This paper is useful because it tests a problem that is very close to the one I am now facing: what happens to wearable physiological data when the person is moving through a real-world environment?

The study is about driving rather than autism or airports, but the methodological problem transfers well. The researchers ask whether wearable signals can distinguish different driving contexts, and they are careful to separate a signal reflecting the driver's physiological response to context from the road environment itself.

That distinction matters for my project. If a wearable changes during a busy airport journey, the signal is not a direct measurement of the airport and it is not a direct measurement of sensory overload. It is a bodily response occurring within a particular context.

What the paper actually tested

Em and Teoh analysed wearable data collected during real-world driving using an Empatica EmbracePlus.

The dataset included:

• 11 drivers;
• 18 driving sessions;
• 1536 non-overlapping 60-second windows;
• three fixed real-world route types: rural, highway and urban.

The wearable recorded physiological, motion and temperature-related information. The paper then tested whether those signals could classify the driving context.

This is important because the study is not asking whether the wearable can diagnose stress or drowsiness. It is asking something more basic: does context leave a detectable signature in wearable data?

The signal reflects the person responding to context

One of the strongest points in the paper is the authors' warning that a wearable signal should not be confused with the environment itself.

A physiological or motion signal reflects the driver's response to the context, not the road environment directly.

For my airport project, the same principle applies.

A change in EDA or heart rate while somebody is moving through security might be associated with the situation, but the signal does not tell me:

• what feature of security caused the change;
• whether the change was sensory, emotional, cognitive or physical;
• whether the traveller experienced it negatively;
• or whether they want support.

This gives me another reason to design around context + uncertainty + user confirmation, rather than trying to classify an internal state.

What looked significant at first became much weaker

At the window level, many of the features appeared statistically different across contexts.

However, the authors argue that the correct statistical unit is not each 60-second window as though it were independent. Windows taken from the same driving session are correlated and therefore behave more like pseudo-replicates.

Once the analysis was corrected to match the experimental structure and multiple comparisons were controlled properly, only 2 of 32 features remained significant:

• skin conductance response count;
• accelerometer movement count.

This is a useful warning for me because wearable datasets can look much more informative than they really are if repeated measurements from the same person are treated as independent evidence.

Important methodological lesson

A large number of sensor windows does not necessarily mean a large number of independent observations. Real-world wearable studies can overstate significance if repeated data from the same person or session are treated as separate participants.

The 90.6% accuracy result is misleading

The most striking comparison in the paper is between two validation methods.

Under a naive, ungrouped cross-validation approach, the model reached around 90.6% accuracy.

That initially looks extremely strong.

But this approach allowed data from the same drivers to appear in both training and test folds, meaning the model could partly learn who the driver was rather than learning a pattern that generalised across people.

When the authors used leave-one-driver-out cross-validation, performance fell to only 34.9% accuracy, with a macro-F1 of 0.33.

The difference was around 55.6 percentage points.

For me, this is one of the most important findings in the paper.

It shows how easy it is for a wearable model to look impressive when it is actually recognising individual-specific patterns.

Personalisation and generalisation are different problems

This result creates an interesting design question for my project.

A system that recognises patterns for one individual over time may be useful even if it performs poorly across different people.

That is different from claiming that one model can understand all autistic travellers.

This distinction fits my emerging direction well.

I am not trying to build a universal classifier that says:

“This physiological pattern means sensory overload.”

A more defensible possibility is a system that learns what is unusual relative to one person's own baseline, while still treating the meaning of that change as uncertain.

So the weak person-independent classification does not make wearable sensing irrelevant. It makes personal baseline and cautious interpretation more important.

No single sensor channel dominated

The researchers also compared different groups of signals:

• physiological;
• accelerometric;
• thermoregulatory;
• and all channels combined.

None clearly dominated.

Importantly, the combined-channel model did not outperform accelerometry alone.

This challenges a simple assumption that adding more sensors automatically creates a better model.

For my own sensing architecture, this is important. I am currently considering EDA, PPG/heart rate and accelerometry, but this paper reminds me that every additional signal needs a reason to be there.

More sensors mean:

• more processing;
• more battery demand;
• more opportunities for artefact;
• more complex interpretation;
• and potentially more burden in the wearable form.

The design question is therefore not “how many signals can I collect?” but “what does each signal contribute that I cannot get more simply?”

HRV features appeared important, but need caution

My notes from the paper highlighted HRV-related measures such as heart rate, inter-beat intervals and measures including SDNN, RMSSD and pNN50.

These appeared statistically informative in parts of the analysis, but many apparent feature differences weakened once the statistical structure was corrected.

That is particularly relevant because I have already been questioning whether HRV or PRV can be interpreted confidently during movement.

This paper does not establish that HRV is useless. It reinforces the need to treat it carefully when the person is moving, and not to isolate one physiological feature from the rest of the context.

Why accelerometry matters to my project

Accelerometer movement count was one of the two features that remained significant after the stricter analysis.

I find that important because accelerometry has a different role from EDA or heart rate.

It is not necessarily trying to infer an internal state.

Instead, it can help the system understand what the body is doing.

For an airport traveller, this could help distinguish between:

• sitting and waiting;
• walking normally;
• rushing;
• climbing stairs;
• carrying luggage;
• standing in a queue;
• or making an abrupt movement.

That does not solve the interpretation problem, but it can help stop the system from treating every physiological change as if it happened under the same conditions.

What this means for my airport system

This paper strengthens the architecture I am moving towards:

physiological change + movement + airport context → cautious interpretation → transparent prompt → traveller decides

It also makes the role of movement data clearer.

Accelerometry may not only be another feature for a classifier. It can help the system understand whether a physiological change occurred while the traveller was physically active or relatively still.

For example:

heart rate rises + accelerometry shows rapid walking

→ physical activity may explain much of the change.

Whereas:

heart rate / EDA changes + accelerometry shows little movement + context shows an increasingly crowded queue

→ the system may have more reason to check in.

Even in the second case, the correct conclusion is still “maybe ask”, not “we know what this person feels.”

Design implication

Movement and context should help the system interpret physiological change, not turn that change into a diagnosis.

What this paper does not establish

This study does not show:

• that wearable signals detect sensory overload;
• that the same patterns would occur in autistic adults;
• that driving contexts are equivalent to airport contexts;
• that one classifier can generalise reliably across people;
• that more sensor channels automatically improve prediction;
• or that a system should automatically intervene when a signal changes.

The strong naive accuracy is actually a warning about how easily classification performance can be overstated.

What changed in my thinking

This paper made me think more carefully about the difference between personalisation and generalisation.

A model may perform poorly when asked to classify a new person, while still potentially noticing meaningful changes within one individual's own data.

That fits the direction I am taking better than trying to build a population-level “overload detector”.

The main takeaway for my project is:

Real-world context and movement affect wearable physiological signals, and once analysis is properly separated across people, classification becomes much weaker.

That supports using movement and contextual information to interpret biosignals cautiously.

It also reinforces the idea that the system should probably be designed around personal baseline + context + a transparent check-in, rather than a universal hidden classifier.

Source

Em, P.P. and Teoh, T.S. (2026). Driving-Context Classification from Wearable Physiological and Motion Signals During Real-World Driving. Sensors. Published 30 September 2026.

Key points from my reading: real-world wearable context classification; 11 drivers and 18 sessions; Empatica EmbracePlus; correction for pseudo-replication and multiple comparisons; skin conductance response count and accelerometer movement count remaining significant; leave-one-driver-out performance of 34.9% accuracy versus 90.6% under naive ungrouped cross-validation; no dominant modality; combined sensing not outperforming accelerometry alone.