Physiology and high-performing archers: Taha 2018

Fifty young Malaysian archers, 37 boys and 13 girls aged 13 to 20, were divided into three groups, high, medium and low potential, by a cluster analysis that combined shooting score with five measures taken at rest or self-reported: heart rate, systolic and diastolic blood pressure, breathing rate and calorie intake. A classification algorithm, k-nearest neighbour, then tried to identify each archer's group from the same measures.

The version using a single neighbour (k = 1) classified 84% of the archers correctly, the one with 10 neighbours 78%, and the one with 100 neighbours 56%; the last placed everyone in the high-potential group, which must therefore have made up about 56% of the sample. According to the authors the best group had the highest resting breathing rate, calorie intake and score, while heart rate and blood pressure did not separate the groups. The paper reports neither the mean values of the measures nor the number of archers in each group: the differences can be seen only in a chart.

The score came from a single end: four practice arrows and then six scored arrows at 50 metres, on a range set up to simulate competition. Bow type is not given. Blood pressure and heart rate were measured in the morning, seated, before any exercise, with an automatic Omron monitor on the left arm; breathing rate by counting chest movements by hand for one minute, three times, keeping the highest value. Diet came from a three-day food diary, two weekdays and one weekend day, kept by the archers and analysed with nutrition software.

The three groups were formed by hierarchical cluster analysis. Classification used three versions of k-nearest neighbour, with 1, 10 and 100 neighbours and Euclidean distance, assessed by five-fold cross-validation; the outcomes are overall accuracy and precision and recall for each group.

The design has a basic flaw: the groups were built from the same variables that the algorithm then uses to recognise them, score included. That a classifier recovers groups drawn by another analysis on the same data is close to inevitable, and does not show that those measures predict an archer's level. The 100-neighbour version, with 50 archers, makes no practical sense: with so little data it assigns everyone to the largest group by construction.

Six arrows at 50 metres are far too few to define an archer's level, and calling a group built on one day's score "high potential" is a choice of words, not a finding. There are no tables of the measured values and no statistical tests of differences between groups. The reading that a faster resting breathing rate indicates better control is the authors' own and is debatable, since a lower resting rate is usually associated with a more relaxed state. The descriptive data also raise doubts: an age range of 13 to 20 with a mean of 17.0 and a standard deviation of 0.56 is hard to reconcile, and the figure caption refers to two classes instead of three.

For archers the study offers no guidance. The authors' message, that resting breathing and calorie intake distinguish the better archers, is not supported by the design: the groups were also built on score, and the values and statistical comparisons are missing. Resting blood pressure and heart rate did not separate the groups even by the authors' account.

For a coach who hears about algorithms that can spot talent from a few physiological measures, it is a useful example of what to check: how the groups were defined, on how many arrows, and whether the model was tested on a separately collected set of archers. Here the answers are, in order, with the same variables used for classification, on six arrows, and no.