DLProjectCNNLSTM/eeg-seizure-project/CNN-LSTM - Storage Bucket

Agentic Active Learning for Annotation-Efficient EEG Seizure Detection

CNN-LSTM Baseline Pipeline on Laptop 2

Autonomous end-to-end baseline implementation for pediatric EEG seizure detection on the CHB-MIT Scalp EEG database using a 1D Convolutional Neural Network combined with a Bidirectional Long Short-Term Memory (CNN-LSTM) recurrent network.

---

1. Project Overview & Objectives

In clinical electroencephalography (EEG), continuous long-term monitoring produces hundreds of hours of multi-channel recordings per patient, but true epileptic seizure activity constitutes less than 1% of the data. Full expert manual review is prohibitively labor-intensive.

This repository establishes the deterministic supervised baseline for "Agentic Active Learning for Annotation-Efficient EEG Seizure Detection", benchmarking model performance when trained on the full eligible training cohort under a strictly disjoint subject-wise partition.

---

2. Remote Data Source & Inventory

The preprocessed multi-channel EEG windows were fetched from Hugging Face at:

• Direct Starting Location: https://huggingface.co/buckets/DLProjectCNNLSTM/eeg-seizure-project/tree/processed/10s
• Parent Repository: https://huggingface.co/buckets/DLProjectCNNLSTM/eeg-seizure-project

Remote Artifact Inventory & Verification Status

File Name	Remote Size	Local Verified Size	SHA/Xet Hash	Status
eeg_windows.npz	6,570,592,069 B (~6.12 GB)	6,570,592,069 B	3aeaa0efc12595ab...	Verified & Downloaded
window_metadata_full.csv	8,153,781 B (7.78 MB)	8,153,781 B	24ade44ac6982ba0...	Verified & Downloaded
processing_manifest.json	766 B	766 B	469e9dc090267cdd...	Verified & Downloaded
window_metadata.csv	1,008,569 B (0.96 MB)	1,008,569 B	69acb6741b4ae5f0...	Verified & Downloaded (Pilot)
download_manifest.json	—	1,994 B	—	Generated locally

---

3. Verified Dataset Specifications

• Total Windows: 38,291
• Window Duration: 10.0 seconds
• Sampling Frequency: 256.0 Hz
• Samples per Channel: 2,560 samples
• Number of Channels: 18 standard bipolar montage channels
• Array Shape: (38291, 18, 2560) of dtype float32
• Integrity: 0 NaN values, 0 Inf values across 1,764,489,216 data points (Global Min: -0.0054, Global Max: 0.0047)
• Channel Montage Ordering (Consistent across all windows):
  1. FP1-F7
  2. F7-T7
  3. T7-P7
  4. P7-O1
  5. FP1-F3
  6. F3-C3
  7. C3-P3
  8. P3-O1
  9. FP2-F4
  10. F4-C4
  11. C4-P4
  12. P4-O2
  13. FP2-F8
  14. F8-T8
  15. T8-P8
  16. P8-O2
  17. FZ-CZ
  18. CZ-PZ

Label Semantics & Distribution

• 0: Non-seizure background EEG (37,548 windows)
• 1: Seizure activity (275 windows)
• -1: Ambiguous / unresolved border transition (468 windows)

Supervised Training Rule: Label -1 windows are strictly excluded from ordinary supervised training loss and binary evaluation, but are preserved in metadata for downstream active learning and uncertainty sampling experiments.

---

4. Subject-Wise Partitioning (Zero Patient Leakage)

To prevent data leakage caused by patient-specific EEG signatures, windows are partitioned strictly by subject:

Partition	Patient Subject IDs	Total Windows	Supervised Windows (0 & 1)	Non-Seizure (0)	Seizure (1)	Ambiguous (-1)	Seizure %
Train	chb01–chb15, chb21 (16 subjects)	28,691	28,616	28,421	195	75	0.68%
Validation	chb16–chb19 (4 subjects)	4,322	4,309	4,270	39	13	0.91%
Test	chb20, chb22–chb24 (4 subjects)	5,278	4,898	4,857	41	380	0.84%
Total	All 24 Subjects	38,291	37,823	37,548	275	468	0.73%

Note: chb01 and chb21 originate from the same pediatric patient and are strictly co-partitioned within the training split. Zero window IDs or subjects cross partition boundaries.

---

5. CNN-LSTM Architecture

The network processes multi-channel 1D time-series data: (batch_size, 18, 2560).

Input Window: (B, 18, 2560)
  │
  ▼ [Conv Block 1]: Conv1d(18 -> 32, k=7, s=2, p=3) + BN + ReLU + MaxPool1d(2)
  Shape: (B, 32, 640)
  │
  ▼ [Conv Block 2]: Conv1d(32 -> 64, k=5, s=2, p=2) + BN + ReLU + MaxPool1d(2)
  Shape: (B, 64, 160)
  │
  ▼ [Conv Block 3]: Conv1d(64 -> 128, k=3, s=2, p=1) + BN + ReLU + MaxPool1d(2)
  Shape: (B, 128, 40)
  │
  ▼ Transpose Channels & Time: permute(0, 2, 1)
  Shape: (B, 40, 128) [40 temporal feature steps]
  │
  ▼ [Bidirectional LSTM]: 2 layers, hidden_size=128, dropout=0.3
  Shape: (B, 40, 256)
  │
  ▼ Global Average Pooling over time: mean(dim=1)
  Shape: (B, 256)
  │
  ▼ Dense Head: Linear(256 -> 64) + ReLU + Dropout(0.3) + Linear(64 -> 2)
  Output Logits: (B, 2)

• Total Trainable Parameters: 715,330
• Loss Function: CrossEntropyLoss with class weights derived strictly from eligible training windows: $$w_0 = \frac{N}{2 \times N_0} = 0.5034, \quad w_1 = \frac{N}{2 \times N_1} = 73.3744$$

---

6. Execution Commands (Windows PowerShell)

Setup Virtual Environment

python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
pip install -r requirements.txt

1. Download Preprocessed Data

python download_processed_data.py

2. Verify Dataset Integrity

python verify_processed_data.py

3. Train Baseline Model

python train.py --config configs/baseline_config.json

4. Resume Interrupted Training

python train.py --config configs/baseline_config.json --resume

5. Evaluate on Held-out Test Set

python evaluate.py --config configs/baseline_config.json --checkpoint checkpoints/best_model.pt

---

7. Actual Experimental Results

Baseline Training Run Summary

• Compute Device: NVIDIA T1000 8GB GPU (CUDA 12.1 via PyTorch 2.5.1+cu121)
• Mixed Precision: Automatic Mixed Precision (torch.amp.autocast)
• Epoch Training Time: ~24 seconds per epoch (448 batches)
• Early Stopping: Triggered after Epoch 14 (Patience: 5 epochs without improvement over Epoch 9)
• Total Training Duration: 338.6 seconds (5.64 minutes)
• Best Validation Metric: PR-AUC = 0.3932 (Epoch 9)
• Best Checkpoint: checkpoints/best_model.pt

Held-out Test Partition Results (chb20, chb22, chb23, chb24)

The optimal classification threshold ($\tau = 0.9790$) was determined strictly on validation data (maximizing validation F1) and applied to the held-out test partition:

Metric	Baseline Score	Notes
Accuracy	99.16%	High baseline due to extreme negative class prevalence
Specificity	99.69%	4,842 / 4,857 non-seizure windows correctly classified (only 15 false alarms)
Precision	50.00%	15 / 30 window detections were true seizure windows
Recall / Sensitivity	36.59%	15 / 41 held-out seizure windows detected
F1-Score	0.4225	Harmonic mean of precision and recall
PR-AUC	0.3351	Over 40× higher than random chance (baseline prevalence 0.84%)
ROC-AUC	0.7850	Strong discrimination across all classification thresholds

Confusion Matrix (Test Set)

                  Predicted Non-Seizure (0)    Predicted Seizure (1)
True Non-Seizure (0)        4,842 (TN)                   15 (FP)
True Seizure (1)               26 (FN)                   15 (TP)

---

8. Generated Deliverable Artifacts

• Environment Report: reports/environment_report.txt
• Download Manifest: data/processed/10s/download_manifest.json
• Dataset Verification Report: data/processed/10s/dataset_verification_report.json
• Split Manifest & Files: data/processed/10s/split_manifest.json, train_split.csv, val_split.csv, test_split.csv
• Best Model Checkpoint: checkpoints/best_model.pt (Epoch 9)
• Latest Checkpoint: checkpoints/latest_checkpoint.pt (Epoch 14)
• Training Logs & CSV: logs/training.log, logs/metrics.csv
• Test Metrics Report: reports/test_metrics.json
• Confusion Matrix Figure: reports/confusion_matrix.png
• Training Curves Figure: reports/training_curves.png
• Model Summary: reports/model_summary.txt

---

9. Active Learning Readiness (Phase 12)

The baseline is architected for seamless integration into active learning experiments:

• Simulation Oracle: Training labels ($y \in {0, 1}$) act as hidden ground truth and are only revealed when queried by an acquisition policy.
• Acquisition Functions:
  • Random Sampling (Passive baseline)
  • Predictive Entropy: $H(y|x) = - \sum_{c=0}^1 p_c \log p_c$
  • Margin Sampling / Least Confidence
• Budgets: Evaluated across 10%, 20%, 30%, 40%, and 50% annotation budgets against this full supervised baseline.
• Fixed Partitions: Validation (chb16–chb19) and Test (chb20, chb22–chb24) subjects remain frozen and are never accessed during the acquisition phase.

---

10. Clinical Limitations & Known Challenges

1. Extreme Class Imbalance: Seizure windows constitute only 0.73% of the dataset. While class weighting enables the model to learn, seizure representations can suffer from high false-negative rates on unseen patients with distinct epileptiform morphology.
2. Patient Heterogeneity: Different pediatric patients exhibit markedly different seizure onset patterns (e.g. generalized rhythmic slowing vs. focal high-frequency spike trains).
3. Window-Level vs. Event-Level Performance: Window-level classification metrics (10-second segments) evaluate isolated window discrimination. They do not directly represent patient-level seizure event sensitivity or clinical alarm false positive rates (FPR/h). In clinical practice, post-processing filters (e.g. collar smoothing, persistence thresholding) are mandatory to avoid duplicate alarms on consecutive seizure windows.