Aiming patterns and motion sensors: Quan 2016 archery

Four girls from the archery team of a Korean middle school, between 14 and 16 years old and with an average score of 56.8 out of 60 the previous season, each shot 36 arrows at 30 metres wearing nine inertial sensors. For every arrow the authors took the two seconds before the release and drew twenty-seven figures from them, one per axis per sensor, calling them the "aiming pattern" of that shot. They then used stepwise regression to find which of those figures predicted the score of the arrow.

The outcome is that a model can be found for all four, but it differs for each one. Ten variables were selected for the first archer, three for the second, eleven for the third and five for the fourth; the share of score variability explained is 64.0, 27.9, 72.0 and 45.6 per cent respectively, with models significant at 0.12, 1.61, 0.55 and 0.40 per cent. For the first archer the upper arm and forearm of the bow arm weigh most, for the fourth the drawing arm, for the third the pelvis. A final model, built on the averages of the four athletes, keeps only the bow forearm and the left shank, with the leg weighing 3.87 times the arm.

The nine sensors of the Noraxon myoMOTION system, about four centimetres by five and weighing 34 grams, sat in elasticated band pockets on the head, the upper back, the left forearm and upper arm, the right forearm and upper arm, the pelvis and both shanks, and sampled at 100 hertz. Each archer shot six ends of six arrows, with the usual pause to walk up and collect them, on her own school range.

The delicate part is cutting out of the continuous trace of an end the two seconds of aiming before each arrow. The authors do it with a dynamic time warping algorithm, which searches the long recording for the stretches most similar to a reference stretch chosen by hand and marks their end point, where the acceleration changes abruptly. Over each two second window, that is 200 samples, the mean is computed for each of the twenty-seven axes. Automatic extraction did not succeed on every arrow: 36 windows for the first archer, 35, 29 and 33 for the others, 133 in all instead of 144, because of recording errors or because the algorithm could not find the stretch. The analysis is a stepwise multiple regression run in SAS, with the score of the single arrow as the variable to explain, and for each coefficient the authors also report the standardised value, which allows the weight of different body parts to be compared.

Four participants, all girls, all from the same school team and of similar level: any regularity found describes those four and not archers in general. The individual models carry ten and eleven variables over some thirty arrows, a ratio that almost inevitably chases noise, and stepwise selection inflates the significance of the coefficients that survive, because it tries many and keeps the best. The authors report no test of the models on data other than those used to build them.

The pooled model is the weakest point: it is computed on the averages of the four athletes, that is on four points, with two variables. A near-perfect fit, with explained variance close to one hundred per cent and an F of 56,898, means nothing in those conditions, and the ratio of 3.87 between leg and arm that follows from it should be read as a figure with no degrees of freedom behind it. The text also contradicts itself on the significance of that model, 0.30 per cent in the results and 0.75 per cent in the discussion, speaks of 29 variables where the sensors give 27, and in the table for the first archer a positive coefficient carries a negative standardised value.

A doubt remains about what was measured. The mean acceleration over two nearly motionless seconds mostly describes the tilt of the sensor relative to gravity, that is how that body segment is held, rather than how it moves; the authors call it motion and read it as sway, but they measure neither variability nor tremor within the window.

The contribution that holds up is methodological: the aiming phase can be cut automatically out of a continuous recording from wearable sensors, and what the body does in those two seconds can be related to the score of the arrow. Looking at the aiming phase rather than the instant of release, which is what almost every earlier study observed, is a reasonable move and is now the basis of several analysis tools.

On content, what the work shows best is that the model changes from one archer to another: for one it is the bow arm that counts, for another the drawing arm, for a third the pelvis. The authors conclude that individual work is needed, and across four girls that is a hypothesis rather than a result. The collective advice that closes the article, that holding the legs still matters more than holding the arm still, comes from a calculation on four averages and will not carry the weight the authors put on it.