Cross-Patient Evaluation of CNN-Based Facial Expression Recognition for Intubated ICU Patients Using Leave-One-Patient-Out Validation

Intubated Intensive Care Unit (ICU) patients experience severe communication limitations due to medical device occlusions and altered physiological conditions, which make facial expression interpretation challenging for healthcare professionals. Facial Expression Recognition (FER) based on Convolutional Neural Network (CNN) offers a promising solution; however, its reliability in real ICU environments critically depends on the model's ability to generalize across patients with heterogeneous clinical characteristics. This study evaluates the cross-patient performance of a CNN-based FER system using a two-stage transfer learning approach and a rigorous patient-independent validation protocol, Leave-One-Patient-Out Cross-Validation (LOPOCV). Four ImageNet backbones, ResNet50, DenseNet121, MobileNetV2, and EfficientNetB0, were trained on 33 videos from 10 intubated ICU patients mapped into three FER classes. Conventional frame-level training yielded high accuracies, with DenseNet121 achieving 100% under non-patient-independent evaluation. An ablation study showed that partially unfreezing the final backbone layers produced the most stable configuration for small-scale clinical datasets. However, LOPOCV revealed a marked performance decrease (mean accuracy ≈ 45%), highlighting identity leakage inherent in frame-level evaluation and limited cross-patient generalization under high occlusion and inter-patient variability. These findings establish a more realistic, patient-independent performance baseline for FER in intubated ICU patient settings and underscore the need for temporal architectures and multimodal strategies to improve robustness and clinical reliability.