In this guide
How Wrist Motion Sensors Track Parkinson's Symptoms: Tremor Bands, Rest Gates, and Dose-Time Plots
A wrist sensor cannot see Parkinson's disease. It can only feel movement: acceleration and rotation, sampled many times a second. Everything useful that follows — a tremor score, a slowness proxy, a day of fluctuation — is signal processing applied to that movement, plus rules about when the data is trustworthy. This guide explains that chain, why each step exists, and where honest student measurement stops and clinical measurement begins.
Why a wrist, and why continuous tracking
Parkinson's symptoms change through the day. Tremor comes and goes, movement slows, and involuntary movements appear, often in patterns a single clinic visit misses. A wrist device worn for days records movement continuously, so the result is a timeline rather than a snapshot. That timeline is the whole point: it can show fluctuation around dose times, quiet periods, and restless nights that a person would struggle to recall accurately. Empatica's WellPD, launched in late October 2026, is the current commercial reference for this idea: an FDA-cleared and CE-marked wrist device with a companion app that summarises tremor, dyskinesia, bradykinesia, sleep and activity into daily, weekly and monthly views. Its algorithms are clinically validated products of years of work. A student build borrows the shape of the idea, not the validation, and this guide keeps that distinction explicit throughout.
The sensor: what an IMU actually measures
The core sensor is an inertial measurement unit, usually a small module such as the MPU6050 that combines a 3-axis accelerometer with a 3-axis gyroscope. The accelerometer measures acceleration along three axes, including gravity, so it reports both tilt and linear motion. The gyroscope measures rotation rate around three axes, so it reports how fast the wrist is turning. Together they give six streams of numbers. A typical student build samples them at around 100 Hz, meaning 100 readings per second per axis, which is fast enough to capture the few-hertz rhythms of tremor while staying light enough for a small microcontroller such as an ESP32 to process and serve over Wi-Fi.
Raw IMU data is messy. Gravity is always present in the accelerometer, walking shakes every axis, and a sensor strapped loosely adds its own wobble. No single reading means anything on its own; the work is all in how the streams are filtered, windowed, and gated before any score is computed.
From wiggles to a tremor band: the FFT idea
Tremor is rhythmic, and rhythm is best seen in frequency, not in time. The standard tool is the Fast Fourier Transform (FFT), which takes a short window of motion data — a few seconds of acceleration or rotation — and reports how much energy sits at each frequency. A resting tremor typically concentrates its energy in a narrow band around 4 to 6 Hz, while walking and voluntary movement spread energy across much wider ranges. So a tremor index can be defined simply: compute the FFT of a rest window, add up the energy in the 4 to 6 Hz band, and compare it with the energy outside that band. A high share in the band means the motion looks tremor-like; a low share means it does not. That is a demonstration index built from a stated rule, and its honesty comes from stating the band, the window, and the formula openly so anyone can reproduce or challenge it.
The rest gate: the step students most often skip
The FFT band trick only works on still hands. The moment a person walks, reaches, or gestures, the motion swamps the tremor signal and the band calculation becomes meaningless. Clinical and student systems therefore need a rest gate: a rule that admits data into the tremor score only when the hand has been verifiably still. A simple, testable version watches gyroscope magnitude and opens the gate only after it has stayed below a fixed threshold for a full 2 seconds; walking raises the magnitude above the threshold within about a second and closes the gate, freezing the score. The gate is the most important design decision in the whole build, because without it the device reports walking as tremor. A session with no still periods honestly produces no tremor score at all, and that gap is a property of the design, not a fault to hide.
Slowness, involuntary movement, sleep, and activity: four proxies, four caveats
Bradykinesia, or slowness of movement, cannot be read directly from one sensor trace, so student builds use proxies. The most honest is a timed finger-tap test: count taps in 10 seconds through a button, plus a movement-speed score from gyroscope magnitude during a slow guided reach. Both numbers are sensitive to fatigue and practice, so a single count is unreliable and repeats need a rest between them. Involuntary movement is harder still; a student build can only flag raised movement variability during rest windows, which is a proxy indicator and never a clinical dyskinesia measure. Sleep and activity come from actigraphy, the old idea that movement counts over time approximate rest and activity: overnight movement counts give a rest log, and daytime counts give step and activity totals. Neither stages sleep, scores sleep quality, or detects falls, and claiming otherwise turns a proxy log into a safety claim.
Dose-time plots: correlation display, never an efficacy claim
The most revealing view in a symptom monitor is the day plotted against manually logged dose times: a tremor curve with markers showing when doses were taken. It lets a student see fluctuation around those times, which is exactly what a clinician wants from a timeline. But the plot shows timing only. It cannot judge whether any medicine worked, because dose timing, activity, fatigue, and time of day all move together, and a student dataset has no controls. The report must say so wherever the plot appears. The same discipline applies to every number in the build: each index carries its formula, its gate, and its limit, so an examiner can test the claim rather than trust it.
How a student build differs from a clinical device
A clinical device such as WellPD pairs validated algorithms with large patient datasets, regulatory clearance, and a companion app refined over years; its reported figures, such as a 14-day battery life and an 11-gram device weight, describe a finished medical product. A student build uses an ESP32 and an MPU6050, fixed thresholds chosen by the builder, one wrist, and one wearer per calibration, with no clearance and no validation. Its value is different: it demonstrates the full measurement chain — sensing, gating, frequency analysis, proxy design, and honest reporting — at a scale a student can build, test, and explain. If you want to see that chain as a working project, the companion build is the WellPD-inspired wrist motor symptom monitor, which implements the rest gate, the 4 to 6 Hz tremor band, the tap test, and the dose-time plot as a testable prototype with every limit stated. The same project page also documents the calibration checks — the still-hand test, the walking-exclusion test, and the tap-test repeat — that turn the guide's ideas into a demonstration an examiner can watch.
Practical tips for building one
Start with the gate, not the score: get the still-hand detection working and visible on a dashboard before computing any FFT. Test the gate by walking, because a gate that never closes is worse than no gate at all. Keep the vocabulary of tasks small — one rest test, one tap test, one reach task — so every number has a procedure behind it. Log everything to CSV from day one, including excluded segments, so the report can show which windows were scored and which were rejected. Charge and mount the battery safely with a protected holder and supervised charging, and keep the strap comfortable enough for a long session, since an unworn device records nothing.
Frequently asked questions
Can a wrist sensor diagnose Parkinson's disease?
No. A wrist sensor measures movement only, and tremor-like motion has many causes. Diagnosis is a clinical judgement that combines history, examination, and tests, and no student build, however careful, performs it. Student projects demonstrate measurement, not diagnosis, and their reports must say so.
Why 4 to 6 Hz for the tremor band?
Resting tremor in Parkinson's disease typically concentrates around 4 to 6 Hz, which makes that band a useful, literature-grounded starting point for a demonstration index. It is a design choice, not a fingerprint: other movements can leak energy into the same band, which is exactly why the rest gate must run first.
What breaks a tremor score most often?
Movement, in that order: walking, gesturing, and a loose strap. Then comes a gate threshold set too high, which admits walking as rest, or too low, which admits nothing at all. The fix is procedural: calibrate the threshold on the actual wearer, test it by walking, and log excluded segments so failures are visible.
Is a tap test a real bradykinesia measurement?
It is a proxy, not a clinical measure. Taps per 10 seconds plus a reach speed score track slowness-like behaviour under controlled conditions, but fatigue, practice, and motivation shift the numbers. Treat repeats as a small dataset with a stated procedure, never as a single headline result.
Can the same device track sleep properly?
No. Movement counts approximate rest and activity, but they cannot stage sleep or score its quality. A student build should call its overnight record a movement-count log and leave sleep staging to clinical systems.
Limitations
This guide describes demonstration measurement, not medical practice. Student-built symptom monitors are not FDA-cleared, not CE-marked, not diagnostic tools, and not clinically validated, and their indices must never guide treatment decisions or be quoted as a person's medical state. Thresholds transfer poorly across people and wrists without recalibration, rest-gated scores vanish without still periods, and dose-time plots show timing, not efficacy. Commercial figures quoted here are reported references for context only, and the student build's own numbers come only from its stated rules and its own demonstration runs.
See the working student version here: WellPD-inspired wrist motor symptom monitor.