Aditya369-WESAD — wrist stress/affect classifier

by AGK FIRE INC

Small 1D-CNN (~225K params) on wrist Empatica E4 signals — 3-class baseline / stress / amusement, leave-one-subject-out evaluated.

Results (LOSO, 15 folds)

metric	mean
accuracy	0.669
macro F1	0.572

subject	acc	macro F1	n_test
S10	0.747	0.590	75
S11	0.895	0.894	76
S13	0.653	0.549	75
S14	0.400	0.217	75
S15	0.707	0.637	75
S16	0.616	0.633	73
S17	0.333	0.254	75
S2	0.718	0.545	71
S3	0.575	0.514	73
S4	0.556	0.578	72
S5	0.767	0.563	73
S6	0.890	0.876	73
S7	0.616	0.529	73
S8	0.730	0.585	74
S9	0.833	0.622	72

Usage

import torch, json
from model import StressCNN

labels = json.load(open("label_map.json"))          # {"1": "baseline", ...}
model = StressCNN(n_ch=6, n_cls=3)
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=baseline, 1=stress, 2=amusement

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
• 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.