Aditya369-WESAD-Binary — wrist stress detector

by AGK FIRE INC

Small 1D-CNN (~225K params) on wrist Empatica E4 signals — BINARY stress vs non-stress (baseline + amusement merged, the WESAD paper's binary setup), leave-one-subject-out evaluated.

Results (LOSO, 15 folds)

metric	mean
accuracy	0.871
binary F1	0.781

subject	acc	F1	n_test
S10	1.000	1.000	75
S11	0.737	0.474	76
S13	0.987	0.978	75
S14	0.533	0.222	75
S15	0.853	0.744	75
S16	0.973	0.952	73
S17	0.547	0.292	75
S2	0.930	0.865	71
S3	0.658	0.419	73
S4	1.000	1.000	72
S5	0.973	0.950	73
S6	1.000	1.000	73
S7	0.945	0.913	73
S8	0.932	0.902	74
S9	1.000	1.000	72

Usage

import torch, json
from model import StressCNN

labels = json.load(open("label_map.json"))          # {"0": "non-stress", "1": "stress"}
model = StressCNN(n_ch=6, n_cls=2)
model.load_state_dict(torch.load("best_model.pt", map_location="cpu"))
model.eval()
# x: (1, 6, 1920) float32 — 60s window @32Hz, per-subject z-scored,
#    channels [BVP, EDA, TEMP, ACCx, ACCy, ACCz]
with torch.no_grad():
    pred = model(x).argmax(1).item()                # 0=non-stress, 1=stress

Training

• Data: WESAD, wrist only (BVP 64Hz, EDA/TEMP 4Hz, ACC 32Hz), resampled to 32Hz
• 60s windows, 50% overlap, per-subject z-norm, majority-vote labels
• Binary labels: stress (WESAD 2) -> 1; baseline (1) + amusement (3) -> 0
• Leave-one-subject-out, early stopping on a val split of train subjects
• best_model.pt = state_dict of the best fold

---

© 2026 AGK FIRE INC. Released under Apache 2.0.