Wearable sensors and the archer's shot: Liu 2025

A Beijing company that makes wearable sensors put its own wristband on the bow-hand wrist of six adolescent archers (4 girls and 2 boys, mean age 15.5) under official competition conditions, and tried to use the signals for two separate jobs: recognising the phases of the shot automatically, and estimating how much pressure the archer felt under.

For the phases, an LSTM neural network reading only the three components of acceleration picked out the windows containing a complete shot with an accuracy of 84.53%. Adding total acceleration and a smoothed measure of how it changes, which the authors call SmoothDiff, raised accuracy to 95.99% (F1 95.95%), and a bidirectional version of the network reached 96.80%. When the count is instead how many individual windows are classified correctly under conditions closer to real use, the figure falls to 75%.

For stress, twelve different classifiers tried to separate the ends the archer had rated as high stress from those rated low, working from heart rate variability. Accuracies ranged from 68 to 82%. The model the authors chose, a multi-layer perceptron (MLP), stopped at 80% (F1 75%) and at 77.11% on held-out data; logistic regression and linear discriminant analysis reached 82%, so they did slightly better than the model the preprint highlights.

The wristband (Ergosensing, developed by the authors themselves) recorded a three-axis accelerometer and a green-light optical sensor (photoplethysmography), from which heartbeats are derived, both at 20 Hz. Everything took place on a single day, at a competition shot in ends of three arrows; the number of ends differed between archers, because those who went through to the next stage shot more. The experimenter marked the start and end of each end, each draw and each release on a tablet; then, looking at the acceleration trace, every shot was cut by hand into draw, aim and release, with the end of the release set at the first peak in the signal. After each end the archer rated the tension they had felt from 1 to 5: 1 to 3 became low stress, 4 and 5 high stress.

For phase recognition the signal was split into 4-second windows (80 samples) moving forward one second at a time; a window counted as a shot if more than half its points fell inside an annotated shot, and the windows without a shot, far more numerous, were thinned out while those with one were duplicated. For stress the windows lasted 30 seconds, and from each one the authors computed heart rate, SDNN, RMSSD, pNN20, pNN50, high-frequency and total power, SD1, SD2 and sample entropy. In both cases the data were split 70% for training and 30% for testing.

Six archers, one day, and no number of arrows or ends reported: this is a prototype, and the preprint has not been through peer review. The most serious weakness lies in how the data were split, with no statement of whether windows from the same archer or the same end ended up in both the training and the test sets. With windows that overlap by three quarters and only six people, the model was probably tested on signals almost identical to ones it had already seen, and in that case the accuracies come out inflated. The figure for held-out data is not explained either.

The numbers do not quite line up: the summary credits the LSTM with 96.8%, while the table gives that figure to the bidirectional version, with the LSTM at 95.99%. The 'stress' is a rating given at the end of each end, so every three arrows, while the heart-rate windows last 30 seconds and therefore include time between shots. One point of my own, not the authors': an optical sensor at 20 Hz takes a sample every 50 milliseconds, a coarse resolution for measuring differences of 20 milliseconds between successive beats such as those behind pNN20. Finally, nothing measures the score, so there is no way to tell whether the ends rated more stressful were shot better or worse.

For anyone shooting today the study changes nothing: it is not about technique or score, and the wristband is not a validated training tool. What it shows is that an accelerometer on the bow-hand wrist can tell the windows that contain a shot apart from the rest of the time spent on the line. Splitting the shot into draw, aim and release is described in the method, but how accurately it works is not measured. That is a first step towards counting shots or timing the aim automatically, things a coach currently notes by hand or works out from video.

The stress part calls for more caution. A classifier that guesses eight times out of ten whether an archer will report feeling more or less tense, on six teenagers and with a data split that probably flatters the model, is not yet an indicator to rely on in competition. For anyone who already wears a watch that estimates heart rate variability, this work is not enough to turn that number into a measure of mental tension during an end.