This paper presents a deep learning approach for detecting head impacts in ice hockey from 2D game videos. Traumatic brain injuries, including concussions and repetitive head impacts, pose significant health risks in sports. Understanding the relationship between head impact features (magnitude and frequency) and outcomes such as mental health decline, cognitive deficits, and Chronic Traumatic Encephalopathy (CTE) requires extensive datasets. Tracking head impacts during games is challenging, and available tools are impractical for most leagues due to cost and equipment constraints. Utilising game videos for head impact detection offers a viable solution. A methodology combining player detection and tracking with a Long-term Recurrent Convolutional Network (LRCN) for head impact detection is proposed. Our player detection model achieved high precision and recall scores, facilitating accurate tracking. The YOLOv8x object detection model yielded precision, recall, mAP50, and mAP50-95 scores of 0.97, 0.97, 0.99, and 0.95, respectively. The StrongSORT tracking algorithm used for player tracking minimised ID switches, important for precise tracking in dynamic sports environments. The LRCN-based head impact detection model showed promising results, with an accuracy of 87% and a loss of 0.04. Future work involves refining dataset creation to address data imbalance and exploring alternative deep learning models like CONVLSTM and 3DCNN for improved performance.