Catastrophic injuries in Thoroughbred flat racing pose significant challenges to equine welfare and the sustainability of the racing industry, particularly in Korea where year-round racing on sand tracks is practiced. This study analyzed 223,117 race starts at Seoul and Busan racetracks from 2013 to 2024, identifying 1,869 injury cases (0.84%). Statistical analysis revealed five principal risk factors associated with catastrophic injuries: reduced racing frequency in the prior year, older racing age, wet track conditions, history of veterinary examinations, and prolonged intervals between races. A predictive model was developed using the XGBoost algorithm with 10-fold cross-validation, achieving outstanding performance metrics including 99.9% accuracy, 99.1% precision, 99.7% recall, 99.4% F1-score, and an area under the curve of 1.000. The highest risk group exhibited injury rates approximately ten times greater than the baseline, effectively identifying horses at elevated risk. Key high-risk scenarios involved older fillies racing in spring after trainer changes and trainers associated with elevated injury incidences. This study demonstrates that advanced machine learning techniques can enable early detection of high-risk horses, facilitating targeted preventive measures that may reduce catastrophic injuries and improve equine welfare. Further developments involving deep learning and integration of real-time data are anticipated to enhance prediction accuracy and practical application in racing in Korea.