Temporal difference encoding for subject-independent mental workload assessment from fingertip PPG

This is the authors' abstract. We don't add key points for this paper.

Photoplethysmography (PPG) can support unobtrusive mental-workload monitoring, but person-specific physiology and evaluation design complicate subject-independent claims. We present a fingertip-PPG framework for classifying 0-back versus 3-back task conditions, used here as proxies for lower versus higher experimentally induced workload, in previously unseen people. The strict inductive mode represents each five-minute MAUS trial with six 50-s windows and five adjacent temporal-difference blocks, yielding 1276 features, and evaluates a linear SVM with fully nested leave-one-subject-out (LOSO) validation. The revised primary analysis uses a regular 16-candidate grid and achieved 79.55% accuracy (70/88), 79.50% macro-F1, ROC-AUC 0.821, a subject-bootstrap 95% CI of 71.56%–87.50%, 100% coverage, and zero rejection. The originally submitted narrowed candidate set produced 81.82% accuracy and is retained as a secondary sensitivity/development result. A locked 5000-permutation test gave p = 1 / 5001 ≈ 0.000200 . An optional label-free 696-dimensional subject-batch RBF-SVM mode reached 85.23% accuracy and ROC-AUC 0.890. Two controls separated representation from normalization: adding subject-batch normalization to the 1276-dimensional temporal representation reduced accuracy to 69.32%, whereas removing it from the matched 696-dimensional RBF pipeline reduced accuracy to 62.50%. These contrasts show that the effect of subject-batch normalization is representation dependent and are consistent with a role for reducing subject-specific variation, but they do not establish one sole causal mechanism or mechanistic equivalence between the two pipelines. Independent held-out CogLoad replication using pulse-derived HR/RR reached 85.00% accuracy and ROC-AUC 0.934.