Justifying model complexity: evaluating transfer learning against classical models for intraoperative nociception monitoring under anesthesia

BackgroundAccurate intraoperative detection of nociceptive events is essential for optimizing analgesic administration and improving postoperative outcomes. While deep learning models promise to capture complex temporal dynamics of physiological signals, their added complexity may not always yield clinically meaningful gains compared to well-engineered classical approaches. MethodsWe evaluated two classical supervised models--L1-regularized logistic regression and Random Forests (with and without drug dosing features)--against a Temporal Convolutional Network (TCN) transfer-learning framework. We used a dataset of 101 adult surgical cases (~50,000 annotated nociceptive events over ~18,500 minutes) sourced from PhysioNet that tracked 30 physiologic and 18 drug-related features in 5-second windows. All models were assessed under a leave-one-surgery-out cross-validation, with AUROC and AUPRC as primary metrics. We further examined probability calibration (Platt scaling, isotonic regression) and four ensemble strategies--including a meta-learner, MLP, and a feature-conditioned gated network--to quantify the benefit of deep personalization. ResultsDrug-aware Random Forests achieved the highest discrimination (AUROC 0.716; AUPRC 0.399), significantly outperforming the TCN transfer-learning model (AUROC 0.649; AUPRC 0.311). Isotonic calibration reduced expected calibration error by over 80% but did not alter discrimination. None of the ensemble methods surpassed the standalone Random Forest, and the gated network consistently assigned > 84% weight to the classical model. Permutation importances revealed critical mechanistic features related to sympathetic physiologic response. ConclusionsIn this head-to-head benchmark, interpretable classical models on expertly curated features matched or exceeded the performance of a complex deep learning approach, while offering superior computational efficiency and transparency. These findings underscore the importance of rigorous comparative evaluation before adopting high-complexity AI solutions in clinical practice. Data Availability StatementAll data was sourced from Subramanian et al. on PhysioNet under data usage agreement and proper citations in the manuscript. All code and analysis can be provided upon reasonable request. The authors plan to upload their code on GitHub. Competing Interests StatementThe authors declare no conflict of interests or financial stakes in this work. Funding DisclosuresThere is no funding to declare for this work.