🏃 Football Player Fatigue Prediction — Wearable IoT Sensors & ML

A machine learning framework that predicts three-class fatigue states in football players from wearable IoT sensor data — with Karvonen heart rate labeling, SMOTE balancing, LOSO validation, and a real-time coach substitution-alert dashboard.

Features • How It Works • Results • Installation • Usage • Architecture • Tech Stack • Contact

📌 Overview

Football Player Fatigue Prediction repurposes the PAMAP2 physical activity dataset for sports-specific fatigue monitoring. Instead of binary high/low labels, this system introduces a three-class Karvonen-based labeling scheme (LOW / MEDIUM / HIGH) grounded in cardiac physiology — with zero data leakage into model training.

The project offers three core contributions:

🏷️ Karvonen Auto-Labeler — assigns fatigue states from heart rate reserve % and movement intensity, then discards intermediates before training to prevent leakage
🌲 Personalized Random Forest — per-subject model fine-tuning that handles large inter-individual physiological variation
⚽ Coach Dashboard — 90-minute match simulation with real-time fatigue phase tracking and substitution alerts

✨ Features

🏷️ Karvonen Fatigue Labeling

Heart rate reserve % computed via the Karvonen formula (HRrest=60, HRmax=195)
Movement intensity as Euclidean norm of chest accelerometer
Three classes: LOW (light cardiac load + high movement), HIGH (cardiac drift), MEDIUM (otherwise)
Intermediate features discarded after labeling — fully leakage-free

📊 Sliding Window Feature Extraction

512-sample windows (~5.12s at 100 Hz), 50% overlap
11 statistical descriptors per channel: mean, std, min, max, median, Q1/Q3, RMS, range, MAD, energy
8 channels × 11 features = 88 features per window
Majority vote label assignment per window

⚖️ SMOTE Class Balancing

Original distribution heavily skewed toward MEDIUM (83.4%)
SMOTE (k=3) applied exclusively on training set
Balanced to 451 samples per class (1,353 total)

🌲 Four Models Evaluated

Random Forest — 200 trees, max depth 20, balanced class weights
SVM (RBF kernel) — C=10, gamma=scale, balanced class weights
LSTM — 2-layer (64→32 units), BatchNorm + Dropout, early stopping
Personalized RF — dedicated per-subject model with subject-level SMOTE

🔍 LOSO Cross-Validation

LeaveOneGroupOut with subject IDs as groups
Subject-independent generalization evaluation
Exceeds original PAMAP2 benchmark (97.96% vs ~88%)

⚽ Coach Decision Support Dashboard

90-minute match simulation with phase-by-phase fatigue profiling
LOW → MEDIUM transition through first half, HIGH accumulation in final 20 minutes
Visual heart rate, movement intensity, and fatigue class timeline
Substitution alert triggered when HIGH fatigue detected

🖥️ Demo

Fatigue Prediction

Window  → HR: 158bpm | Accel norm: 8.2 | Over 70 mins played
Output  → 🔴 HIGH Fatigue — Substitution Recommended

Personalized vs Global (Subject 3)

Global RF (LOSO) → F1: 48.3%
Personalized RF  → F1: 100.0%   (+51.7 pp improvement)

🚀 Installation

Prerequisites

Python 3.8 or higher
GPU recommended (Kaggle P100)

Step 1 — Clone the Repository

git clone https://github.com/bk1210/football-fatigue-prediction.git
cd football-fatigue-prediction

Step 2 — Install Dependencies

pip install -r requirements.txt

Step 3 — Download the Dataset

Get the PAMAP2 dataset from UCI: 👉 PAMAP2 Physical Activity Monitoring Dataset

Place the subject .dat files inside a PAMAP2_Dataset/Protocol/ folder in the project root.

Step 4 — Run the Notebook

jupyter notebook iot_fatigue.ipynb

Or upload directly to Kaggle and run with P100 GPU.

📖 Usage

Running the Full Pipeline

Open iot_fatigue.ipynb and run all cells — the notebook handles:

Data loading from PAMAP2 .dat files (9 subjects, 100 Hz)
Preprocessing — remove transitional segments, interpolate missing values, clip HR to [30, 220]
Karvonen-based three-class fatigue labeling (leakage-free)
Sliding window feature extraction (88 features per window)
SMOTE balancing on training set
Training and evaluating all four models (RF, SVM, LSTM, Personalized RF)
LOSO cross-validation across all 9 subjects
Feature importance analysis
90-minute football match simulation + coach dashboard

🔄 Pipeline

🏗️ Architecture

End-to-End Pipeline

PAMAP2 Raw Data (9 subjects, 100 Hz, 54 channels)
    │
    ▼
Preprocessing
[Remove activityID=0 | Interpolate NaN | Clip HR to 30-220 bpm]
    │
    ▼
Fatigue Label Creation (Karvonen Formula) ← Core Contribution
    ├─► HRR% = (HR - HRrest) / (HRmax - HRrest)
    └─► Movement Intensity = √(ax² + ay² + az²)
         │
         ▼
    LOW / MEDIUM / HIGH  →  intermediates discarded (leakage-free)
    │
    ▼
Sliding Window Feature Extraction
[512 samples | 50% overlap | 8 channels × 11 stats = 88 features]
    │
    ▼
SMOTE Balancing (training set only, k=3)
    │
    ├──► Random Forest (200 trees, depth 20)
    ├──► SVM RBF (C=10, gamma=scale)
    ├──► LSTM (64→32 units, BatchNorm, Dropout)
    └──► Personalized RF (per-subject, 80/20 split)
    │
    ▼
Evaluation: 70/30 Test Split + LOSO Cross-Validation
    │
    ▼
Coach Dashboard — 90-min Match Simulation + Substitution Alerts

Project Structure

football-fatigue-prediction/
│
├── iot_fatigue.ipynb                    # Full pipeline — labeling, training, LOSO, dashboard
├── requirements.txt                     # Python dependencies
└── README.md                            # Project documentation

📊 Results

Model Comparison — 70/30 Test Split

Model	Accuracy	F1-Macro
Random Forest	97.56%	76.23%
SVM (RBF)	94.15%	63.96%
LSTM	91.22%	57.85%
Personalized RF	97.87%	79.93%

LOSO Cross-Validation vs Baseline

Method	LOSO Accuracy	LOSO F1-Macro
Reiss & Stricker (PAMAP2 Benchmark)	~88.0%	—
This Work (RF)	97.96% ± 2.57%	87.55% ± 18.69%

Comparison with Base Papers

Criterion	Krishnaleela & Prakash (2025)	Liu et al. (2023)	This Work
Dataset	PAMAP2	Own (proprietary)	PAMAP2
Task	HAR (18 classes)	Fatigue (binary)	Fatigue (3-class)
SMOTE	No	No	✅ Yes
LOSO	No	No	✅ Yes
Personalized	No	No	✅ Yes
Performance	99.86% Acc	94.15% Acc	97.87% Acc / 79.93% F1

🛠️ Tech Stack

Technology	Purpose
Python 3.8+	Core language
scikit-learn	RF, SVM, SMOTE, LOSO, metrics
TensorFlow 2.19	LSTM model training
imbalanced-learn	SMOTE oversampling
NumPy / Pandas	Feature engineering, sliding window
Matplotlib / Seaborn	EDA, training curves, dashboard
Kaggle (P100 GPU)	Training environment

📦 Dependencies

numpy>=1.24.0
pandas>=2.0.0
scikit-learn>=1.3.0
imbalanced-learn>=0.11.0
tensorflow>=2.19.0
matplotlib>=3.7.0
seaborn>=0.12.0

Install with:

pip install -r requirements.txt

🔮 Future Improvements

Collect a football-specific dataset with GPS + continuous HR from professional players
Raw 100 Hz stream modeling with CNN-BiLSTM on full time-series
On-device edge inference using optimized BiLSTM
Multi-task learning — joint activity type + fatigue state prediction
Prospective validation linking HIGH-fatigue episodes to injury occurrence
Real-time MQTT streaming pipeline for live match deployment

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.

👤 Contact

Bharath Kesav R

📧 Email: bharathkesav1275@gmail.com
🐙 GitHub: @bk1210
🎓 Institution: Amrita Vishwa Vidyapeetham, Coimbatore

🙏 Acknowledgements

Reiss & Stricker (2012) — for the PAMAP2 dataset
Liu et al. (2023) — Base paper on wearable fatigue detection
Op de Beeck et al. (2018) — Personalization motivation

⭐ If you found this project useful, please give it a star on GitHub! ⭐

*Built with ❤️ for football/sports analystics and data analytics *

Name		Name	Last commit message	Last commit date
Latest commit History 6 Commits
Football Player Fatigue Prediction.ipynb		Football Player Fatigue Prediction.ipynb
README.md		README.md
pipeline.png		pipeline.png
requirements.txt		requirements.txt

Folders and files

Latest commit

History

Repository files navigation