WATCH / RESEARCH NOTES · ROSALIND PICARD / AFFECTIVE COMPUTING

I watched three pieces of work connected to Rosalind Picard and the MIT Media Lab’s Affective Computing group: Emotion Technology: From Autism to Customer Experience and Decision Making, An AI smartwatch that detects seizures, and Research Group Overview: Affective Computing. Watching them together was much more useful than treating them as separate references because they show the progression from physiological sensing, to real-world wearable systems, to the wider field of affective computing.

I also found Picard’s work genuinely fascinating. What I like about it is that she moves very comfortably between engineering, physiology, AI, healthcare and human experience. The technology is clever, but the interesting part is always what that technology can actually do for a person.

Affective computing and the autism connection

Picard uses the term affective computing for technology that attempts to sense, interpret or respond to information connected to emotion, arousal and human state. This immediately felt relevant to my project because some of her earlier work looks specifically at autism and at situations where a person’s internal physiological state may not be obvious from outward behaviour.

That matters for my project because airport support often depends on the traveller recognising that they need help and then communicating that need. A wearable system could potentially notice a meaningful physiological change earlier and offer support privately, without assuming that the person must visibly appear distressed first.

However, this does not mean the technology can simply “read emotion”. The stronger interpretation for my project is that it can notice changes in physiological activity that may be meaningful when combined with context and the user’s own interpretation.

Understanding the autonomic nervous system

One of the parts I found confusing in the first talk was Picard’s discussion of the autonomic nervous system.

The autonomic nervous system controls many bodily processes that happen largely without conscious effort, including heart rate, sweating, digestion and parts of breathing.

The sympathetic nervous system is associated more with activation and preparing the body for action, while the parasympathetic nervous system is associated more with slowing, recovery and regulation. They are not simply a stressed/calm switch, but this distinction helps me understand why signals such as EDA, heart rate and HRV change.

Useful way for me to think about it: the wearable would not be reading an emotion directly. It would be reading changes in autonomic activity that may indicate increased physiological demand or arousal.

Skin conductance and why one signal is not enough

Across Picard’s work, skin conductance / electrodermal activity (EDA) appears repeatedly. EDA changes with sweat-gland activity and is closely linked to sympathetic arousal.

What I found especially useful is that Picard repeatedly shows how easily physiological signals can be misinterpreted. EDA can change with emotional stress, but also with cognitive load, excitement, neurological activity and other forms of activation.

This is directly relevant to my project. If an autistic traveller’s EDA or heart rate changes, the system cannot automatically know whether the cause is sensory overload, anxiety, walking quickly, carrying luggage, excitement or something else.

That reinforces the direction I have already been moving towards: personal baseline + multiple signals + context + user interpretation.

Picard’s earlier 24-hour recordings also produced surprisingly large EDA peaks during sleep. That was useful because it showed me that a strong physiological response does not necessarily map neatly onto what I would intuitively label as “stress”. The body can behave very differently across sleep, movement, attention, emotion and neurological activity.

Real-world context: the driving studies

Picard’s wider work also used real-world driving to explore physiological sensing. Drivers were monitored using signals including ECG / heart activity, skin conductance, respiration and muscle activity while researchers compared calmer and more demanding driving conditions.

I think this is a particularly useful comparison for airports because both are complex, changing environments. If a traveller’s heart rate rises, the system cannot know whether that came from anxiety, walking quickly, carrying luggage, cognitive effort, excitement or sensory demand. This reinforces why the airport context needs multiple signals and contextual information, rather than a single threshold labelled “stress”.

The seizure work: from unexpected data to useful intervention

The most memorable example was Picard’s work on seizure detection.

A student from MIT borrowed two skin-conductance wristbands for his non-speaking autistic brother. At one point, one wrist showed a very large EDA peak while the other remained relatively flat. Initially, this looked like a faulty sensor.

The family diary later showed that the peak occurred shortly before a generalised tonic-clonic seizure. This unexpected observation led the research in a completely different direction.

Further work combined EDA with movement / accelerometry and used machine learning to recognise seizure-related patterns. This eventually contributed to wearable seizure-detection technology associated with Empatica.

For me, the important part is the architecture:

collect signal → identify pattern → validate it → combine signals → train algorithm → trigger a useful response

This makes AI much easier to understand in relation to my project. AI is not “reading the body” in some magical way. It is a pattern-recognition layer between sensing and response.

The seizure work also introduced me to SUDEP — Sudden Unexpected Death in Epilepsy. Picard explains that after some generalised tonic-clonic seizures, breathing can stop before the heart stops. The value of the wearable is therefore not that it treats the seizure itself, but that it can detect an event and alert another person quickly enough for action to happen.

That is a useful design lesson for my project: the sensing system does not necessarily need to “solve” the physiological state. Its value may be in recognising a meaningful change and triggering the right next step.

The amygdala and the one-sided signal

The word I could not hear properly in the seizure talk was amygdala.

Picard explains that we have an amygdala on each side of the brain, and research involving direct stimulation has shown that stimulating one side can produce a stronger skin-conductance response on the corresponding side of the body.

I do not need to investigate the amygdala in depth for my project, but this example was useful because it shows how complex physiological data can be. A strange signal does not necessarily mean a sensor is broken, and a skin-level measurement can sometimes reflect processes happening elsewhere in the nervous system.

Biofeedback and influencing physiological state

Another idea I initially found confusing was the suggestion that people can learn to influence some of these signals.

The autonomic nervous system is largely automatic, but some of the processes it affects can be influenced indirectly through things such as breathing, relaxation and biofeedback.

That gives me another possible role for my project. The system would not only notice a meaningful change; it could offer a user-controlled regulation response such as a breathing prompt, haptic pattern, audio intervention or another support option.

The key is that the traveller remains involved in the loop rather than being passively monitored.

What the wider Affective Computing group adds

The MIT group overview helped me see that affective computing is much broader than one smartwatch or one medical application. The field sits across wearable sensing, machine learning, mental health, communication, privacy, wellbeing and human-computer interaction.

That actually helps me understand where my own project sits. My project is not simply “a wearable”. It sits somewhere between:

biosensing · affective computing · human factors · interaction design · accessibility

This feels much closer to the kind of work I am interested in doing.

Privacy and control are part of the system

The wider Affective Computing work also makes the privacy question impossible to ignore. Physiological data can reveal information that a person has not deliberately chosen to communicate, so the design cannot treat consent as an afterthought.

For an airport accessibility system, I would need to consider who can access the data, whether airport staff ever see it, whether it stays on the user’s own device, how long it is stored, whether sensing can be paused, and how the user can reject or correct an interpretation.

That makes privacy, transparency and control part of the interaction design itself rather than a policy layer added at the end.

Why I find Picard’s work so interesting

What I find particularly clever about Picard’s work is that the technology is rarely the end point.

The research might begin with questions such as can this signal be measured?, can an algorithm recognise a pattern? or can a wearable detect a change?, but the more important question becomes:

what can that information do for the person?

That is exactly the shift I need to make in my own Studio Practice. I do not need to become the engineer building the sensor. My design role can sit in the translation between physiological data and human experience: how the system communicates, what response it offers, how uncertainty is shown and how the traveller stays in control.

Personal reflection: Picard’s work makes this direction feel much more real to me. It shows that meaningful work can sit between technology, physiology and human experience rather than belonging neatly to one discipline.

What I am taking forward

The three videos together have clarified several things for me. Physiological signals can indicate that something has changed, but they do not automatically explain why. Multiple signals and contextual information are more useful than a single metric. AI can recognise patterns without creating certainty. Biofeedback creates the possibility of a closed loop in which the user responds rather than simply being monitored. Privacy, consent and user control have to be part of the system architecture.

Most importantly, I am now thinking less about “Can technology detect stress?” and more about:

How could a system notice meaningful physiological change, communicate that uncertainty clearly, and give an autistic traveller more control over what happens next?

That feels much closer to the actual design problem.

Sources