{Reference Type}: Journal Article
{Title}: Football Movement Profile-Based Creatine-Kinase Prediction Performs Similarly to Global Positioning System-Derived Machine Learning Models in National-Team Soccer Players.
{Author}: Schuth G;Szigeti G;Dobreff G;Pašić A;Gabbett T;Szilas A;Pavlik G;
{Journal}: Int J Sports Physiol Perform
{Volume}: 19
{Issue}: 9
{Year}: 2024 Sep 1
{Factor}: 4.211
{DOI}: 10.1123/ijspp.2024-0077
{Abstract}: <B>OBJECTIVE: </B>The relationship between external load and creatine-kinase (CK) response at the team/position or individual level using Global Positioning Systems (GPS) has been studied. This study aimed to compare GPS-derived and Football Movement Profile (FMP) -derived CK-prediction models for national-team soccer players. The second aim was to compare the performance of general and individualized CK prediction models.<BR><B>METHODS: </B>Four hundred forty-four national-team soccer players (under 15 [U15] to senior) were monitored during training sessions and matches using GPS. CK was measured every morning from whole blood. The players had 19.3 (18.1) individual GPS-CK pairs, resulting in a total of 8570 data points. Machine learning models were built using (1) GPS-derived or (2) FMP-based parameters or (3) the combination of the 2 to predict the following days' CK value. The performance of general and individual-specific prediction models was compared. The performance of the models was described by R2 and the root-mean-square error (RMSE, in units per liter for CK values).<BR><B>RESULTS: </B>The FMP model (R2 = .60, RMSE = 144.6 U/L) performed similarly to the GPS-based model (R2 = .62, RMSE = 141.2 U/L) and the combination of the 2 (R2 = .62, RMSE = 140.3 U/L). The prediction power of the general model was better on average (R2 = .57 vs R2 = .37) and for 73% of the players than the individualized model.<BR><B>CONCLUSIONS: </B>The results suggest that FMP-based CK-prediction models perform similarly to those based on GPS-derived metrics. General machine learning models' prediction power was higher than those of the individual-specific models. These findings can be used to monitor postmatch recovery strategies and to optimize weekly training periodization.