Young archers and physiology data: Taha 2018 study

Fifty young Malaysian archers, 37 boys and 13 girls aged 13 to 20, shot six scoring arrows at 50 metres after four practice shots. Before shooting, resting heart rate, systolic and diastolic blood pressure and breathing rate were measured, and calorie intake was worked out from three-day food diaries. The authors split the archers into two groups, high and low potential, with an automatic clustering algorithm (k-means), then trained a support vector machine to recognise the two groups from the five physiological measures.

The linear model classified 94% of the archers correctly, with a sensitivity of 88.89%, a specificity of 96.88% and a Matthews coefficient of 0.8690; the quadratic and cubic models reached 88 and 84%. The percentages fit two groups of 18 and 32 archers, although the paper does not say so. According to the graphs, the high-potential group had higher values on every measure except diastolic pressure, which means higher resting heart rate, systolic pressure and breathing rate as well.

Blood pressure and heart rate were taken in the morning, before any exercise, with an automatic Omron monitor on the left arm and the archer seated. Breathing rate was counted by a therapist with a hand on the chest for one minute, three times, keeping the highest value. For diet the archers filled in a three-day diary, two weekdays and one weekend day, analysed with the Nutritionist Pro software. The k-means clustering used the physiological measures and the shooting score together; the classification used the five physiological measures, with fivefold cross-validation. There is an inconsistency over the models: the abstract and results refer to linear, quadratic and cubic functions, the methods to three radial basis variants.

The basic flaw is that the reasoning goes in a circle. The groups were formed partly from the same physiological measures the model then uses to recognise them, so a high rate of correct classification was largely to be expected and does not show that these measures say anything about shooting ability. The score, moreover, comes from just six arrows, and the mean values of the two groups are never given as numbers.

The results also run against the authors' reading: the better group had higher resting heart rate, systolic pressure and breathing rate, and the authors interpret faster breathing as a sign of good control without any measurement to support it. The conclusions stress the importance of nutrition, but the only dietary figure is calorie intake from a three-day diary filled in by teenagers. The standard deviation for age (0.56 years with ages from 13 to 20) is not plausible. The model is never tested on archers other than those used to build it, and bow type is not reported.

For an archer or coach the study offers nothing usable. Resting heart rate, blood pressure and breathing, measured once in the morning, do not become markers of talent just because an algorithm recognises groups built from those same measures. That the better archers in the sample had higher resting values does not fit the authors' reading, but with this design no conclusion can be drawn from it either way.

The paper is mainly of interest as an example of machine learning used in a way where testing the model is not independent of the data that built it.