Multi‑Season Injury Prediction in Football Using Machine Learning

---

Tezin Türü: Yüksek Lisans

Tezin Yürütüldüğü Kurum: Uluslararası Final Üniversitesi, Engineering, Software Engineering, Kıbrıs (Kktc)

Tezin Onay Tarihi: 2026

Tezin Dili: İngilizce

Öğrenci: Abdalrahman Siddig Ali Bashir

Danışman: Ghazaal Sheikhi

Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu

Özet:

Every season, thousands of footballers at the elite level get injured, and the cost of those injuries is enormous, with a review covering the five seasons up to 2024/25 recording 22,596 injuries across Europe’s top five men’s leagues at a combined financial toll of more than £2.97 billion in lost income and wasted resources. Researchers and clubs have tried to get ahead of this problem using wearable monitoring systems and machine learning, yet most published studies run into the same two obstacles, namely that they draw on small single-club datasets of only 25 to 40 players over one or two seasons which makes it hard to generalise findings, and that the wearable hardware they depend on is expensive with data that is proprietary and out of reach for most clubs and researchers. This thesis takes a different route by relying entirely on publicly available match statistics, pulling data from Understat and Transfermarkt to build a dataset of 47,477 player-match records covering 1,017 English Premier League players across five seasons from 2020/21 to 2024/25, and since injury risk does not come from a single match but builds up gradually over time each player’s appearances are treated as a time series, which allows the model to pick up on workload accumulation, fatigue patterns, and the toll of repeated match exposure. A Long Short-Term Memory (LSTM) network with an attention mechanism handles this sequential task and is compared against Decision Tree and Random Forest baselines, with the prediction framed as a binary classification problem where, given a player’s recent run of matches up to a specific game-week, the model determines whether they will be injured within the next 30 days. Across the five seasons, the models achieve AUC values ranging from 0.44 to 0.75, with the LSTM with attention reaching the best AUC in three of the five seasons, while the Random Forest achieves the single highest result at 0.746, and a three-tier explainability framework combining SHAP, attention weights, and Integrated Gradients consistently points to cumulative workload and recent performance load as the dominant injury- risk signals across all model families.