Painless

Opportunity

Nearly one in four US adults lives with chronic pain [1], yet pain is usually measured from memory, which misses when it happens and what triggered it. In 2023, 24.3% of US adults lived with chronic pain, yet most of what doctors learn about it comes from memory: a rating at an appointment, or a diary filled in at the end of the day. The moments that matter, when the pain actually hits and what the person was doing, are usually lost.

Our team brought together biomedical engineering, AI research and product design to create an objective measurement to pain, and we all kept coming back to one question: Why is pain still measured by asking "on a scale of 0 to 10, how bad is it?" Painless uses a wrist wearable to passively monitor voice (pain sounds, pain words, vocal strain) and movement (jerks, abrupt stops, stiffness, restlessness). It learns each person's usual range during their first week and labels moments Normal, Moderate or High relative to it. When both signals agree a moment is unusual, the wearable asks "What were you doing?" and logs the spoken answer. A pattern engine then links moments to activities, times and body areas. We describe the design, a two-stage evaluation plan and key limitations. Painless is a wellness tool, not a medical device.

Why now? In hospitals, nurses ask that question every few hours. Patients who can't answer, such as sedated patients, people with dementia or newborns, are scored by a nurse watching their face, voice and movement. Outside the hospital it's even harder. We noticed that the clinical scales nurses already trust, like CPOT and PAINAD, score two things a wearable can sense continuously: vocalization and movement. The Vexo bracelet captures both. So we asked: what if the wearable noticed the moment, and Vexo takes note of the users activity in the moment. Vexo can ask for user input if it needed it.

What we learned

Voice carries real pain information. The AVPENet study found that while facial cues were the stronger signal on their own, adding voice made pain estimation noticeably more accurate- Everyone's "normal" is different. Research from MIT and the PANDIA study showed that models adapted to each person outperform one-size-fits-all models. That pushed us to compare every user only with themselves. Mild pain is hard to detect. Accuracy drops as you add more pain categories, so we settled on three levels (Normal, Moderate, High) and let the user's own words provide the detail. There is a line between wellness and medicine. A product that diagnoses or guides treatment is an FDA-regulated medical device. We designed Painless as a personal wellness tool, with a clear non-clinical disclosure on every screen. Simple beats statistical. Our first design showed bell curves. We realized most users wouldn't understand them, so we replaced them with a three-zone bar and a "Now" marker that reads like a fuel gauge.

How we built it

1. Sensing. The Vexo bracelet streams audio and movement. Movement intensity comes from the acceleration magnitude, and sudden jerks from how quickly it changes:

For voice, we detect speech-to-text, and measure vocal strain through pitch and loudness

2. A personal baseline. The first week of wear sets each user's usual range.

3. Combining voice and movement.

This means one signal alone can never trigger High, which protects against false alarms from a sigh (voice only) or a workout (movement only).

4. Check-in. Vexo logs the activity that spiked movement and audio

5. Pattern engine. Logged moments are grouped by activity, time of day and body area to surface insights like "8 of your 10 high moments came after 2+ hours of sitting."

6. Pilot experiment. To test whether pitch and volume change with pain, two of us each said the phrase "Vexo Hacks is cool" under three conditions:

at rest (Normal)
with a hand in an ice bucket (Moderate)
in the ice bucket while our arm was pinched (High)

We extracted pitch and volume from the six recordings and trained a residual network (ResNet) to classify the three levels, following the audio approach used in AVPENet.

Challenges we faced

Tiny data. With two participants and six recordings, our pilot shows the pipeline works end to end, but it can't show that it generalizes. The conditions also ran in a fixed order, so repetition effects can't be ruled out. We designed larger lab and field studies as the next step.
No ground truth for pain. Pain is subjective, so we had to induce it safely (cold water is a standard lab method) and treat the user's own report as the reference. -We did not want to simulate data in order for it to be tailored to human experiences
Telling pain from everyday life. Coughing, laughing, exercise and background voices can all look like pain. Requiring voice and movement to agree was our main defense.
Clinical vs. wellness. We started out designing a hospital tool, then realized it would need FDA clearance. Reframing Painless as an everyday wellness product changed our language, our levels and our disclosures.
Privacy. Always-on listening is sensitive, and some states require everyone's consent to record conversations. We designed the system to turn audio into features on the device and never store raw recordings.
Choosing the right number of levels. We debated adding a "Very High" level, but it would rest on roughly 1 in 1,000 moments, too rare to measure reliably. Three levels, plus the user's own rating, gave us both reliability and detail.

Accomplishments that we're proud of
We tested on real human voices, not simulated data. In one day we designed and ran a safe induced-pain experiment (ice bucket, then ice bucket plus a pinch) so our model learned from genuine human reactions. -** We designed a feature that has not yet been adopted by any wearable device**
We trained a residual network on our own recordings. It goes from raw audio to pitch and volume features to a three-level classifier, the same family of models used in published pain research.
We grounded every design choice in clinical research. Our signals mirror what nurses already score on validated scales like CPOT and PAINAD, and our personal-baseline approach follows studies showing that per-person models outperform one-size-fits-all ones.
We built a system that adapts to each person. Painless compares you only with yourself, so a naturally loud talker or an active person isn't flagged for being who they are.
We worked across disciplines. Two biomedical engineers, an AI engineer and a product designer turned an idea into research, a working pilot, an app design and a written paper in a single day.

What's next for Painless

A bigger, better-controlled study. First, a lab study with about 20 participants, randomized condition order and more phrases. Then a four-week field study with about 30 people who live with recurring pain, to measure how often Painless catches real moments and how useful its patterns are.
Add heart signals. Use heart rate and heart rate variability as a supporting channel, since research shows consumer wearables can track pain this way.
A clinical version. With validation, adapt Painless for patients who can't self-report, such as people after surgery or living with dementia, through the FDA pathway already used by AI pain monitors.
Fair for every voice. Test across ages, sexes, accents and languages so Painless works equally well for everyone.
From insight to action. Turn patterns into gentle, well-timed nudges, like a stand-up reminder before your usual long-sitting threshold, and measure whether they reduce high-pain moments.