José Eduardo Teixeira ¹ Samuel Encarnação ² Luís Branquinho ³ Ricardo Ferraz ⁴ Daniel L. Portella ⁵ Diogo Monteiro ⁶ Ryland Morgans ⁷ Tiago M. Barbosa ⁸ António Miguel Monteiro ⁸ Pedro Forte ^2*

• ¹ Instituto Politécnico da Guarda, Guarda, Guarda, Portugal
• ² Higher Institute of Educational Sciences of the Douro, Penafiel, Portugal
• ³ Polytechnic Institute of Portalegre, Portalegre, Portalegre, Portugal
• ⁴ University of Beira Interior, Covilhã, Castelo Branco, Portugal
• ⁵ Universidade Municipal de São Caetano do Sul, São Caetano do Sul, Sao Paulo, Brazil
• ⁶ ESECS - Polytechnique of Leiria, Leiria, Leiria, Portugal
• ⁷ Institute of Coaching and Performance, University of Central Lancashire, Preston, United Kingdom
• ⁸ Polytechnic Institute of Bragança (IPB), Bragança, Braganca, Portugal

A promising approach to optimizing recovery in youth football has been the use of machine learning (ML) models to predict recovery states and prevent mental fatigue. This research investigates the application of ML models in classifying male young football players aged under (U)15, U17 and U19 according to their recovery state. Weekly training load data were systematically monitored across three age groups throughout the initial month of the 2019-2020 competitive season, covering 18 training sessions and 120 observation instances. Outfield players were tracked using portable 18 Hz global positioning system (GPS) devices, while heart rate (HR) was measured using 1 Hz telemetry HRbands. The Rating of Perceived Exertion (RPE 6-20), and Total Quality Recovery (TQR 6-20) score were employed to evaluate perceived exertion, internal training load and recovery state, respectively. Data pre-processing involved handling missing values, normalization, and feature selection using correlation coefficients and Random Forest (RF) classifier. Five ML algorithms [K-nearest Neighbors (KNN), Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), RF, and Decision Tree (DT)], were assessed for classification performance. K-fold method was employed to crossvalidate the ML outputs. The results indicated high accuracy for this ML classification model (73 to 100%). Feature selection highlighted critical variables, we implemented the ML algorithms considering a panel of ten variables (U15, U19, body mass, accelerations, decelerations, training weeks, sprint distance, and RPE). These ten features were included according to their percentage of importance (3-18%). The results were cross validated with good accuracy across five folds (79%). In conclusion, the five ML models in combination weekly data demonstrated the efficacy of wearable device-collected features was an efficient combination in predicting fooball players' recovery states.