Wrist vs. Chest — How Much Does Sensor Placement Change the Answer?

Every experiment on this blog so far has used scalp EEG. This one is a deliberate change of signal, to see whether the pattern I keep finding — pooling people inflates a score, evaluating properly brings it back down — shows up somewhere that isn’t EEG at all.

The data

Schmidt et al., 2018 recorded 15 people wearing two devices at once during a lab protocol with a stress-inducing task and a resting baseline: a RespiBAN chest unit (ECG, EDA, EMG, respiration, temperature, 3-axis accelerometer, all at 700 Hz) and an Empatica E4 wrist band (blood volume pulse at 64 Hz, EDA and temperature at 4 Hz, accelerometer at 32 Hz). Same person, same fifteen minutes, two different sensor placements.

I cut both signals into 30-second windows, kept only windows labeled entirely “baseline” or entirely “stress,” and computed the mean and standard deviation of each channel per window — 12 features for chest, 8 for wrist. Same logistic regression, same subject-wise 5-fold evaluation (GroupKFold by subject) as the EEG experiments.

The result

WESAD baseline vs. stress, subject-wise balanced accuracy
Device	Features	Balanced accuracy	Null (shuffled)
Chest (RespiBAN)	12	0.807 ± 0.066	0.500 ± 0.021
Wrist (Empatica E4)	8	0.694 ± 0.060	0.494 ± 0.010

Both devices clearly separate from chance — this isn’t a case of one sensor working and the other not working at all. But the gap between them is real: 0.114 balanced accuracy, over 11 points, lost by going from the chest rig to the wrist band, on the same subject-wise evaluation that the last two posts used.

Why this matters for the “sense, interpret, adapt” loop

This connects back to the two posts before it, not just by reusing the same evaluation method. A wrist-worn device is the realistic form factor for anything that has to run outside a lab — which means any real adaptive system built on this kind of signal is choosing to pay this 11-point cost for comfort and practicality. Whether that tradeoff is worth it depends entirely on what the system needs to be right about, and that’s a design question, not a modeling one.

Sources & papers I used while learning this

Schmidt et al.2018

Introducing WESAD, a Multimodal Dataset for Wearable Stress and Affect Detection

The dataset used in this post: 15 subjects, chest (RespiBAN) and wrist (Empatica E4) physiological sensors recorded simultaneously during a lab stress protocol.
Read paper