This paper caught my attention because it is very close to the system architecture I am now considering. Rather than relying on physiology alone, Heng and Yavari combine environmental conditions, movement/context and physiological data in an attempt to model personal comfort in real-world conditions.

That is directly relevant to my emerging airport proposition. I am currently asking whether physiological change becomes more useful when it is interpreted alongside what the traveller is doing and what is happening around them, rather than treating a heart-rate or electrodermal change as meaningful on its own.

The paper is useful partly because its results are not especially strong. It shows that building a multimodal system is technically possible, but it also shows how difficult it is to infer a person's state reliably once sensing moves out of the laboratory.

What the paper actually tested

Heng and Yavari developed a Comfort Framework that combines three layers of information:

1. Macro-environmental conditions from location-based APIs.
2. Kinematic and micro-environmental context from smartphone and wearable sensors.
3. Physiological responses from wearable sensing, including heart rate and heart-rate variability.

The framework was intended to cover thermal, acoustic, visual and air-quality conditions. However, the pilot only provided meaningful empirical validation for thermal and acoustic signals.

The study was also very small: it was a single-subject pilot involving 18 free-living sessions in Melbourne. The system produced 418 labelled 60-second windows, with 10 sessions contributing valid leave-one-session-out evaluation folds.

This means I need to treat the work as a proof of pipeline and methodological exploration, not evidence that the model generalises to other people.

Why the real-world setting matters

One of the most useful parts of this paper is that the authors deliberately moved away from a controlled chamber.

In a real environment, physiology is being affected by many things at once. Walking, posture, temperature, noise, movement, location and ordinary activity can all alter the signals a wearable records.

The authors therefore included kinematic context rather than pretending the physiological signal could be interpreted in isolation.

This is particularly relevant to an airport. A traveller may be walking quickly, carrying luggage, queuing, sitting, rushing after a gate change or standing still at security. Any of those activities could affect heart rate or HRV without meaning that the person needs sensory or accessibility support.

For my project, this strengthens the idea that movement data is not just another sensor channel. It may be necessary for interpreting whether a physiological change is even unusual in context.

The prediction performance is an important warning

The strongest Random Forest model achieved an F1 macro score of 0.456 ± 0.151 across the valid evaluation folds.

That is modest performance.

I find this result more useful than a much higher laboratory accuracy would have been because it makes the limitations of real-world physiological inference visible. The comfortable and uncomfortable classes overlapped substantially even though there were dataset-level differences in measures such as heart rate, RMSSD and wrist temperature.

The paper therefore does not show that combining enough sensors allows a computer to reliably understand how somebody feels.

Instead, it shows that multimodal sensing can reveal useful patterns while uncertainty remains high.

Key lesson for my project

Putting physiology, movement and environmental information together does not remove ambiguity. A multimodal system may have more context, but that does not make its interpretation certain.

The person still provided the label

Another detail I think is important is how the researchers established whether a period was “comfortable” or “uncomfortable”.

The system did not independently discover the person's comfort state. The participant labelled it using a minimal-disruption button press on a smartwatch.

That distinction matters for my project.

The sensors produced information about physiology and context, but the person remained the source of meaning against which the system was evaluated.

This reinforces my current position that the autistic traveller should remain the final interpreter. Biosensing may contribute to deciding when it could be useful to ask, but the system should not claim that it already knows what the traveller is experiencing.

Context can also create false confidence

The authors found that some apparently useful features were actually acting as session or location identifiers rather than genuine comfort signals. Features such as altitude, phone gravity components and individual magnetometer axes had to be removed because the model risked learning environmental fingerprints instead of comfort.

This is particularly important for an airport system.

A model could appear to learn that a person becomes uncomfortable at security, at a particular gate or in a particular terminal, when it may actually just be learning the distinctive sensor pattern associated with that location.

That means adding context is not automatically better. The system still needs to distinguish between:

context that helps interpret the person

and

context that simply makes the situation easy for the model to recognise.

For me, this is another reason to avoid designing around a hidden automated classification score. The interaction should communicate uncertainty rather than conceal it.

What signals appeared useful

After the authors removed features that were likely to produce leakage, the strongest contributors included:

• acoustic features;
• HRV-related features;
• wrist skin temperature;
• heart rate;
• wrist movement;
• smartphone kinematic features.

This does not mean these are the “best sensors” for my airport project. The pilot involved one person and measured comfort, not autistic sensory overload or support need.

However, it does strengthen the general architecture I am considering:

physiology + movement + environment/context

rather than physiology alone.

It also supports my decision to keep movement sensing alongside candidate physiological channels, because activity is a major confound in free-living data.

Air quality and environmental sensing

The framework was designed to include thermal, acoustic, visual and air-quality conditions. The authors discuss environmental factors such as particulate matter, CO₂ and other air-quality conditions within the wider comfort problem.

However, the pilot did not validate localised indoor air quality or the visual domain. The authors identify additional light and wearable air-quality sensing as future work.

For my project, air quality could therefore remain part of the wider contextual design space, particularly in enclosed airport environments, but I should not treat this paper as evidence that air-quality sensing improved the model.

What this means for my airport system

This paper gives me one of the clearest technical precedents so far for the architecture I am moving towards.

A simplified airport version could be:

physiological change + movement/activity + airport/environmental context → cautious interpretation → transparent check-in → user decision

The value of the sensing is not that it gives the airport or the wearable access to a hidden internal truth.

Its possible value is that several weak pieces of information may together provide a reason to check in at a useful moment.

For example, a physiological change occurring while the traveller is stationary in a suddenly crowded, noisy gate area means something different from the same physiological change while they are running to catch a connection.

Even then, the system still does not know whether support is wanted.

A prompt such as:

“A change was detected while this area became busier. Would support help?”

fits the evidence much better than:

“You are becoming overloaded.”

What this paper does not establish

This study does not demonstrate:

• that physiological signals detect autistic sensory overload;
• that multimodal sensing can reliably infer anxiety, stress or dysregulation;
• that the framework generalises beyond the single participant;
• that a contextual biosensing system improves airport accessibility;
• that air-quality or visual sensing improved the pilot model;
• or that a person should receive support automatically when a model detects change.

The modest classification performance is actually a reason to keep these limitations visible.

What changed in my thinking

Before reading this paper, I already thought context would be important. This study makes that argument much more concrete.

The strongest takeaway for me is not “use more sensors.”

It is:

Real-world physiology only becomes interpretable when movement and environment are considered, and even then the interpretation remains uncertain.

That strengthens my emerging contextual-airport direction, but it also makes the interaction principle clearer.

I am increasingly thinking that the most defensible role for biosensing is timing and awareness, not state classification.

The system might notice that something has changed and have enough context to make a check-in more relevant, but the traveller should still decide what that change means and whether any support is useful.

Source

Heng, S. and Yavari, A. (2026). Multimodal Fusion of Environmental and Physiological Data for Real-World Personalised Comfort Modelling. Sensors, 26(10), 2940.

Key points from my reading: three-layer environmental/kinematic/physiological architecture; single-subject 18-session free-living pilot; modest Random Forest performance; acoustic, HRV, wrist-temperature and movement features; feature leakage/location identity; thermal and acoustic domains validated while visual and localised indoor air quality remained unresolved.