SleepLiteCNN: Lightweight Sleep Analysis Models

• SleepLiteCNN is a family of compact deep learning models designed for processing raw sleep biosignals like EEG and ECG under resource constraints.
• It achieves robust performance in multi-class sleep staging and apnea subtype classification through end-to-end learning and subject-specific adaptation.
• The architecture variations—from pure CNNs to CNN-Transformer hybrids—demonstrate adaptability across different signal modalities and deployment platforms.

SleepLiteCNN denotes a family of lightweight deep-learning architectures for automated sleep analysis rather than a single immutable network. In the literature considered here, the name is associated with compact models for single-channel EEG sleep stage scoring, two-channel EEG staging, edge-oriented CNN-Transformer inference, and single-lead ECG sleep apnea subtype classification. Across these usages, the unifying theme is end-to-end learning from raw physiological time series under explicit resource or annotation constraints, but the sensing modality, temporal context, output space, and deployment target vary substantially (Tsinalis et al., 2016, Fernandez-Blanco et al., 2021, Yao et al., 2022, Mohammadi et al., 1 Aug 2025).

1. Scope, nomenclature, and problem settings

Across the cited works, SleepLiteCNN is applied to several related but non-identical formulations. The term therefore refers more reliably to a design philosophy—compact CNN-centric inference on sleep biosignals—than to a single canonical topology.

Work	Signal and labels	Technical profile
(Tsinalis et al., 2016)	Fpz-Cz EEG; Wake, N1, N2, N3, REM	Five-epoch context, raw EEG input, two-stage convolution with stacking, class-balanced SGD
(Fernandez-Blanco et al., 2021)	Fpz-Cz and Pz-Oz EEG; Wake, N1, N2, N3, REM	Seven 1D convolutional blocks, dropout, leave-one-patient-out evaluation
(Yao et al., 2022)	Fpz-Cz EEG; Wake, N1, N2, N3, REM	Four Conv1D layers, one Transformer encoder layer, subject-specific fine-tuning, embedded deployment
(Mohammadi et al., 1 Aug 2025)	Single-lead ECG; Normal, OSA, CSA, MSA	Three Conv1D layers, quantization-aware training, FPGA synthesis, energy-focused design

A common source of confusion is the assumption that SleepLiteCNN always denotes sleep stage scoring from EEG. The 2025 paper uses the name for a different task: sleep apnea subtype classification from single-lead ECG at 1-second resolution (Mohammadi et al., 1 Aug 2025). Another potential misconception is that all SleepLiteCNN variants are purely convolutional. The 2022 instantiation couples a CNN front end with a Transformer encoder, retaining the lightweight emphasis while altering the sequence-modeling mechanism (Yao et al., 2022).

2. Single-channel EEG SleepLiteCNN for sleep staging

The 2016 single-channel EEG model is the clearest early formulation of the SleepLiteCNN pattern: raw Fpz-Cz EEG is provided directly to a CNN without hand-engineered features, and the network learns discriminative filters end-to-end for five-class sleep staging (Tsinalis et al., 2016). The input is a continuous 1-D EEG signal comprising five 30-second epochs—two preceding, the target, and two succeeding—sampled at 100 Hz, for a total length of 15,000 time points with shape $(1,1,15000)$. The first convolutional layer uses 20 filters of size $(1 \times 200)$ with ReLU activation and valid convolution, followed by max-pooling with pool size $(1 \times 20)$ and stride $(1 \times 10)$. A stacking layer rearranges the 20 pooled feature maps into a 2D image of shape $(1,20,1479)$. A second convolutional layer then applies 400 filters of size $(20 \times 30)$, again with ReLU and valid padding, followed by max-pooling with pool size $(1 \times 10)$ and stride $(1 \times 2)$. The flattened representation feeds two fully connected layers of 500 units each, and the final softmax outputs class probabilities $p_\theta(y|x)$ over Wake, N1, N2, N3, and REM. No batch-normalization or dropout layers were used in the published architecture.

Training used multiclass cross-entropy with L2 regularization,

$$L(\theta) = - \sum_{i=1}^N \sum_{c=1}^5 y_{i,c}\log p_\theta(y_{i,c}|x_i) + \lambda \|\theta\|_2^2,$$

optimized by stochastic gradient descent with backpropagation. Early stopping was applied based on validation performance. A defining element was class-balanced random sampling within SGD: at each iteration, minibatches were constructed by randomly sampling an equal number of examples from each sleep stage, explicitly mitigating skew toward majority stages such as N2.

The evaluation used an openly available dataset from 20 healthy subjects, comprising 39 polysomnograms total, with movement and “Not Scored” epochs removed and N3 and N4 merged into a single N3 stage per AASM guidelines. The protocol was 20-fold cross-validation at subject level: in each fold one subject’s recordings were held out for testing; of the remaining 19 subjects, 4 were randomly chosen for validation and 15 for training. Metrics were computed from a class-balanced, or normalized, confusion matrix and included precision, sensitivity, F1-score per stage, per-stage accuracy, and overall accuracy.

Reported performance was a mean F1-score across stages of 81% with range 79–83% and 95% confidence interval [79, 83], mean per-stage accuracy of 82% with range 80–84% and 95% confidence interval [80, 84], and overall accuracy of 74% with range 71–76% and 95% confidence interval [71, 76]. The worst-stage F1-score was for N1 at 61% with range 59–63%. The normalized confusion structure was physiologically plausible: N1 was classified as N1 in 60% of cases and was misclassified mainly as REM or Wake, N3 was classified correctly in 91% of cases, REM in 74%, and Wake in 70%. No significant correlation was found between scoring performance and either sleep efficiency ($(1 \times 200)$0, $(1 \times 200)$1) or the fraction of transitional epochs ($(1 \times 200)$2, $(1 \times 200)$3).

Interpretability was an explicit component of the analysis. First-layer filters were extracted from each fold, transformed via Fourier analysis, normalized per stage and per filter, and grouped by the sleep stage on which they showed maximal activation. The learned filters aligned closely with AASM scoring criteria: N3-associated filters showed strong power at 1–1.5 Hz, often paired with 12.5–14 Hz; N2-associated filters emphasized 13–14.5 Hz and 2–4 Hz; REM-associated filters showed 2–5 Hz activity with absence of 20–50 Hz; Wake-associated filters emphasized sub-1 Hz activity; and N1-associated filters showed peaks near 7 Hz and 9 Hz, often with 20–50 Hz components. Filters with activity in 10.5–12 Hz and 15–17 Hz were absent across folds, consistent with avoidance of ambiguous alpha–spindle overlap. This establishes the model as an end-to-end learner whose internal representations map onto physiologically meaningful sleep-scoring rules.

3. Compact EEG variants and related lightweight architectures

A later lightweight EEG staging model on the expanded Sleep-EDF corpus retained the raw-signal, compact-CNN philosophy while changing the topology substantially (Fernandez-Blanco et al., 2021). Each 30-second epoch was represented by 3000 samples per channel, yielding input tensors of $(1 \times 200)$4 for the dual-channel configuration and $(1 \times 200)$5 for single-channel variants. The network comprised seven one-dimensional convolutional blocks, each followed immediately by 2-point max-pooling, then a dropout layer and a final fully connected softmax. All convolutions used stride 1, valid padding, and ReLU activations. The kernel schedule was $(1 \times 200)$6, with 20 output feature maps at each convolutional stage. After the seventh pooling layer, the representation was flattened to 400 dimensions, followed by dropout with rate 0.5 and a 5-neuron softmax head. The total trainable parameter count was 13,485.

The training loss was categorical cross-entropy, optimized with Adam at initial learning rate $(1 \times 200)$7. Training ran for up to 100 epochs with early stopping if validation loss did not improve for 10 consecutive epochs. No batch-normalization was used. The evaluation protocol was true leave-one-out cross-validation at patient level, ensuring zero subject overlap between train and test. In the dual-channel setting, mean accuracy was $(1 \times 200)$8, Cohen’s $(1 \times 200)$9 was $(1 \times 20)$0, and macro-$(1 \times 20)$1 was $(1 \times 20)$2. An ensemble of three models—Fpz-Cz only, Pz-Oz only, and both channels—used majority voting and reached essentially the same accuracy, 92.65%, but with slightly lower mean $(1 \times 20)$3. Single-channel results were lower, and Wilcoxon tests indicated that the two-channel improvement was statistically significant with $(1 \times 20)$4.

A related lightweight architecture, LightSleepNet, occupies the same design space while not using the SleepLiteCNN name (Liao et al., 2024). It operates on single-channel Fpz-Cz EEG and uses five successive 1-D convolutional layers with two residual blocks, group convolution, channel shuffle, dropout, global average pooling, and a final 128-to-5 fully connected layer. The reported footprint is 43.08 K parameters and 45.76 MFLOPs, and the paper adds a personalized unsupervised domain-adaptation procedure combining Adaptive Batch Normalization and gradient re-weighting. Reported performance for LightSleepNet with personalization was ACC = 83.8%, MF1 = 75.3%, and $(1 \times 20)$5. Its relevance here is architectural rather than terminological: it shows how the lightweight sleep-staging line evolved toward parameter-efficient residual and adaptation mechanisms under mobile constraints.

4. CNN-Transformer edge-oriented instantiation

The 2022 real-time classifier extends the SleepLiteCNN idea by combining four sequential convolutional layers with a Transformer encoder to capture time-variant dependencies in single-channel EEG (Yao et al., 2022). The input is a non-overlapping 30-second Fpz-Cz epoch sampled at 100 Hz, represented as a 3000-sample vector. The convolutional front end reduces the temporal dimension from 3000 to 19 and expands the channel depth to 128, producing a $(1 \times 20)$6 feature map. This map is interpreted as a sequence of 19 tokens of dimension 128 and passed to a single Transformer encoder layer. A final dense-softmax head outputs probabilities over Wake, N1, N2, N3, and REM.

The paper provides the standard mathematical formulation for the Conv1D stage and for Transformer self-attention. For the encoder, given input sequence $(1 \times 20)$7, the query, key, and value projections are

$(1 \times 20)$8

followed by scaled dot-product attention

$(1 \times 20)$9

The loss is standard categorical cross-entropy, and no further regularization by dropout or weight decay is reported.

Preprocessing uses per-epoch z-score standardization,

$(1 \times 10)$0

with $(1 \times 10)$1 and $(1 \times 10)$2 computed over each 3000-point epoch. Optimization uses Adam with default $(1 \times 10)$3, $(1 \times 10)$4, and $(1 \times 10)$5, batch size 64, and 5-fold cross-validation over 77 subjects, with a 90% train / 10% validation split within each fold. A distinguishing feature is subject-specific training before inference: after standard training, 10% of each subject’s test-set epochs are randomly sampled and used to fine-tune the model, and the remaining 90% are then inferred with the fine-tuned model.

Before subject-specific training, overall accuracy was 0.775 and per-class F1 scores were Wake 0.90, N1 0.32, N2 0.82, N3 0.75, and REM 0.69. After subject-specific training, overall accuracy rose to 0.795, and F1 scores became Wake 0.91, N1 0.37, N2 0.84, N3 0.77, and REM 0.73. The full model contains approximately 300,000 parameters, corresponding to about 2 MB in 32-bit floats. A reduced version, smaller than 1 MB after quantization, was deployed on an Arduino Nano 33 BLE board and yielded 68% accuracy in preliminary tests. The paper states that the reduced model was fully functional and that the full end-to-end pipeline was observed to run in real time on 30-second windows.

5. SleepLiteCNN for single-lead ECG sleep apnea subtype classification

In 2025, the name SleepLiteCNN was used explicitly for an energy-efficient model that classifies sleep apnea subtypes—Obstructive, Central, and Mixed—alongside Normal rhythm from single-lead ECG (Mohammadi et al., 1 Aug 2025). This setting differs from EEG sleep staging in both modality and label semantics. The input is an 11-second ECG snippet sampled at 128 Hz, producing 1,408 samples. Windowing follows WIN-11: an 11-second window where the first second is labeled and windows slide with 1-second resolution. No explicit denoising or normalization is applied; the raw ECG windows are fed directly into the network.

The architecture is a three-layer 1D CNN with batch normalization, dropout, and a four-way softmax output. The sequence is BatchNorm1D, Conv1D_1 with kernel size 5 and 5 filters, MaxPool1D(2), Conv1D_2 with kernel size 5 and 45 filters, MaxPool1D(2), Conv1D_3 with kernel size 5 and 25 filters, MaxPool1D(2), flatten to 4,300 features, Dropout(0.5), and a dense softmax layer with 4 outputs. The total trainable parameter count is approximately 39 K.

Training uses categorical cross-entropy over the four classes $(1 \times 10)$6Normal, OSA, CSA, MSA$(1 \times 10)$7, optimized with Adam with $(1 \times 10)$8, $(1 \times 10)$9, and $(1,20,1479)$0. The learning rate is $(1,20,1479)$1 with decay on plateau, batch size is 64, and training proceeds for 600 epochs with early stopping on validation macro-$(1,20,1479)$2. Dropout is applied before the final dense layer, and quantization-aware training is performed with QKeras for the 8-bit model. No data augmentation beyond the intrinsic overlap of sliding windows is reported.

A central contribution is explicit energy modeling. Per-inference energy is defined as

$(1,20,1479)$3

where $(1,20,1479)$4 is the number of MAC operations in layer $(1,20,1479)$5 and $(1,20,1479)$6 is the energy per MAC in that layer after quantization. Using QKeras’s energy estimator configured for a 45 nm CMOS process, full-precision SleepLiteCNN consumed 25.76 $(1,20,1479)$7J per inference, whereas the 8-bit quantized model consumed 1.80 $(1,20,1479)$8J per inference. On the held-out test set, the 8-bit model achieved 94% accuracy and macro-$(1,20,1479)$9. The per-class breakdown was Precision/Recall/$(20 \times 30)$0 of 0.92/0.95/0.93 for Normal, 0.90/0.93/0.91 for OSA, 0.93/0.90/0.91 for CSA, and 0.89/0.86/0.88 for MSA.

The hardware study targeted an Artix-7 FPGA (xc7a200tfbg484-3) using hls4ml. Quantization reduced resource usage from 31.56% to 23.21% for LUT, from 33.2% to 23.92% for FF, from 22.47% to 19.32% for BRAM, and from 26.08% to 13.38% for DSP. In the same comparison table, SleepLiteCNN had 39 K parameters, 95% full-precision score in the “Prec” column, 92% full-precision F1-score, 94% quantized score in the “Prec_q” column, 91% quantized F1, and 1.80 $(20 \times 30)$1J quantized energy. The paper positions this operating point against larger models such as VGG-11 and MobileNet-v1, emphasizing the combination of four-way subtype classification, 1-second resolution, and low per-inference energy.

6. Methodological themes, interpretability, and limits of comparison

Several methodological patterns recur across the literature. First, raw biosignal ingestion is prominent. The 2016 EEG model uses raw Fpz-Cz EEG directly as input, the 2021 EEG model applies no further band-pass filtering or hand-crafted feature extraction, and the 2025 ECG model applies no explicit denoising or normalization before inference (Tsinalis et al., 2016, Fernandez-Blanco et al., 2021, Mohammadi et al., 1 Aug 2025). Second, the architectures are intentionally compact relative to broader deep-learning practice: the 2021 two-channel model has 13,485 trainable parameters, LightSleepNet reports 43.08 K parameters, the 2025 ECG SleepLiteCNN has approximately 39 K parameters, and the 2022 CNN-Transformer model is larger at roughly 300,000 parameters but still designed for constrained hardware (Fernandez-Blanco et al., 2021, Liao et al., 2024, Mohammadi et al., 1 Aug 2025, Yao et al., 2022). Third, personalization or class-balancing is treated as a first-class design problem. The 2016 system uses class-balanced minibatch sampling, the 2022 model applies subject-specific fine-tuning, and LightSleepNet introduces AdaBN with gradient re-weighting for unsupervised domain adaptation (Tsinalis et al., 2016, Yao et al., 2022, Liao et al., 2024).

Interpretability is most explicit in the 2016 study, where learned filters were shown to recover spectral patterns corresponding to AASM scoring rules, including slow waves, sleep spindles, K-complex-related activity, REM-associated mixed theta, and ambiguity-reducing gaps in the alpha–spindle overlap region (Tsinalis et al., 2016). Later lightweight works focus more heavily on efficiency, cross-subject generalization, or deployment, but the earlier filter analysis remains a benchmark example of physiologically grounded end-to-end representation learning.

Direct performance comparison across all SleepLiteCNN instances is limited by task heterogeneity. EEG models solve five-class sleep staging with 30-second epochs or multi-epoch context, whereas the ECG model solves four-class apnea subtype classification with 1-second decision resolution and 11-second input windows (Tsinalis et al., 2016, Yao et al., 2022, Mohammadi et al., 1 Aug 2025). Evaluation protocols also differ materially: 20-fold subject-wise cross-validation, leave-one-out patient-level validation, 5-fold cross-validation, and held-out test evaluation are all represented in the cited works (Tsinalis et al., 2016, Fernandez-Blanco et al., 2021, Yao et al., 2022, Mohammadi et al., 1 Aug 2025). This suggests that parameter count, accuracy, macro-$(20 \times 30)$2, and hardware efficiency should be interpreted within each sensing and labeling regime rather than as interchangeable indicators of a single benchmark model.

The most defensible encyclopedia-level characterization is therefore that SleepLiteCNN names a compact, raw-signal, sleep-oriented deep-learning lineage with multiple architectural realizations. In some papers it is a pure CNN for EEG staging; in others it is a CNN-Transformer hybrid for edge inference; and in the most explicit recent usage it is a quantized ECG classifier for apnea subtype detection. The continuity lies in computational restraint, end-to-end feature learning, and clinically aligned sleep-analysis objectives, not in a fixed layer blueprint.