From:  Artificial intelligence for pain assessment via facial expression recognition (2015–2025): a systematic review

 Summary of the included studies.

Source/YearDataset/nPopulationApproach/Inclusion criteriaAI methodOutcome (as reported)
Bargshady et al. [17], 2024Lab datasets (AI4PAIN: 51; BioVid: 87)AdultsAcute Pain Datasets (video-based)Video vision transformers (ViViTs)Accuracy 66.9% (AI4PAIN), 79.9% (BioVid), outperforming ResNet baselines
Bargshady et al. [18], 2020UNBC-McMaster, MIntPAINAdultsBenchmark datasetsEnsemble DL model (CNN + RNN hybrid, EDLM)Accuracy > 89%, ROC 0.93; robust vs. single-stream CNN
Bellal et al. [19], 2024ICU, 30 patientsCritically ill, non-communicative adultsNEVVA® pilot device calibrationAI-based computer vision integrated in devicesFeasible, device calibrated against expert assessment
Benavent-Lledo et al. [20], 2023UNBC, BioVidAdultsPublic pain expression datasetsTransformer-based computer visionAccuracy > 96% (UNBC), > 94% (BioVid); high precision, recall
Cascella et al. [21], 2024Oncology + public datasets (Delaware, UNBC)Cancer patients, adultsBinary classifier using AUsNeural network (17 AUs, OpenFace)Accuracy ~94%; AUROC 0.98
Cascella et al. [22], 2024OncologyAdult cancer patientsVideo + audio (facial + speech)Multimodal AI (speech emotion + facial expression)Feasibility shown; early accuracy promising
Cascella et al. [23], 2023Clinical feasibility (real-time)AdultsReal-time pain detection from facial videosYOLOv8 object detectionFeasible, metrics reported with good accuracy (JPR)
Casti et al. [24], 2019Clinical/Lab settingAdultsAutomatic pain detection calibrationDL-based system (CNN)Benchmarked; addressed inter-/intra-observer variability
Casti et al. [25], 2021Public dataset (video pain sequences)AdultsLandmark time-series analysisTransfer entropy (TE) + ML classifiersTE-based approach improved accuracy, robust to noise
Chen et al. [26], 2022UNBC + lung cancer datasetAdults, including patients with lung cancerPain-related AUsWeakly supervised MIL/MCILAccuracy 87%, AUC 0.94 (UNBC); validated also on clinical lung cancer data
Dutta and M [27], 2018UNBC + live videoAdultsReal-time video-based pain recognitionHybrid DL modelValidated in real-time; high accuracy reported
Ghosh et al. [28], 2025UNBC, BioVid + VIVAE (audio)AdultsMultimodal (facial + audio)Ensemble DL with CNN + fusionAccuracy up to 99.5% (3-class), 87.4% (5-class); audio peak 98%
Guo et al. [29], 2021Cold pressor experiment; 29 subjectsAdultsCold pain inductionCNN (Inception V3, VGG-LSTM, ConvLSTM)F1 score 79.5% (personalized ConvLSTM)
Heintz et al. [30], 2025Perioperative, multicenter (503 pts)Adults perioperativeComputer vision nociception detectionCNN-basedStrong AUROC, external validation, and feasibility proven
Mao et al. [31], 2025UNBCAdultsPain intensity estimationConv-Transformer (multi-task joint optimization)Outperformed SOTA; improved regression + classification
Mieronkoski et al. [32], 202031 healthy volunteers, experimentalAdultsPain induction + sEMGML (supervised on muscle activation)Modest c-index 0.64; eyebrow/lip muscles most predictive
Morsali and Ghaffari [33], 2025UNBC, BioVidAdultsPublic Pain DatasetsErAS-Net (attention-based DL)Accuracy 98.8% (binary, UNBC); 94.2% (4-class); cross-dataset BioVid 78%
Park et al. [34], 2024155 pts post-gastrectomyPostoperative adultsClinical recordingsML models (facial, ANI, vitals)AUROC 0.93 (facial); better than ANI/vitals
Pikulkaew et al. [35], 2021UNBC datasetAdultsSequential facial imagesCNN (DL motion detection)Precision: 99.7% (no pain), 92.9% (becoming pain), 95.1% (pain)
Rezaei et al. [36], 2021Dementia patients, LTC settingOlder adults, dementiaUnobtrusive video datasetDeep learning + pairwise/contrastive trainingOutperformed baselines; validated on dementia cohort
Rodriguez et al. [37], 2022UNBC + CKAdultsRaw video framesCNN + LSTMOutperformed SOTA AUC (UNBC); competitive on CK
Semwal and Londhe [38], 2024Multimodal datasetAdultsFacial + multimodal integrationMulti-stream spatio-temporal networkShowed robust multiparametric pain assessment
Tan et al. [39], 2025200 patientsAdults perioperative/interventionalVideo recording (STA-LSTM)STA-LSTM DL networkAccuracy, sensitivity, recall, F1 ≈ 0.92; clinical feasibility
Yuan et al. [40], 2024ICU, public + 2 new datasetsCritically ill adults (ventilated)Facial occlusion managementAU-guided CNN frameworkSuperior performance in binary, 4-class, regression tasks
Zhang et al. [41], 2025503 postop patients + volunteersAdults postoperativeClinical Pain Dataset (CPD; 3,411 images) + Simulated Pain Dataset (CD)VGG16 pretrainedAUROC 0.898 (CPD severe pain), 0.867 (CD); software prototype developed

AI: artificial intelligence; ResNet: Residual Network; UNBC: University of Northern British Columbia Pain Expression dataset; MIntPAIN: Multimodal International Pain dataset; DL: deep learning; CNN: convolutional neural network; RNN: recurrent neural network; EDLM: ensemble deep learning model; ROC: receiver operating characteristic; ICU: intensive care unit; NEVVA: Non-Verbal Visual Analog device; AUs: action units; AUROC: area under the receiver operating characteristic curve; YOLOv8: You Only Look Once version 8; JPR: Journal of Pain Research; ML: machine learning; AUC: area under the curve; MIL: multiple instance learning; MCIL: multiple clustered instance learning; VIVAE: Visual and Vocal Acute Expression dataset; VGG: visual geometry group; LSTM: long short-term memory; ConvLSTM: convolutional long short-term memory; SOTA: state-of-the-art; sEMG: surface electromyography; ErAS-Net: enhanced residual attention-based subject-specific network; ANI: analgesia nociception index; LTC: long-term care; CK: Cohn-Kanade dataset; STA-LSTM: Spatio-Temporal Attention Long Short-Term Memory; CD: Control Dataset.