Automatic analysis of neonatal facial behavior in real Neonatal Intensive Care Unit (NICU) environments is challenged by high data variability, visual noise, and the need for interpretable models suitable for clinical use. Facial expressions are a core component of neonatal pain, discomfort, and alertness assessments; however, their evaluation remains largely subjective in routine practice. In this study, a dataset of neonatal video recordings acquired under routine NICU conditions was analyzed, comprising 991 annotated videos after preprocessing. An attention-based convolutional neural network was employed for the frame-level binary classification of three clinically relevant facial traits: eye opening, eye frowning, and mouth opening. To address interpretability requirements in safety-critical settings, a dual explainability framework combining intrinsic attention mechanisms and post-hoc SHAP analysis was adopted. The proposed approach achieved strong performance, particularly for eye-related traits, despite challenging acquisition conditions. Explainability analyses showed that predictions were driven by anatomically meaningful facial regions, supporting model transparency and trustworthiness for clinical deployment.