4). Labels for training were derived from the Global, Frontal, and Local SO assigned values using our original algorithm built using a much higher density EEG. This expansion brought our total electrode number of models to 10, with five different sets of electrodes and either the “full” or “base” feature-banks.

We summarize the performance of our full battery of models in Fig. 5. As expected, the models always performed better when leveraging information from additional channels, with the full-feature, 8-channel model achieving the highest test- and validation-set accuracies of respectively 76.3, 75.8%. Nonetheless, the six-node models (FCP and FCO, base or full features) showed only a modest drop in performance, with accuracy remaining over 70% in all scenarios, with relative losses of < 2% for FCP and < 4% for FCO when compared to models of the FCPO set. Furthermore, the FC models (four channels) also showed acceptable performance, averaging around 70% accuracy. We show model-performance under all configurations and datasets in Supplementary Table S5.

Having showed that we could achieve classification leveraging SOs that were detected in sleep stage N3, we were curious whether the same approach might be generalizable to other stages. To probe this question, we repeated our entire training and testing procedure on the set of SOs captured during N2 sleep, and then tested both N2- and N3-based models on validation data from both sleep stages.

This study revealed that our models’ efficacy was basically undiminished when trained and tested on N2 rather than N3 data (full results in Supplementary Table S6). Eight-channel models based on N2 achieved accuracy within 1.5% of their N3 counterparts on their corresponding validation sets under both the full- and base-feature frameworks. Similar results held for models built on restricted channel-sets, except the 2-node-model. Furthermore, and somewhat surprisingly, the N3-trained models proved proficient at classifying SOs from the N2 stage, again with a relative loss in accuracy of at worst around 1.5%. However, the N2-trained models were unable to generalize cleanly to N3, with relative accuracy dropping by 7% under most configurations. Thus, our N3 models’ direct portability to novel data from stage N2 indicates a physiology-driven stability of SO attributes across different phases of sleep.