StARS DCM Platform: Real-Time Sleep Monitoring

• StARS DCM Platform is a modular hardware and software system that integrates a wearable EEG patch with peripheral biosignal devices for precise sleep-stage decoding and intervention.
• It employs advanced signal processing and deep learning (CNN + Bi-LSTM) to achieve near real-time sleep stage classification with accuracy benchmarks around 88.7%.
• The platform's closed-loop design supports bespoke auditory and thermal interventions, making it ideal for experimental, translational, and clinical sleep research.

The StARS DCM Platform is a modular hardware and software architecture supporting real-time sleep physiology monitoring and intervention, centered around an open-source, wearable EEG patch and integrated peripheral biosignal devices. Developed for precise sleep-stage decoding and closed-loop modulation, the platform combines a high-fidelity, forehead-adhered biosignal device (the DCM patch), a high-throughput real-time messaging framework (ezmsg), and extensible algorithms for real-time signal analysis and intervention. It enables experimental and translational research in neuroscience, sleep engineering, and brain-computer interface (BCI) domains, providing a customizable, synchronized data stream and intervention control stack suitable for both laboratory and field deployments (Coon et al., 3 Jun 2025).

1. Hardware System Architecture

The DCM forehead EEG patch employs an ADS1299-based front end, providing 24-bit precision and supporting 4–16 differential biopotential channels. Modular expansion is achieved by snap-on daughterboards for EEG, EMG, EOG, and ECG acquisition. The device's conformal substrate is implemented with a flexible two-layer PCB (6 mil trace/spacing), designed for adhesion to the forehead or as a foldable stack, approximating the area of two US quarters.

Integrated features include:

• MicroSD storage, multi-color LED status indicator, and an embedded linear resonant actuator for haptic feedback.
• Power via a 300 mAh LiPo coin cell, with ≈8–10 hours of continuous EEG streaming per charge, and 20-minute USB fast-charging with 5 kV isolation for safety.
• Wireless communication over BLE 5.0 via the nRF52840 MCU, supporting multiprotocol operation and NFC-based tap-to-pair authentication for seamless smartphone or hub association.
• Onboard IMU (6-axis), microphone, and ambient light sensors facilitate multimodal context awareness (motion, snoring, environmental light).
• Peripheral integration: Smart rings (PPG for HR, HRV, SpO₂) and thermal actuators (microcontroller-controlled water-cooled bedding or conductive polymer films) synchronize via BLE and ezmsg topics.

All modular electronics and mechanical files are released under open hardware licenses, permitting adaptation for diverse biosignal acquisition scenarios and sensor montages (Coon et al., 3 Jun 2025).

2. Real-Time Software and Data Handling: ezmsg Framework

The core software layer is implemented in the ezmsg pub/sub framework, optimized for sub-millisecond end-to-end latency on consumer hardware. ezmsg nodes interact through named, timestamped message topics, achieving synchronized acquisition, preprocessing, feature extraction, classification, and intervention command streaming.

Key ezmsg nodes:

• Data Acquisition Node: Streams raw biosignals from the DCM at 500 Hz per channel, and from rings (PPG/IMU).
• Synchronization Node: Aligns device timebases using NTP-like BLE/NFC protocols and publishes coherent timestamps.
• Preprocessing Node: Implements DC removal, notch filtering (50/60 Hz), band-pass selection (0.4–5 Hz for slow-wave detection), artifact rejection (±200 µV or excessive IMU activity), and cleans up the raw stream.
• Feature Extraction Node: Computes spectral (Welch PSD across bands δ/θ/α/σ/β), time-domain, and peripheral-derived features (HR, HRV, motion, respiration).
• Sleep-Stage Classifier Node: Implements neural network inference (CNN + Bi-LSTM), producing sliding-window sleep stage outputs.
• Control & Intervention Node: Consumes state predictions to generate real-time commands for effectors (audio, thermal).

The system is compatible with ROS and LabStreamingLayer, supports YAML-based configuration, and can run custom user callback code through the Python API. End-to-end loop latency is empirically measured at ≲ 30 ms: $t_{\rm acq} ≈ 2$ ms, $t_{\rm proc} ≈ 5$ ms, $t_{\rm infer} ≈ 10$–20 ms, and $t_{\rm actuation} ≈ 1$–2 ms (Coon et al., 3 Jun 2025).

3. Signal Processing and Feature Computation

The processing pipeline begins with high-pass and notch filtering to remove drift and power-line contamination. Slow wave activity (SWA) relevant for sleep-stage N2/N3 detection is isolated using a zero-phase FIR band-pass (0.4–5 Hz). Artifact detection excludes samples with high-amplitude noise or movement-induced corruption.

For feature extraction:

• Spectral power is estimated using Welch’s method:

$$\mathrm{PSD}(f)\;=\;\frac{1}{L}\left|\sum_{n=0}^{L-1} x[n]\,e^{-j2\pi fn/L}\right|^2$$

with 1-second windows and 50% overlap.

• Canonical bands (δ 0.5–4 Hz, θ 4–8 Hz, α 8–13 Hz, σ 12–15 Hz, β 15–30 Hz) are computed per 30-s epoch.
• Time-domain features include zero-crossing rates, line length, and envelope variance.
• Peripheral signals: Ring-based HRV (RMSSD, SDNN), respiration rate (mic spectral peaks), and motion count supplement central features (Coon et al., 3 Jun 2025).

4. Machine Learning for Sleep Stage Decoding

Sleep stage classification leverages a hybrid deep neural network:

• Input: 30 s per-channel feature vectors ($\sim$100 features).
• Convolutional blocks (Conv1D, 32 and 64 filters) for local pattern extraction, with batch normalization and dropout.
• Temporal modeling with Bi-LSTM (128 units), followed by dense ReLU and softmax output for 5-way stage classification (Wake, N1, N2, N3, REM).
• Transfer learning: Pretrained on gold-standard PSG EEG; CNN layers are frozen, with Bi-LSTM and classifier fine-tuned using 5–10 nights of subject data from the DCM and ring.
• Optimization: Cross-entropy loss with L2 regularization.

Performance benchmarks illustrate decoding accuracy of 88.7% and macro $F_1$ of 0.85 with DCM + transfer, closely approaching PSG-EEG baselines (93.2%/0.91), while outpacing ring-only scoring (72.5%/0.68) (Coon et al., 3 Jun 2025).

5. Closed-Loop Auditory and Thermal Interventions

Two primary closed-loop effectors are implemented:

• Auditory Slow-Wave Stimulation (aSTIM): During N2/N3, negative EEG deflections (−60 µV, 0.4–5 Hz) trigger 50 ms pink noise bursts at 400 ms after threshold crossing (aligned to up-state), with a 10 s refractory period. A >3-minute sustained N2/N3 state is required for enabling stimulation.
• Dynamic Thermal Modulation: Smart bedding modules receive setpoint updates (e.g., 32 °C→28 °C ramp over 5 min post-Wake→N2 confirmed). PID parameters are user-tunable to maintain surface temperature (stability < 0.2 °C overshoot).

The controller node executes these logic flows in response to real-time state transitions, with all effectors accessible through exposed ezmsg topics (Coon et al., 3 Jun 2025).

6. Modularity, API, and Open-Source Tooling

Hardware modularity is realized via open KiCAD board files and plug-in daughterboard connectors (e.g., custom ADCs, environmental sensors). The Python API (ezmsg) enables streaming, feature extraction, and control customization:

from ezmsg import Node, Topic class MyEEG(app=Node): def __init__(self): self.eeg_topic = Topic("/dcm/clean") @subscribe(self.eeg_topic) def handle(self, msg): # custom processing pass app = MyEEG() app.run()

YAML-based configuration governs sample rates and control thresholds. Command-line tools (stars-record, stars-playback) enable dataset recording and conversion to standardized formats (LSL, CSV) for subsequent analysis (Coon et al., 3 Jun 2025).

7. Experimental Validation and Application Scenarios

A validation protocol with $n=12$ adults demonstrated aSTIM efficacy: δ-power increase of +18% during N3, +25% memory recall improvement, and unchanged arousal index. Example applications include:

1. Full StARS Suite: Multi-modal sleep research with patch, ring, and bedding modules.
2. Low-SWaP Mode: At-home monitoring/intervention using ring plus smartphone.
3. Clinical Pilot: aSTIM in TBI for glymphatic enhancement monitored by CSF biomarkers.
4. Field Deployment: Audio stimulation for restorative sleep compression in operational contexts.

The published stack consistently provides synchronized, extensible, low-latency, and open-access infrastructure for neurophysiological monitoring and intervention (Coon et al., 3 Jun 2025).