Patient-independent Evaluation of CNN-LSTM for Facial Expression Recognition in Intubated ICU Patients

This study evaluates CNN-LSTM architectures for facial expression recognition (FER) in intubated ICU patients under strict patient-independent Leave-One-Patient-Out Cross-Validation (LOPOCV), aiming to establish a conservative baseline for clinical video analysis.The dataset comprises 33 videos from 10 ICU patients annotated by medical experts, with ten clinical expression labels consolidated into three FER target classes based on clinical relevance.Spatial facial features were extracted using ImageNet-pretrained CNN backbones, while temporal dependencies were modeled using an LSTM module.Controlled baseline comparisons were conducted between static CNN models, temporal aggregation (CNN-GTP), and sequential temporal modeling (CNN-LSTM) under identical preprocessing and training configurations.Four CNN backbones were evaluated with fully frozen feature extractors to isolate the effect of temporal modeling.The results indicate that temporal modeling improves macro-averaged performance, while overall performance remains constrained by limited data size, class imbalance, and inter-patient variability.These findings underscore the importance of patient-independent evaluation and establish a conservative, clinically grounded baseline for future ICU-oriented FER research.