Biomechanics and talent in archers: Taha 2018 study

Fifty young Malaysian archers, 37 boys and 13 girls aged 13 to 20, shot one end of six arrows at 50 metres while wearing a set of sensors, and the authors tried to pick out those with the most potential using a machine learning algorithm. There were five measures: body sway, bow movement, the activity of one finger flexor and one finger extensor muscle, and static balance. A first algorithm (k-means) split the archers into two groups, high potential and low potential, on the basis of these measures and the score.

Three support vector machine models, differing in the scale of the function used, were then trained to recognise the two groups from the five measures. With five-fold cross-validation the finest model classified 70% of the archers correctly and the other two 64%. The authors write that the high-potential group does better on almost every measure, but in the box plots the only clear difference is the score: for sway, bow movement, muscle activity and balance the two groups' distributions largely overlap, and no statistical test is reported.

After four practice arrows, each archer shot six arrows at 50 metres in conditions set up to resemble competition. The sensors were four Shimmer units: an accelerometer fixed to the pelvis for body sway, a gyroscope inside a glove on the hand for bow movement, and two pairs of electrodes recording electromyography from the finger extensor on the left and the finger flexor on the right. Data were streamed over Bluetooth to an Android phone at 51.2 samples per second. Static balance was measured separately with the stork test, standing on one leg.

For classification the authors used radial basis function models at three different scales, calculated from the number of predictors, which is five. The only outcome reported is the percentage of archers classified correctly, with the corresponding confusion matrices.

The main limitation lies in the design: the two groups were created from the same measures that the models then use to recognise them, so the result shows how well the algorithm recovers a split built from those data, not whether those measures indicate real talent. Potential was never checked over time, for instance against competition results in later years. The paper does not say how many archers ended up in each group: if one of them made up around 64% of the sample, a model that always gave the same answer would match the accuracy of the two simpler models.

Much basic information is missing: the discipline, handedness, how the measures were calculated, the groups' numerical values and any statistical test between them. The score comes from just six arrows. Electromyography recorded at 51.2 samples per second is far below the rates normally used for this signal, and a mean age of 17.0 years with a standard deviation of 0.56 is hard to reconcile with a range of 13 to 20. In the graphs, moreover, extensor amplitude and bow velocity are plotted as negative values, with no explanation. The authors write in the discussion that muscle activity, bow movement and sway influence accuracy, but this design cannot show that.

For anyone who shoots it changes nothing, and for anyone coaching young archers it does not offer a selection tool. It is an example of the early attempts to apply machine learning to talent identification in archery, using cheap wearable sensors. The underlying idea, that youngsters who are steadier in body and bow score more, is in line with other studies on balance and aiming stability, but this paper does not strengthen it, because the groups were built by combining stability and score in the first place, and in the graphs the two groups look very similar on exactly the stability measures.

A model that misclassifies three or four youngsters in ten, and has never been compared with what those youngsters went on to do, says nothing about who will become a good archer.