Estimating Stress Levels Using Only Consumer Smartwatches

A validation study of 1 Hz heart rate for binary stress classification

Oleksandr Solovei (No 126784), MSc in Data Science

University of Aveiro, DETI / IEETA

Supervisor: Prof. José Maria Fernandes

July 9, 2026

Good morning, and thank to the jury for being here. My name is Oleksandr Solovei, and this is my Master's dissertation in Data Science, completed under the supervision of Professor José Maria Fernandes.

The work is a validation study: I ask whether meaningful stress information can be extracted from the one-hertz averaged heart-rate stream that a consumer smartwatch exposes — without any access to continuous ECG or raw PPG. Many consumer wearables already provide internal 'stress scores', but those scores are proprietary and not validated against the clinical gold standard. My work tried to remove that black box.

The whole thesis can be compressed into a single sentence, and I'll put it up front so you know where we're going: a one-hertz smartwatch heart-rate stream is too coarse to measure heart-rate variability — but it is not too coarse to classify stress.

Motivation

Why stress, why now

A rising problem

Chronic stress affects 25–40% of working adults (Eurofound 2021), and is causally linked to cardiovascular disease and immune dysfunction (McEwen 2007).

Self-reported daily stress climbed 31% → 43% over the decade to 2020.

A device already on every wrist

Wearables are ubiquitous and always-on: 611 M units shipped in 2025, forecast to grow toward ~690 M by 2030 (IDC), smartwatches a leading segment.

The platform to measure stress at population scale already exists.

Share of employees experiencing a lot of stress, 2009–2020 (Gallup) — Share of employees feeling stress "a lot" the previous day. Source: Gallup.

Worldwide wearables shipment forecast 2026–2030 (IDC) — Worldwide wearables forecast, next years by category. Source: IDC 2026.

Let's talk about what stress is and why it matters. Epidemiological data shows that daily stress was rising even before recent global crises, becoming a defining public health challenge of the 21st century.

The medical literature, McEwen's allostatic-load framework in particular, establishes chronic stress as a causal contributor to cardiovascular disease, immune dysfunction, depression, and early mortality. So, stress is not a soft variable; it is a measurable, expensive physiological problem.

And like many modern problems, it has a modern potential solution: the consumer wearable. The hardware to measure stress at a population scale is already deployed. Global smartwatch shipments continue to break records, with tech companies now leading traditional watchmakers, and the market expanding rapidly across all price tiers. So, we have a real public-health need, we have the device on the wrist, and the only thing standing between them is what the device actually exposes to a researcher. That is the empirical gap this thesis steps into.

The constraint

What 1 Hz Heart Rate Loses

Same 60-second WESAD window, two very different signals

ECG

RMSSD ≈ 66 ms

1 Hz (pseudo)

RMSSD ≈ 2.6 ms

Mean HR level partly survives. Beat-to-beat HRV structure does not.

The core constraint of this work is the signal itself. Garmin and similar devices expose a one-hertz integer BPM stream — no IBI, no raw PPG, no ECG. The one-second averaging acts as a low-pass filter before the data is even sampled, effectively killing the high-frequency respiratory band. This is why every feature I will show today is strictly prefixed with the word 'pseudo'.

Let's look at a concrete example from the WESAD dataset. The green trace is the ECG-derived beat-to-beat heart rate. It clearly retains the rapid variability needed for classical HRV metrics. The red trace is the smartwatch-equivalent one-hertz stream. It preserves the slow mean heart-rate level, but the beat-to-beat structure is entirely flattened. In this specific 60-second window, the true ECG RMSSD is about 66 milliseconds, while the pseudo-RMSSD from the one-hertz stream drops to just 2.6 milliseconds.

This perfectly illustrates why I use the term pseudo-HRV. The formulas may resemble classical HRV, but the input signal is fundamentally different. Therefore, this thesis does not ask whether a smartwatch can replace an ECG. It asks what discriminative stress signal still survives in this heavily constrained one-hertz stream.

Scope

Research question + three hard constraints

Can meaningful stress information be extracted from 1 Hz HR alone, with no ECG / PPG?

Signal

1 Hz averaged HR: the only stream every device exposes.

Validation

Must check against ECG-HRV ground truth (Bland-Altman, ICC, effect size).

Generalisability

LOSO-CV: test on subjects never seen in training.

This brings us to the exact scope of the thesis. The central research question is clear: can meaningful stress information be extracted from a one-hertz heart rate stream alone, without any access to ECG or raw PPG? To answer this without fooling ourselves, I built this work around three hard constraints.

First, the signal constraint. We strictly limit our input to the one-hertz averaged heart rate. Why? Because it is the lowest common denominator — the only data stream that virtually every consumer device on the market guarantees to expose to third-party developers.

Second, the validation constraint. Any claim we make about our pseudo-HRV features must be rigorously evaluated against an ECG ground truth. We do not just look for loose correlations; we rely on Bland-Altman analysis, Intraclass Correlation Coefficients, and effect sizes to measure absolute agreement and true discriminative power.

Third, the generalisability constraint. It is very easy to build a model that looks highly accurate on data it has already seen. To prove this system can work for new users, all classification is evaluated using Leave-One-Subject-Out Cross-Validation. The model is tested exclusively on individuals it has never encountered during training.

These three constraints make the machine-learning task significantly harder, but they ensure that our findings reflect real-world deployment conditions, not just an overfit laboratory experiment.

Position

Related work

Physiology: why HR is the only channel

SAM
sympatho-adrenal

seconds ; HR +15–25 bpm
watch CAN see

HPA
cortisol

15–30 min ; saliva/blood
inaccessible

Prior work, by required signal detail

Gjoreski 2016: raw PPG + EDA + ACC92% (within-subj)

Schmidt 2018 (WESAD wrist): raw BVP + EDA + temp70.0% 3-class

Hernando / Plews: beat-to-beat IBIn/a

Vendor scores (Garmin / Fitbit / Samsung)black-box

This thesis: 1 Hz averaged HR only79.3%

This brings us to the related work, which I've divided into two columns to position our contribution clearly.

On the left, we have the physiology that justifies our exclusive focus on heart rate. The stress response has two main axes. The SAM, or sympatho-adrenal axis, reacts within seconds, driving the heart rate up by 15 to 25 beats per minute. This is exactly the signal a smartwatch can, in principle, see. The HPA axis, on the other hand, operates on a 15-to-30-minute timescale, releasing cortisol that is only measurable in saliva or blood — making it completely inaccessible to a wrist wearable. Therefore, heart rate is the only viable channel.

On the right, prior work is arranged by the level of signal detail required. Gjoreski achieved an impressive 92% accuracy, but used a rich combination of raw PPG, EDA, and accelerometry, and evaluated within-subject — which we know inflates generalisation. Schmidt's foundational WESAD paper reached 70% on a three-class task using raw BVP, EDA, and skin temperature. Other studies rely on beat-to-beat IBI, while commercial vendor scores from Garmin or Fitbit remain proprietary black boxes.

The highlighted bottom row represents the specific gap this thesis fills: using a 1 Hz averaged heart rate as the sole input.

I want to volunteer one important caveat right now, and I will repeat it when the headline number appears: the 79% accuracy I will report later is for a binary baseline-versus-stress task, whereas Schmidt's 70% baseline was a three-class task. They are not directly comparable. I present my result not as a state-of-the-art claim over Schmidt, but as competitive evidence that the severely constrained 1 Hz signal is, in fact, usable.

Data landscape

Available stress datasets

Very few open datasets pair a validated stress protocol with chest ECG and a wrist signal on the same subjects.

Dataset	N	Stressor / protocol	Chest ECG	Wrist signal
SWELL-KW	25	Work overload + interruption (office task)	✓	✗
DriveDB	17	Simulated driving	✓	✗
AffectiveROAD	15	Real-road driving	✗	✓
Nurses' Stress Dataset	15	24 h work shift (ambulatory)	✗	✓
WESAD	15	TSST (validated lab protocol)	✓	✓

WESAD is the only one offering all three on the same subjects: a laboratory-validated stressor, chest ECG ground truth, and a wrist signal. These are the ingredients this thesis needs for both the agreement analysis and the 1 Hz classification task.

Empirical substrate

WESAD dataset

Datasets / protocol	Value
Subjects (S12 excluded)	15 (12 👨 / 3 👩)
Chest ECG (ground truth)	700 Hz
Wrist signal	Empatica E4 BVP, 64 Hz
Stressor	Trier Social Stress Test (validated)
Binary windows (60 s, 30 s stride)	587332 919 total · base / stress
Class balance	64%36%

Backup ; TSST protocol

TSST: Trier Social Stress Test

A validated laboratory stressor (Kirschbaum 1993). Two main phases before an evaluative panel:

Baseline

5–15 min rest, seated

→

Anticipation

speech preparation

→

Mock interview

5 min, free speech, evaluators

→

Serial arithmetic

5 min, 1024 − ... publicly

→

Recovery

rest

Drives reliable +15–25 bpm sympathetic tachycardia + cortisol surge. The cardiac signal the watch can in principle see. Social-evaluative threat → SAM + HPA co-activation.

Backup B4

Why 15 subjects, not 16?

WESAD recorded 16 subjects; S12 excluded here for data quality (sustained ECG / wrist-signal artefact in the relevant windows). 15 usable.

Applied consistently across agreement, effect-size, and LOSO classification analyses.

Feature set

The 8 pseudo-HRV features

60 s windows, 30 s stride. HR → pseudo-IBI: \(\text{IBI} \approx 60000 / \text{HR}\).

Class	Feature
Level	mean HR
Level	mean pseudo-IBI
Dispersion	\(\sigma_{HR}\)
	HR range
	\(CV_{HR}\)
	\(\sigma_{\text{pseudo-IBI}}\) (SDNN proxy)
Dynamics	HR slope
Dynamics	pseudo-RMSSD

"pseudo-" is not cosmetic: these are 1 Hz HR statistics, not clinical HRV (Camm et al. 1996).

The 8 features collapse to ≈ 3 effective axes: level / dispersion / slope. Mean HR vs Mean IBI r = −0.985.

Backup B3

Why 60 s windows, 30 s stride?

Classical HRV uses 5-min stationary records, but at 1 Hz that is 300 samples per window, and the TSST stress phase is shorter than 5 min. 60 s balances: enough beats to estimate dispersion, short enough to track state change. 50% overlap → CIs bootstrap subjects, not windows.

Backup B6

8 features or 3? Gini importance

Level: 43.3%

Mean HR23.26%

Mean IBI19.99%

Dispersion: 50.2%

HR range13.06%

\(\sigma_{HR}\)11.41%

pseudo-RMSSD8.92%

\(CV_{HR}\)8.59%

SDNN proxy (\(\sigma_{\text{pseudo-IBI}}\))8.18%

Slope: 6.6%

HR slope6.59%

Pseudo-RMSSD is informationally inside the dispersion cluster (\(r \approx 0.77\text{–}0.91\) with \(\sigma_{HR}\) / range).

Methods

Method pipeline

Wrist / watch pathway

Empatica BVP

wrist 64 Hz

→

Peak detection

NeuroKit2

→

1 Hz HR downsample

Garmin-equivalent

→

Filters

range / successive-diff / gap

→

60 s windows

30 s stride

→

8 pseudo-HRV

features

→

LOSO-CV

SVM-RBF ; LR ; RF

ECG ground-truth pathway

ECG RR intervals

700 Hz chest

→

True HRV

RMSSD / SDNN / mean HR

→

Agreement

Bland-Altman ; ICC

Within-subject baseline normalisation enters before LOSO: baseline-only → no label leakage. Motion-gating tiers exist for real deployment; every WESAD window is low-motion, so the gate does not touch these numbers.

Statistical rigour

Validation framework

Bland-Altman

bias + limits of agreement; subject-cluster bootstrap CIs (5,000 resamples of subjects, not windows).

Applied per feature, pseudo-HRV vs ECG-HRV. Reports direction and magnitude of bias, not just significance.

ICC(2,1)

absolute agreement; Koo-Li bands (poor < 0.5 / moderate / good / excellent).

Two-way random-effects, single-measures. Penalises both bias and inconsistency between raters/methods.

Wilcoxon signed-rank

on subject means + Bonferroni α = 0.05 / 8 = 0.00625.

Non-parametric, so it makes no normality assumption across the 15 subject-level baseline-vs-stress differences.

Cohen's d

effect size (negligible 0.2 / medium 0.5 / large 0.8).

Computed on the cluster-bootstrap distribution, not pooled raw SD, which avoids inflating effect size from within-subject correlation.

LOSO-CV

subject-independent accuracy + AUC + balanced accuracy.

Three classifiers compared: SVM-RBF, Logistic Regression, Random Forest. Held-out subject never seen in training.

Calibration

Brier Skill Score; isotonic vs raw probabilities.

Checks whether predicted probabilities are trustworthy, not just whether the predicted class label is correct.

Result A, the honest negative

Pseudo-HRV ≠ ECG-HRV

Feature vs ECG	Bias	r	ICC(2,1)
Mean HR	−3.04 bpm	0.73	≈ 0.70 (pooled)
Mean pseudo-IBI	+20 ms	0.76*	≈ 0.74*
\(\sigma_{\text{pseudo-IBI}}\) vs SDNN	large	low	0.02
pseudo-RMSSD vs RMSSD	> 120 ms LoA	0.13	0.00

* mostly arithmetic inverse of HR; not an independent HRV signal

For absolute HRV measurement, the 1 Hz stream fails, and we can say exactly by how much.

This is the slide that prevents this thesis from over-claiming. I want to deliver it plainly: for absolute Heart Rate Variability measurement, the one-hertz stream fails.

Let's look at the table. For Mean HR, we see a slight under-reading — a bias of minus 3.04 bpm — and a pooled ICC of 0.70. This looks like a partial win. However, I want to volunteer a critical nuance: if we strip out the between-subject variance, that ICC drops to a per-subject median of just 0.40. This means the pooled 0.70 mostly reflects that the watch correctly ranks different people's baseline rates, not that it perfectly tracks a single person's window-to-window state changes.

But the real collapse happens in the highlighted rows: the variability proxies. When we compare pseudo-SDNN and pseudo-RMSSD against true ECG ground truth, absolute agreement vanishes. The ICCs drop to 0.02 and a literal 0.00. For pseudo-RMSSD, the limits of agreement span 120 milliseconds, which swallows the entire physiological range.

So, why is this a contribution? Because instead of vaguely stating that 'consumer watches aren't perfect,' we are quantifying exactly how much they fail, feature by feature, using rigorous cluster-bootstrapped confidence intervals.

But this leaves us with a fascinating paradox. Pseudo-RMSSD has zero absolute agreement with clinical RMSSD. Yet, as I will show on the next slide, it still has a significant statistical effect size for separating stress from baseline. Let's look at how both of these things can be true at once.

Backup B1

Pooled ICC 0.70 but median 0.40: why?

The pooled ICC absorbs between-subject HR spread (different people sit at different resting HR).

It reflects ranking different people similarly, not tracking within-person state changes. Strip the between-subject variance and the median per-subject ICC falls to 0.40.

So even mean HR is not window-level interchangeable with ECG.

Backup B12

What could a better wrist signal do?

Wrist PRV from the raw 64 Hz Empatica signal (not downsampled):

Metric (wrist PRV vs ECG)	r	Band
Mean HR	0.92 (vs pseudo 0.73)	Strong
SDNN	0.47–0.53	Moderate
RMSSD	0.47–0.53	Moderate

A 64 Hz wrist signal partly recovers variability, but Garmin does not expose it to third parties.

Result B, the positive

Discrimination survives

7 of 8 features pass Bonferroni-corrected Wilcoxon.

Feature	Δ stress − base	\|d\|
Mean HR	+11.4 bpm	1.13
HR range	↑	1.04
\(\sigma_{HR}\)	↑	0.98
\(CV_{HR}\)	↑	0.83
Mean pseudo-IBI	−66 ms	1.08
\(\sigma_{\text{pseudo-IBI}}\)	↑	0.63
pseudo-RMSSD	↑	0.60
HR slope	≈ 0	0.05

Ok, pseudo-HRV fails as a clinical measurement. But this slide proves why it still succeeds as a stress classifier. Poor absolute agreement does not imply the absence of a stress signal.

When we look at discriminative power, seven of our eight features pass the strict Bonferroni-corrected Wilcoxon test. And crucially, these tests were computed on subject-level means — fifteen genuinely paired observations, completely avoiding the trap of pseudo-replicating 919 windows.

The effect sizes, measured in Cohen's d, are large across the board. Mean HR shows a massive stress shift of plus 11.4 beats per minute, yielding an effect size of 1.10. Heart rate dispersion metrics, like standard deviation and range, also exceed 0.95.

There is only one null feature, highlighted in red: the HR slope, with an effect size of just 0.05. But this null result makes perfect physiological sense. Sympathetic activation triggers a state-level shift in heart rate, not a sustained monotonic ramp. Inside a 60-second window, the slope simply oscillates and averages to near zero.

But the most important takeaway is the paradox I promised you. Look at the pseudo-RMSSD and pseudo-SDNN proxies. They show solid medium-to-large effect sizes of 0.60 and 0.63. Yet, two minutes ago, I showed you they have an ICC of absolutely zero against the ECG ground truth.

How can a feature that is entirely broken as a clinical measurement still reliably discriminate stress? That paradox is the intellectual center of this thesis, and the next slide explains exactly what the smartwatch is actually measuring.

Conceptual centrepiece

Why both are true: the paradox

Agreement and classification answer different questions.

"Do the two numbers match?"

Absolute match: FAIL ×

pseudo-RMSSD vs ECG RMSSD: r = 0.13, ICC = 0.00

"Does it move the right way within a person?"

Within-person direction: PASS ✓

pseudo-RMSSD stress vs baseline: |d| = 0.60

At 1 Hz, pseudo-RMSSD physically measures within-window HR dispersion, not vagal tone, which is why it tracks σ_HR (r ≈ 0.8), not ECG RMSSD.

We are not claiming to measure HRV; we are detecting the direction of sympathetic arousal.

This brings us to the conceptual centrepiece of the thesis: the resolution of the paradox.

On Slide 11, we saw that pseudo-RMSSD has an ICC of strictly zero against true ECG RMSSD. Yet on Slide 12, it discriminates stress with a significant effect size of 0.60. How can a feature that completely fails as a measurement still succeed as a predictor?

The resolution lies in understanding that agreement and classification ask two fundamentally different questions.

The left box asks the agreement question: 'Do these two numbers match?' For pseudo-RMSSD versus clinical RMSSD, the answer is a definitive no. There is no absolute measurement relationship.

The right box asks the classification question: 'Does this feature consistently move in the right direction within a single person when stress is induced?' The answer is yes.

These questions are mathematically and physically independent. The physical reason is locked in the card at the bottom. At one hertz, the beat-to-beat structure that defines parasympathetic vagal tone has been irreversibly averaged away by the watch. Therefore, pseudo-RMSSD is physically incapable of measuring vagal tone.

What it actually measures is within-window heart rate dispersion — the amplitude of fluctuation in the smoothed signal. So, when I report a stress effect for pseudo-RMSSD, I am not claiming to measure parasympathetic withdrawal. I am detecting the direction of sympathetic arousal, encoded as increased dispersion.

This distinction is the ultimate defence of scientific honesty in this work. Now, let's see how well this sympathetic detection generalises to unseen individuals in the machine learning pipeline.

Backup B11

Why keep pseudo-RMSSD at all?

Discriminative at window level (d = 0.60) but informationally redundant with σ_HR / HR range (r ≈ 0.77–0.91).

Dropped for deployment at only −1.1 pp cost. Kept in the paper to complete the agreement story: without it the "pseudo" honesty argument has a hole.

Result C, the headline

LOSO classification

Model	Features	Balanced acc.	AUC
SVM-RBF	8	79.32% ± 16.32	86.23%
Logistic Regression	8	79.31%	86.82%
Random Forest	8	75.06% ± 16.89	85.41%
SVM-RBF (drop pseudo-RMSSD)	7	78.18%	86.55%
Single ΔmeanHR only	1	77.0%	84.4%
+1 dispersion feature	2	78.0%	86.03%

Most of this is a calibrated tachycardia detector. The other five features add ≈ 1 pp. Context: WESAD wrist 70.0% (3-class), not like-for-like.

Here is the headline result, and I want to attach the caveats right at birth.

The highlighted row shows our top performer: an SVM with an RBF kernel reaches 79.3% balanced accuracy and an 86.2% AUC. This is evaluated under strict Leave-One-Subject-Out cross-validation, meaning the classifier is continually tested on individuals it has never seen.

But I want to unpack the story behind this number, starting with the Logistic Regression row. As you can see, it is essentially tied with the SVM. This near-tie is a crucial finding in itself: it tells us that after within-subject normalisation, the decision surface becomes almost perfectly linear. The signal is dominated by a few strong features, not complex non-linear interactions.

In fact, we can break this down even further. A single feature — the within-subject normalized mean heart rate — achieves 77% accuracy all on its own. Adding just one dispersion feature brings it to 78%. The remaining five variability proxies combined add barely one percentage point.

So, to volunteer a phrase before the jury offers it: this pipeline functions fundamentally as a calibrated tachycardia detector. It detects a highly personalized heart-rate rise combined with a slight increase in signal dispersion. And this is exactly what the physiological effect sizes on Slide 12 predicted.

Finally, a note on context. As I promised earlier, we must compare this to the original WESAD wrist benchmark from Schmidt, which achieved 70%. Because their task was three-class and mine is binary, they are not directly comparable. I do not present this 79% as a state-of-the-art claim to beat the benchmark. I present it as conclusive evidence that the severely degraded one-hertz smartwatch signal is, in fact, highly usable for stress classification.

Backup B5

One feature → 77%

A single within-subject z-scored mean-HR feature already reaches 77.0% BA / 84.4% AUC. Add one dispersion feature → 78.0%. The five IBI/variability features together add ≈ 1 pp.

Backup B13

How does this compare?

Work	Signal	Task	Eval	Score
Gjoreski 2016	PPG+EDA+ACC	binary	within-subj	92%
Schmidt 2018 (WESAD)	BVP+EDA+temp	3-class	LOSO	70.0%
This thesis	1 Hz HR only	binary	LOSO	79.3% BA

Different task, different signal, different evaluation: not a direct comparison, only context.

The biggest lever

Personalisation > model choice

Same SVM, same 7-feature set, same LOSO folds. Only the z-score reference used to normalise HR changes.

Population baseline
z-scored vs all subjects pooled

73.76% BA
80.41% AUC

Own baseline
z-scored vs that subject's rest

78.18% BA
86.55% AUC

+4.4 pp BA, +6.1 pp AUC from switching the reference alone: bigger than the entire SVM-vs-Random-Forest gap (4.3 pp).

"Is 90 bpm high?" → "Is this 10 bpm above YOUR baseline?"

Person A rests at 60, stresses to 70. Person B rests at 80: same absolute 70-ish, opposite states. Pooled normalisation can't tell them apart; own-baseline normalisation can.

Deployment consequence: every real user needs a rest-baseline calibration step before classification can run at all.

≈ 5 min proposed (low end of WESAD's 5–15 min rest), but not empirically validated as a minimum; flagged as open.

This slide elevates a design choice to a primary finding, and it is arguably the most actionable result for any product team. The two bars compare our normalisation strategies.

On top, population-level normalisation, where we z-score features against the entire cohort mean, yields 74.9% balanced accuracy. Below it, within-subject normalisation — z-scoring against the user's own baseline — jumps to 79.3%.

That 4.4-point gain is actually larger than the entire performance gap between the best and the worst classifier I tested. This tells us that the choice of normalisation matters far more than the choice of model.

The reason is captured in the card. Population normalisation asks, 'Is 90 bpm high in general?' — which is a broken question, because the answer depends entirely on who you are. Within-subject normalisation asks, 'Is this 10 bpm above your specific baseline?' — and that is a well-defined physiological state.

Concretely: Person A rests at 60 and stresses to 70; Person B naturally rests at 80. An absolute value of 70 means 'stressed' for the first person and 'relaxed' for the second. Without within-subject normalisation, the classifier cannot resolve this ambiguity.

The deployment consequence is simple: every real-world user must undergo a baseline calibration phase before the system can function. I propose a five-minute rest period as a starting point, but I want to be intellectually honest here: the minimum duration for this calibration has not yet been empirically validated, and it remains an open question for future work.

However, even with this personalisation, we see a massive variance in performance across individuals. Let's look at why that is the central failure mode of this approach.

Honest failure mode

Per-subject heterogeneity

Per-subject balanced accuracy ranges 42.11% (S2) → 92.83% (S11), a 50.7 pp spread, the central empirical finding.

11 / 15 above 70%.
4 near or below chance (S2, S9, S15, S6).
~25% of users would get unreliable output with no warning.
S2 AUC = 55%: the signal itself is weak, not a threshold problem (no model recovers what the stream doesn't contain).

Plausibly ~15–20% of healthy people are blunted responders (Dickerson 2004).

This brings us to the honest failure mode of the thesis. While our aggregate accuracy is promising, it hides a serious tail.

As you can see on the chart, per-subject balanced accuracy ranges from 42.1% for Subject 2, all the way up to 92.8% for Subject 11. That is a massive 50.7 percentage-point spread, and I consider this the central empirical finding of the thesis — even more important than our 79% headline.

Eleven of the fifteen subjects exceed the 70% threshold. However, four subjects fall near or below chance. This means roughly a quarter of users would receive unreliable outputs from a deployed system, with no internal warning.

Subject 2 perfectly illustrates why. Their per-subject AUC is just 55%. This means the signal itself is weak. It is not a miscalibrated decision threshold; no machine learning architecture can recover information that the stream simply does not contain. The medical literature, such as Dickerson's 2004 meta-analysis, suggests that about 15 to 20 percent of healthy people are simply 'blunted responders' to the TSST. Their physiology does not produce the expected tachycardia, making them essentially invisible to this classifier.

Finally, I must explicitly state the fairness limitations of this work. PPG sensing quality is known to systematically degrade on darker skin tones. Furthermore, cardiac reactivity varies wildly with age, sex, fitness, and medications — a user on beta-blockers, for example, would silently fail this classifier. Because the WESAD cohort is overwhelmingly young and male, algorithmic fairness across a broader population cannot be claimed here. I am flagging this as a strict threat to validity.

Optional ; Slide 16-A

Confidence cross-check

Calibration + ARIMA confidence

Probability calibration

Raw SVM probabilities already usable: Brier Skill Score 0.397; isotonic → 0.403 (negligible; a safety measure, not a skill gain).

Enables abstention: flag 0.4 < p̂ < 0.6 as "uncertain".

ARIMA cross-check

Forecast each subject's baseline HR into stress; count windows outside 95% PI → anomaly detection rate (ADR).

Per-subject ADR vs SVM accuracy: r = 0.742.

Responder type	ADR	SVM acc.
Strong	> 50%	87.9%
Weak	≤ 10%	52.2%

A 35.7 pp gap: a confidence signal derivable from baseline HR alone. N=15, 95% CI [0.37, 0.91]; internal consistency check, not independent validation.

Backup B8

How is ADR computed?

Per subject: fit ARIMA on baseline HR, forecast into stress period, count 60-s windows whose mean HR falls outside the 95% prediction interval → ADR.

Also a directionality finding: HR stabilises against anticipatory drift once stress begins.

Backup ; proposed gate

Proposed confidence gate

Two branches over the same 1 Hz HR stream:

calibrated stress probability from the classifier;
baseline ARIMA anomaly flag from the 95% prediction interval.

This is a deployment proposal motivated by WESAD, not an on-device validated feature.

Engineering contribution

System: platform, pipeline

Beyond the lab: enables scalable, remote stress-induction without laboratory infrastructure.

Pipeline smoke test

Real Garmin data recorded, synchronised, windowed, and scored without numerical errors.

N=1 classifier result

Gaming session scored P(stress) ≈ 0.18 — the model saw no shift from the idle baseline.

This slide transitions from benchmark validation to real-world engineering. Here, I present the end-to-end system built to bring this classification pipeline out of the laboratory.

First, the stress-induction platform. I built a custom web application featuring six cognitive games. Each game is explicitly designed to trigger a validated stress mechanism, such as time pressure or working-memory overload, while logging every interaction with millisecond precision.

On the hardware side, the pipeline is fully operational. Real Garmin data is recorded, synchronized, windowed into 60-second segments, and scored by our WESAD-trained SVM without any numerical errors. As the green card indicates, this pipeline smoke test passed completely. The engineering architecture works end-to-end.

However, we must look at the red card: the N=1 classifier result. During a single self-experimentation session using a Garmin Fenix 8, the SVM produced a stress probability of approximately 0.18 for both the baseline and the game windows. It found no separation.

I want to frame this exactly as the thesis does: this is a smoke test that passed, not an accuracy claim that failed. This null result is perfectly consistent with three physical and methodological factors: First, Integer BPM Quantization. Garmin devices export heart rate as whole numbers. At 75 beats per minute, a 1-bpm shift equals a massive 10.7-millisecond jump in the inter-beat interval space. This hardware quantization physically crushes the micro-variability features the SVM partially relies on. Second, Unverified Stress Induction. Because this was a self-experiment, I knew the mechanics of the games, and I collected no self-report data to confirm that acute stress was actually triggered. Third, a Short Baseline Anchor. I used only a two-minute pre-game rest period, which is far too short to establish a stable within-subject normalization reference compared to WESAD's 15-minute baseline.

The key takeaway is that the bridge from benchmark to real device is built. The pipeline is fully ready for a powered, multi-participant study upon ethics approval.

Boundaries

Ethics, GDPR & limitations

Art. 9 special category data

HR/HRV is health data under GDPR. A 4-point consent gate (privacy policy, storage, age, voluntary) blocks every game until confirmed. No automated clinical decisions (Art. 22). DPIA drafted.

GDPR posture

Local-first, verified in code: the platform makes zero network calls; interaction logs (timestamps, answers, levels) stay in browser localStorage until manually exported to JSON.

One-click revoke wipes every stored key instantly.

Current state

Only WESAD + author's own N=1 data used so far: no third-party participants.

A fuller multi-participant consent framework (pseudonymisation, EEA storage, retention schedule) is already drafted.

Consent screen 1 from the stress-induction platform

Consent screen 2 from the stress-induction platform

Live consent gate: privacy policy, storage, age, and voluntary-participation checks, enforced before any game starts.

This brings us to the ethical boundaries and limitations of this work.

Because heart-rate data is classified as Article 9 special-category health data under the GDPR, our engineering posture must be strictly privacy-preserving. The platform enforces a four-point consent gate before any game begins, and it operates entirely on a local-first architecture. Zero network calls are made; all interaction logs remain strictly in the browser's local storage until manually exported to JSON, and we ensure no automated clinical decisions are made, completely satisfying Article 22.

To be absolutely clear about the current state: no third-party participants have been recorded. The thesis relies exclusively on the public WESAD dataset and my own self-experimentation data. A full Data Protection Impact Assessment and multi-participant consent framework are already drafted, awaiting submission to the University of Aveiro's ethics committee for future studies.

Finally, I want to state our most important limitation out loud, particularly for the medical members of the jury. This classifier detects sympathetic cardiac arousal — not strictly 'psychological stress' per se. The exact same heart-rate signature — an elevated rate with increased dispersion — is produced by physical exertion, caffeine, fever, or simply walking up a staircase. The WESAD protocol controls for these confounds by keeping subjects seated and sober. A free-living deployment would not have those luxuries, and would absolutely require accelerometer and context fusion to disambiguate those states.

Stating this limitation clearly is, I believe, the only intellectually honest way to interpret the signal.

Answer to the research question

Verdict & what's next

Yes, with qualifications.

79.3% Balanced acc.

86.2% AUC

WESAD-derived 1 Hz signal, LOSO-CV, with per-user calibration

Conditions attached

Within-subject normalisation
Subject-independent (LOSO) evaluation
~25% need extra personalisation
Not a substitute for clinical ECG-HRV
Real-Garmin transfer still to validate

Future work

Multi-modal fusion (ACC, skin temp, EDA)
N-of-1 continuous personalisation
Full CEIC-UA study (10–20 participants)
On-device linear SVM (Monkey C)

Smartwatch HR is too coarse to measure HRV, but not too coarse to support calibrated stress-state classification.

Can meaningful stress information be extracted from the coarse one-hertz stream of a smartwatch?

The answer is yes, with important qualifications. It requires:

Within-subject normalisation.
Subject-independent evaluation.
Designing for the 25% low-performance tail.
Explicitly acknowledging this does not measure clinical vagal tone.
Separate hardware validation to overcome integer quantization.

These define possible future work: multi-modal fusion, N-of-1 continuous learning, and on-device Monkey C deployment.

I will close by reading the final sentence verbatim: Smartwatch heart rate is too coarse to measure HRV, but not too coarse to support calibrated stress-state classification.

Thank you for your time.

Thank you

Questions welcome.

github.com/s126784/dis_companion_code

Oleksandr Solovei ; University of Aveiro, 2026

Acknowledgements: Prof. José Maria Fernandes ; IEETA ; WESAD authors (Schmidt et al., 2018)