Wrist-Worn Sleep Detection & Deep Learning Staging

Sleep Detection & Staging Using Wrist-Worn PPG + Accelerometer Data 1. Objective We are seeking an engineer or researcher to implement a state-of-the-art sleep detection and staging system using PPG and 3-axis accelerometer data from a wrist-worn device. The system should evaluate and compare leading deep learning algorithms, such as PPG-only models (e.g., SleepPPG-Net2, κ ≈ 0.75, high generalization across datasets) and PPG+ACC temporal sequence models (e.g., Olsen et al. U-Net-style, κ ≈ 0.64, validated for night-level metrics). The comparison should be based on both staging accuracy and practical usability for wrist-based wearable signals. The approach must incorporate transfer learning to adapt pre-trained models to real-time raw wrist data, enabling robust performance under diverse users, motion conditions, and signal quality variations. 2. Scope of Work The work includes designing a signal preprocessing pipeline for resampling, noise filtering, motion artifact suppression, and Signal Quality Index (SQI) computation to ensure only reliable data segments are used. Processed data will feed into deep learning models capable of classifying standard sleep stages (Wake, N1, N2, N3, REM) and producing validated night-level metrics: Total Sleep Time (TST), Sleep Efficiency (SE), Sleep Onset Latency (SOL), and Wake After Sleep Onset (WASO). Algorithm selection must be justified in terms of per-epoch accuracy, model generalization, and ability to provide full-night behavioral metrics. Where possible, training should leverage multi-dataset pretraining and fine-tuning to address wrist-specific PPG and accelerometer limitations. 3. Validation and Performance Targets Validation must use subject-wise data splits (with no overlap between training and testing subjects) to ensure realistic generalization. Performance reporting should include Cohen’s κ, macro and per-class F1 scores (with particular emphasis on REM and N3 stages), confusion matrices, and night-level metric errors. The system is expected to achieve at least Cohen’s κ ≥ 0.62 (with a target of ≥ 0.70 in optimal conditions), macro F1 ≥ 0.60 with REM F1 ≥ 0.50 and N3 F1 ≥ 0.55, and median absolute errors within ±20 minutes for TST, WASO, and SOL, and within ±5 percentage points for SE. Additionally, ablation studies should be performed to compare PPG-only versus PPG+ACC inputs, with versus without SQI, and short versus long context sequences, to determine the most effective configuration. 4. Deliverables The final handover should include: 1. Codebase — fully functional training and inference scripts (Python, TensorFlow) with clear directory structure. 2. Preprocessing Module — resampling, noise/artifact handling, SQI computation, with unit tests. 3. Trained Model(s) — checkpoints and configuration files for all tested architectures. 4. Evaluation Report — covering all required metrics, confusion matrices, ablation study results, and discussion of trade-offs between algorithms. 5. Integration Guide — instructions for running preprocessing, training, and inference on sample data, with input/output specifications. 6. Sample Results — per-epoch predictions, confidence scores, and computed nightly metrics for a small test dataset. Ref: 1. A Flexible Deep Learning Architecture for Temporal Sleep Stage Classification using Accelerometry and Photoplethysmography [login to view URL] 2. SleepPPG-Net2: Deep learning generalization for sleep staging from photoplethysmography [login to view URL]

Project ID: 39694865