Proactive Injury Risk Assessment in High-Performance Athletics Using Multi-Modal Wearable Sensor Data: A Comparative Analysis of Ensemble Learning and Deep Temporal Architectures

Warik, Akshada Maneesh (2025) Proactive Injury Risk Assessment in High-Performance Athletics Using Multi-Modal Wearable Sensor Data: A Comparative Analysis of Ensemble Learning and Deep Temporal Architectures. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (2MB) | Preview
Preview	PDF (Configuration Manual) Download (1MB) | Preview

Abstract

In the contemporary landscape of high-performance sports, the margin between peak physical conditioning and musculoskeletal injury is increasingly narrow. As athletes push physiological boundaries, the necessity for a paradigm shift from reactive medical treatment to proactive risk assessment has become paramount. This research presents a robust predictive analytics framework leveraging multi-modal biometric data harvested from wearable sensors. Utilizing a synthetic dataset of 1,000 high-fidelity training sessions, we monitor critical physiological indicators including Heart Rate Variability (HRV), Blood Oxygen Saturation (SpO2), Skin Temperature, Biomechanical Impact Force, and a derived Cumulative Fatigue Index.

We conduct a rigorous comparative analysis between traditional feature-based machine learning models (Logistic Regression, Support Vector Machines, Random Forests) and advanced deep learning architectures (Baseline Long Short-Term Memory networks and LSTM Autoencoders) to detect pre-injury states. Our Exploratory Data Analysis (EDA), reveals statistically significant differences (p ¡ 0.001) in Cumulative Fatigue Index, Impact Force, and Injury Risk Score between injured and non-injured sessions, while other metrics such as SpO2 and Respiratory Rate remain non-discriminative. Furthermore, we introduce a granular time-stamp prediction mechanism capable of identifying risk probabilities at specific millisecond intervals using sequence models.

LSTM Autoencoder attains a Recall of 1.00 with an F1-Score of 0.7616, prioritizing sensitivity to injury cases, while SVM and Random Forest baselines consistently achieve Recall values above 0.88. This study provides a blueprint for integrating real-time AI into coaching dashboards, combining interpretable ensemble methods with highly sensitive temporal deep learning pipelines.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Yaqoob, Abid UNSPECIFIED
Uncontrolled Keywords:	Sports Analytics; Wearable Technology; LSTM; LSTM Autoencoder; Random Forest; SVM; Logistic Regression; Injury Prevention; Biometrics; Time-Series Analysis
Subjects:	T Technology > TK Electrical engineering. Electronics. Nuclear engineering > Biometric Identification Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning G Geography. Anthropology. Recreation > GV Recreation Leisure > Sports
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	09 Sep 2026 11:16
Last Modified:	09 Sep 2026 11:16
URI:	https://norma.ncirl.ie/id/eprint/9928

Actions (login required)

View Item