Automatic pain assessment for non-communicative children is in high demand. However, the availability of related training datasets remains limited. This study focuses on creating a large-scale dataset of pain facial expressions specifically for Chinese critically ill children and evaluating its utility using deep learning models. Data were gathered from two intensive care units at Children's Hospital of Fudan University. The dataset, named pain facial expression of critically ill children (PFECIC), includes 119 pain expression videos and 6951 images collected from 53 children between December 2022 and January 2023. All videos and images were independently triple labeled according to five pain levels. The PFECIC dataset was evaluated through deep learning experiments, demonstrating strong performance metrics: 88.3% accuracy, 88.3% precision, 88.7% recall, an F1-score of 88.5%, and a false-positive rate of 3.0%. Prediction errors were mostly associated with labels close to the true values. Comparative analysis with the classification of pain expressions (COPE) dataset highlighted the superiority of PFECIC in terms of accuracy, validity, and comprehensiveness.