After reading about AI-supported VR systems, I wanted to understand more clearly what machine learning is actually doing when a system claims to recognise anxiety. I followed one of the references into Tayarani-N. and Shahid's Detecting Anxiety via Machine Learning Algorithms: A Literature Review.

The paper reviews a large range of studies that use machine-learning algorithms to identify patterns associated with anxiety. What makes it useful for my project is that the data is not limited to questionnaires or self-report. The review covers information including voice, facial expression, movement, brain activity and physiological signals collected from the body.

What Can the Body Tell the System?

The section on biosignals was the most relevant to my project. The authors describe physiological signals as readings collected from the central and autonomic nervous systems that can contain information about a person's internal state. Examples include cardiac activity, electrodermal activity, muscle activity, temperature and respiration.

Several of the studies reviewed used ECG, heart rate, HRV, breathing or galvanic skin response alongside machine-learning algorithms to identify patterns associated with anxiety. Other studies combined multiple physiological measures rather than relying on one signal alone.

This is interesting because it suggests that a future system does not necessarily need the traveller to continually explain how they feel. Information from the body could potentially provide another layer of information about a changing state.

However, detecting a physiological change is not the same as understanding what that change means.

Wearable Devices

The review also includes a section specifically on wearable devices. Smartwatches, wristbands and body-worn sensors are increasingly able to collect physiological and behavioural information outside laboratory or hospital environments.

Some of the studies use wearable sensors to collect heart rate, ECG, EDA, temperature, movement and breathing information before processing it through machine-learning models. This makes the research particularly relevant to my airport project because a wearable could theoretically collect information while somebody is moving through a real environment rather than only while sitting in a controlled research setting.

This begins to make the idea of live regulation feel more technically plausible. A traveller would not necessarily need to stop and complete an assessment for the system to receive information about physiological change.

The Problem With 'Anxiety Detection'

The most important part of the paper for me is actually its limitations. Tayarani-N. and Shahid explain that anxiety can be measured through self-report, physiological and behavioural methods, but there is low agreement between these different measures. They also warn against assuming that physiological markers can cleanly distinguish anxiety from other psychological states.

This matters enormously for my project. Airports already create many reasons for physiological arousal. Somebody may be walking quickly, carrying luggage, rushing for a gate, excited about travelling, frustrated by a queue or responding to sensory overload. A higher heart rate could appear in all of these situations.

The review also notes that behavioural information can be affected by context and individual differences, while wearable data depends on factors such as the person consistently wearing the device correctly. This means that an automated system cannot be treated as an objective reader of somebody's emotional state.

For my project, I therefore need to be careful with the language I use. I do not think I should describe the technology as a system that can simply "detect sensory overload" or "know when somebody is anxious."

A more accurate idea is that machine learning may be able to identify patterns within physiological and behavioural data that could contribute to a wider understanding of changing arousal.

Multimodal Rather Than One Signal

Another idea that appears repeatedly in the review is the potential value of combining different kinds of information. The authors suggest that biological, clinical and real-time physiological data could potentially create a more holistic picture than relying on one signal by itself.

This connects to the idea of multimodal sensing that appeared in my previous VR research. Rather than designing a system where one heart-rate threshold automatically causes a response, a future concept could potentially consider several pieces of information alongside the person's own settings or baseline.

I do not need to build a machine-learning model for this master's project, but understanding the principle is useful because it changes how I think about the interaction. The intelligence of the system would sit in how it interprets incomplete information rather than simply reacting to one number.

What This Changes in My Project

Before reading this paper, I was at risk of oversimplifying the relationship between the body and anxiety. The research makes it clear that physiological information can be useful, but it is not a direct translation of emotion.

This gives me a much stronger position for the project. The purpose of sensing should not be to diagnose the traveller. It could instead provide an additional source of information that helps a responsive system recognise when the person's physiological state is changing.

The next thing I need to understand is which biosignals are most useful and realistic within an airport context, particularly HRV, heart rate and electrodermal activity, and what their limitations are when somebody is physically moving through the environment.

Key Terminology

Machine Learning (ML) — algorithms that learn patterns from data and use those patterns to classify or predict outcomes.

Anxiety detection — the paper's term for using different forms of data and computational models to identify patterns associated with anxiety. For my project, I need to use this term cautiously because detection does not mean a system perfectly understands the cause of physiological arousal.

Biosignals — physiological signals collected from the body, including cardiac activity, electrodermal activity, respiration, temperature and brain activity.

ECG (Electrocardiogram) — measurement of the electrical activity of the heart.

Heart Rate Variability (HRV) — variation in the time between heartbeats.

Galvanic Skin Response (GSR) / Electrodermal Activity (EDA) — measures associated with changes in the electrical properties of the skin.

Multimodal data — combining more than one source of information rather than relying on a single signal.

Affective computing — computational systems concerned with recognising, interpreting or responding to emotional information.