Heart rate and archery score: Guru 2025 study

Thirty-two male archers from a residential sports institute in India, 15 elite and 17 non-elite, across recurve and compound, shot two sessions of 30 arrows indoors wearing a heart rate strap and filmed on video. For each arrow, 11 heart rate values were extracted, from 5 seconds before to 5 seconds after release: 21,120 values in all. To these the authors added 35 variables spanning body measurements, resting physiological data, psychological traits, sleep quality, previous competition score and the archer's category, to build twelve machine learning models predicting the session score.

The most interesting descriptive finding concerns the course of heart rate around the shot. In both sessions a slowing appears, and the point of maximum deceleration falls 2 or 3 seconds after release, in the follow-through phase, after which the rate rises again. In the correlation map, the heart rate values of all three phases correlate negatively with the score, while age, years of practice, years of training at the institute and previous competition score correlate positively.

Among the twelve models, Cat Boost with the cardiac values treated as linear was chosen as best: root mean squared error 6.262 points on a maximum score of 300, R squared 0.575 and what the authors call accuracy, 98.18%, derived from the mean percentage error. The mean predicted score is 279.4, against observed means of 279.4 and 279.5 in the two sessions.

With Shapley analysis the authors rank the variables by importance. The top ten are years of practice, resting systolic blood pressure, previous competition score, right-hand grip strength, age, heart rate 2 seconds before release, waist-to-hip ratio, concentration-disruption trait anxiety and heart rate 5 seconds after release. Two directions run against what the literature expects: higher values of concentration-disruption anxiety and lower values of grip strength push towards a higher predicted score.

The study is observational, carried out on male archers at a residential sports institute in India with approval from the ethics committee of the Armed Forces Medical College in Pune. The sample averages 21.8 years of age, 7.6 years of practice and 4.1 years of training at the institute; height 168.7 cm, weight 66.7 kg, waist-to-hip ratio 0.9, body fat 10.9%. Grip strength is 41.6 kg on the right and 40.8 on the left, resting heart rate 58.8 beats per minute, resting blood pressure 114.5 over 64.4 mmHg. On the anxiety traits, the mean scores are 13.9 for the somatic component, 13.1 for the cognitive and 8.2 for concentration disruption. The highest score in the previous competition is 660.4 ± 29.68 out of 720, with seven archers coming from international competitions, eleven from national ones and fourteen from below national level.

Each archer shot 30 arrows in two sessions, for 1,920 arrows in all, with a chest heart rate strap and video that allowed each beat to be placed relative to the shot. From every arrow, 11 heart rate values were extracted, one per second from minus 5 to plus 5 relative to release, divided into the three phases: aiming from minus 5 to minus 1, release at 0, follow-through from plus 1 to plus 5. The maximum score per arrow is 10, so the session score is out of 300. The detailed materials and methods are those of a pilot study by the same authors from 2020.

The independent variables number 35 and cover the archer's profile, anthropometric measures, physiological parameters, psychological traits, sleep quality and the 11 cardiac values. The models were built in Python 3.11.4 in two versions: one with the cardiac values as linear variables and one that also adds their second-degree polynomials, to account for the curvilinear course of the deceleration. Six models per version, twelve in all: linear regression, ElasticNet, random forest, XG Boost, Cat Boost and Light GBM, with hyperparameters chosen by grid search with cross-validation.

The session 1 data (960 rows) serve to train, the session 2 data (960 rows) to test. The quality measures are root mean squared error, R squared, mean absolute percentage error and what the authors call accuracy, defined as one hundred minus the percentage error. Variable importance is computed with the Shapley method, which supplies both an overall ranking and a breakdown per individual archer, shown for four archers in pairs within the same category.

The way the data were split means the test is not an out-of-sample test. The rows number 960 per session, that is 32 archers by 30 arrows, while the thing to be predicted is the session score, a single number per archer: the same value is therefore repeated thirty times. And the same 32 archers sit in both the training and the test set, with their constant variables, that is age, years of practice, blood pressure, grip strength, waist-to-hip ratio, anxiety traits and previous competition score. A model like this can recognise which archer it is dealing with from their fixed characteristics and return the total that archer usually shoots. This is not a computational error, it is the design: to know whether the model really predicts, whole archers would have to be held out, not a second session by the same people. The authors do not list this among their limitations.

The 98.18% accuracy means almost nothing. It derives from the mean percentage error on a quantity averaging 279 out of 300 that varies little, with a standard deviation between 7.4 and 9.8: being 5 points out on 279 is less than 2%. The best evidence that this number does not measure predictive ability sits in their own table: the Light GBM model has an R squared of 0.000, that is it explains nothing, and its «accuracy» is 97.27%. Between the model that predicts nothing and the one declared best there is less than one percentage point. The R squared of 0.575 is the only readable measure of the four, and it is moderate.

Among the most important variables, the second thing to note is which ones they are. Seven of the top ten are fixed characteristics of the archer, not things that happen during the shot, and one of them is the previous competition score, that is a past measurement of the very thing being predicted. Including it guarantees predictive power and says nothing useful to a coach: the model is largely stating that those who have shot for longer and have already scored well will score well. Only two of the top ten places are cardiac values taken during the shot, which is the paper's declared novelty.

Two directions of effect run against what the literature expects, and the authors report them without discussing them: a weaker grip and a higher score for concentration-disruption anxiety push towards a higher predicted score. On a model built this way these are more likely artefacts of a 32-person sample with many mutually correlated variables than real effects; either way they cannot be used as practical guidance.

More ordinary points remain. The sample is men only, 32 of them, from a single residential institute, and it is not as varied as the elite and non-elite split suggests: the mean score in the previous competition is 660 out of 720, that is 9.17 per arrow, so even the «non-elite» shoot well. In the text the age is given as 21.8 ± 2.0 years while the table reports 21.8 ± 5.63. Level and bow type enter the model as a category, so group membership appears among the predictors. The ethical approval is from 2016 and the pilot from 2020, so the data are several years old. The limitations the authors declare are two and minor: not having included sport-specific fitness variables, and having worked indoors under controlled conditions, which makes it hard to extend the model to outdoor competition.

The finding that holds regardless of the model is the course of heart rate around the shot, measured over 21,120 beats. There is a slowing, and the lowest point falls not at release but 2 or 3 seconds after it, inside the follow-through; then the rate rises again. It is a refinement of what is usually read, that the heart slows before the string is let go: in these archers the slowing continues past the arrow's departure. Anyone using a heart rate strap in training can look for it in their own traces, and the window to watch is from five seconds before to five after.

On the rest, caution pays. The model was not tested on archers it had not already seen: the same 32 people sit in the training and the test set, and seven of the ten most important variables are their fixed characteristics, the previous competition score included. What the model recognises well, in large part, is who you are, not how you are shooting today. The 98% accuracy does not change that: in the same table, the model explaining zero variance reaches 97.3%, because with scores all close to 279 out of 300 saying «the average» is already nearly right.

Two specific results should not be carried into training. That a weaker grip goes with a higher predicted score, and that more concentration-disruption anxiety goes with a higher score, are directions opposite to what the literature finds, and on 32 archers with dozens of correlated variables that is the kind of result that appears by chance.

A useful reference point remains: 30 arrows indoors, mean score 279 out of 300 in both sessions, that is 9.3 per arrow, in archers whose best competition had been 660 out of 720. And a methodological observation the work demonstrates well: with a heart rate strap and a video, beats can be placed relative to the release without laboratory instrumentation. That is the part of this work a coach can copy.