FEM-PDNet: Evaluating Fixed and Learned Logit Fusion for Pre-Drowsiness Classification on UTA-RLDD

This study evaluates fixed and learned logit fusion for clip-level pre-drowsiness classification on an adaptively curated cohort from the University of Texas at Arlington Real-Life Drowsiness Dataset (UTA-RLDD). The target is an operationally defined visual state, not an independently established physiological condition or a prospective early-warning event. FaceMesh-Enhanced Multi-cue Pre-Drowsiness Network (FEM-PDNet) combines RGB temporal logits with behavior–motion logits derived from FaceMesh, optical flow, and frame differences. A blinded second review yielded 206 agreement-confirmed clips from 51 subjects. Sixteen configurations shared subject-disjoint five-fold partitions, five seeds, validation-selected checkpoints, and a fixed 0.5 threshold. FEM-PDNet obtained macro-F1 0.719±0.143 and pre-drowsy recall 0.578±0.282. Its difference from the tested concatenation implementation was 0.2344 (Holm-adjusted p=0.0260 across 16 contrasts), but fixed-weight controls did not establish an independent benefit of coefficient learning. The behavior–motion multilayer perceptron (MLP) obtained higher mean macro-F1 (0.823±0.017) than FEM-PDNet. Initialization and joint-feature-training analyses further characterized fusion sensitivity. Post-hoc branch diagnostics counted 15 corrected and 99 introduced errors relative to the feature branch across 1030 correlated clip–seed evaluations. These findings are conditional retrospective comparisons on a previously inspected cohort, not independent confirmation of generalization or the superiority of learned fusion.