truong_stft_cnn as a real prediction-mode comparison baseline
Follow-on to 2026-08-17's note,
which left dense_edge_gru prediction mode with a completed, non-confounded real run.
This session first made dense_edge_gru and truong_stft_cnn runs
apples-to-apples, then ran truong_stft_cnn's full prediction pipeline for real as a
comparison baseline.
Making the two pipelines comparable
Window length and step size were keyed off which pipeline was running rather than which task
— dense_edge_gru defaulted to its own untuned 4s/8s window instead of
truong_stft_cnn's contiguous 30s/30s, so only about a third of the interictal signal
was being classified and false-alarm rate wasn't measuring the same monitored time across the two.
Now both use 30s/30s whenever the task is prediction mode, regardless of pipeline; detection mode is
untouched.
truong_stft_cnn real prediction run
Full (non-smoke) prediction-mode run, leave-one-seizure-out across chb01's six seizures, as a
baseline to compare dense_edge_gru's numbers against. Completed, exit 0.
| seizure | n_test | preictal | hit | precision | recall | f1 | roc_auc | avg_precision | FAR/hr |
|---|---|---|---|---|---|---|---|---|---|
| 03 | 750 | 30 | ✓ | 0.206 | 0.867 | 0.333 | 0.887 | 0.147 | 13.0 |
| 04 | 650 | 30 | ✓ | 0.169 | 0.733 | 0.275 | 0.871 | 0.150 | 14.3 |
| 15 | 718 | 30 | ✗ | 0.000 | 0.000 | 0.000 | 0.967 | 0.369 | 1.2 |
| 16 | 743 | 23 | ✓ | 0.920 | 1.000 | 0.958 | 1.000 | 1.000 | 0.0 |
| 18 | 750 | 30 | ✗ | 0.000 | 0.000 | 0.000 | 0.979 | 0.504 | 0.0 |
| 26 | 630 | 30 | ✓ | 0.628 | 0.900 | 0.740 | 0.990 | 0.833 | 1.8 |
Event-level hit rate: 4/6 (66.7%), at both raw and smoothed thresholds. Seizure 16 is the standout
— precision 0.92, recall 1.0, zero false alarms — the same seizure that stood out for
dense_edge_gru in the 2026-08-17 run. Seizures 15 and 18 missed at the operating
threshold but still rank well (roc_auc 0.967/0.979) — the same pattern as 2026-08-17's note:
threshold miscalibration on those folds, not an absence of signal. Mean smoothed false-alarm rate
across all six ≈5.1/hour.
Open items
No label-permutation null control has been run yet for this truong_stft_cnn result
— these numbers haven't been checked against a chance baseline the way 2026-08-17's
dense_edge_gru run was. The result files from this run are currently untracked in git
and will be lost on cleanup if not addressed.