How do we make multi-sensor research setup easier, without losing data quality?

Eye tracking and biosensors are valuable in UX research. They show where people look and how much effort or stress a task causes, which people can't reliably report about themselves. More and more UX, HCI and automotive studies use them.

The tools around them don't get the same attention to UX.

Each device comes with its own software, its own terms and its own steps. A study that uses two sensors means combining two separate processes by hand. In our lab, that meant calibrating the eye tracker in Tobii Core, connecting the EDA sensor and recording a baseline in OpenSignals, then exporting the recording and running a Python script in Jupyter to get the baseline value. None of these tools tells you when everything is ready.

This keeps the methods out of reach for researchers who aren't specialists, and mistakes during setup can make the data unusable. The process needed to be combined into one, and designed for the person doing it.

Sensa is a new web platform for running multi-sensor UX studies in one place: Tobii eye tracking, plus EEG, EDA and ECG through the Biosignalsplux kit. It's being built from scratch by a team of four in the HSRW Usability Lab, and new modules are added as they're ready.

I owned onboarding and calibration, everything that happens before data is recorded. I did the research, the information architecture and the interaction design, and developed the parts closest to the data: the validation layer, the live eye position view and the EDA signal quality indicators.

This case study covers the first full version and the study that tested it. Sensa is still changing, and the study's findings feed into the next version.

I started with what I'd seen in the lab and then checked it against the literature. Prior work names onboarding and calibration as the stage of multi-sensor research most prone to usability problems. But I couldn't find a study that evaluated the setup process itself. Research on these systems looks at signal accuracy and synchronisation, and rarely at the person doing the setup.

So the research had two phases. Discovery shaped the design, and a controlled study tested it against four research questions.

Discovery pointed to five areas where setup goes wrong. Each became a requirement for the flow.

Choosing sensors and metrics takes physiology knowledge that most of Sensa's target users (new researchers, small teams) don't have. So the flow asks about the study instead: the goal, the interface, the environment and the hardware. Each goal lists the signals it affects. Sensa then recommends sensors and explains why.

One participant worked this out mid-task: “I didn't choose these sensors! - oh it's recommending them to me, I get it.”

Many calibration failures start earlier: the participant sits too close, the lighting is wrong, an electrode isn't attached. The current tools never check for any of this.

So before any sensor work, the researcher confirms four things: seating at about 60 cm, lighting and noise, sensor access and electrodes attached. Each item explains its standard in a tooltip. The calibration button stays disabled until all four are checked.

If the participant sits too close, too far or off to the side, the calibration will be poor no matter what else happens. So positioning is a separate step with a live view of where the tracker sees their eyes, and their distance in millimetres.

The calibration itself runs in Tobii's own routine, since the 4C doesn't support the SDK needed to run it inside Sensa. Sensa hands off to Tobii and takes over again for validation.

Finishing a calibration doesn't mean the gaze data is accurate. Quality depends on the person's eyes, the lighting, head position and the calibration itself. The eye-tracking literature recommends checking it with a separate validation step before collecting data, and measured accuracy often differs from what manufacturers claim.

Tobii's consumer tracker reports no quality numbers at all.

Its built-in check is a gaze trace you judge by eye. So I designed and developed a validation layer for Sensa. The participant looks at five targets, one at a time, for about two seconds each, while a live trace shows they're being tracked. Each target turns green or red. The result screen shows where the gaze landed against each target, whether the calibration passed, and the numbers behind it. If it fails, the researcher is asked to recalibrate before going on.

I based the metrics on how eye-tracking data quality is reported in the literature, so a researcher can put them straight into a paper. Errors are converted from screen distance to degrees of visual angle, which makes results comparable across screens and seating distances. The conversion uses the participant's real distance from the screen, read live from the tracker.

The 3° default threshold is a practical choice for a consumer tracker used by non-experts. Research-grade systems aim for under 1°, so the threshold is shown with every result and can be made stricter.

The layer also made the study possible. It measured calibration quality the same way in both conditions, which is how I could check that the new flow didn't make the data worse. It hasn't been benchmarked against a research-grade tracker yet, so its values are relative to this setup.

After the eye tracker, the EDA sensor should feel familiar, so it uses the same stepper, the same rules for moving on and the same button placement.

Each step covers one thing: placing the electrodes, connecting the cable to the right port, checking the signal and recording a baseline. The baseline value is calculated inside Sensa, so there's no export and no Python.

To answer the research questions, I ran a within-subjects study. 12 participants did the full setup twice, once with the current tools and once in Sensa. Half started with each, so learning effects were spread evenly.

Participants followed a printed task sheet without help from me. I observed, logged every breakdown into five categories and screen-recorded each session to check my notes later.

Workload, usability and breakdowns all improved, each with a large effect. Calibration time also went down, but with 12 participants that result isn't statistically confirmed.

Most of the drop in breakdowns came from less uncertainty.

Hesitation and asking for clarification make up 141 of the 142 fewer breakdowns. Procedural errors were rare with both setups. Participants mostly stopped because they couldn't tell whether a step had worked, or what to do next.

I coded the written answers and observation notes into 30 codes and five themes. With the current tools, the hard part was coordinating them. In Sensa, people found it easier because they were led step by step and could see what the system was doing. One participant rated their calibration confidence 3 out of 7 with the current tools and 7 in Sensa:

Showing more also made some people more worried.

Average confidence barely changed. When Sensa showed a reading that looked off, some participants became less sure, and re-recorded baselines went up from 10 to 16. For a research tool I think that's right. Confidence should match the actual state of the setup.

Both setups used the same Tobii calibration engine, so the question was whether the surrounding workflow made results worse. For accuracy, valid data and first-try success, there was no sign of that.

Precision was lower in Sensa. The difference came from two sessions that also had the lowest valid data, which suggests unstable tracking in those sessions. The same link showed up with the current tools. I can't fully rule out the workflow, and the thesis says so.

The study also found problems in my design. Since Sensa is a new product still in development, these feed straight into the next iteration.

Also on the list: benchmarking the validation layer against a research-grade tracker, recommendations that change with the chosen goal (the prototype shows a fixed set), and testing with EEG and ECG to see if the improvements hold as setups get bigger.

I expected to design against mistakes. The research showed that people mostly got stuck because they didn't know whether a step had worked. Showing the system's state, and putting a number on calibration quality, helped more than anything else I designed.

Testing my own design against a real baseline showed what worked, and it also caught a button I forgot to lock. The study has limits: 12 participants, one lab, a consumer eye tracker, and me as designer, developer and analyst. I'd rather state those clearly than overstate the results.