Objective

The objective was to summarize and critically evaluate the evidence of published studies between 2015 and 2025 on the application of AI for pain assessment through the analysis of facial expressions. The review addressed the following question: “What AI-based methods have been applied to detect or assess pain from facial expressions, and what is their accuracy and applicability in clinical or experimental settings?”.

Literature search strategy

The search strategy was developed in line with the PRISMA (Preferred Reporting Items for Systematic reviews and Meta-Analyses) guidelines [10].

The systematic search was conducted in the main biomedical and computer databases, including PubMed/MEDLINE, Scopus, Cochrane Library, and Web of Science, integrated by IEEE Xplore for the literature in engineering and computer science. To ensure completeness, references of the selected articles and relevant citations were also manually checked. Only studies published in the period 1 January 2015–31 July 2025 were considered. The latest electronic search was completed in July 2025. Search records were managed in the Rayyan software [11].

Eligibility criteria

We included articles addressing techniques for APA through facial expression analysis. Studies conducted on adult patients of different clinical conditions, without restrictions related to the care setting (e.g., hospital, intensive care, outpatient, or experimental setting), were considered eligible. A mandatory comparator was not required; however, when present, traditional pain assessment tools (e.g., visual analogue scale, numeric rating scale, clinical observation) were evaluated.

Studies not related to the assessment of pain, opinion-only articles, editorials, conference abstracts without complete data, works that used non-AI-based approaches, or that employed AI in contexts unrelated to facial recognition were excluded.

Only peer-reviewed studies published in English were included. Unpublished articles, not peer-reviewed manuscripts, technical reports, or grey literature were not considered.

Search strategy

The search strategy was developed according to the PCC (Population, Concept, Context) framework and tailored for each database.

The search terms combined controlled vocabulary (e.g., MeSH) and free-text words related to artificial intelligence (“artificial intelligence”, “machine learning”, “deep learning”, “neural networks”, “computer vision”), pain assessment (“pain assessment”, “pain detection”, “pain evaluation”, “pain monitoring”, “pain recognition”, “pain quantification”, “pain scoring”), and facial expressions (“facial expression recognition”, “facial expressions”, “face recognition”, “emotion recognition”, “facial coding”, “nonverbal communication”, “visual perception”).

Filters were applied to restrict results to:

• Years of publication: 2015–2025;

• Language: English;

• Study design: primary research articles (excluding reviews, case reports, letters, and editorials);

• Population: adult humans (excluding neonatal and adolescent populations).

Selection process and data collection

After removing duplicates, titles, and abstracts of all retrieved searches were independently screened by two reviewers (V Cerrone, MC) to assess if they met the inclusion criteria. Full texts of the included studies were then retrieved and reviewed for final selection. Disagreements were resolved through consultation with a third reviewer (DE). For each included study, two reviewers (V Cascella, MRM) independently extracted data using a standardized form developed for this review. Extracted information included study characteristics (authors, year, setting, population), methodological details, AI approaches applied, datasets used, and main outcomes. Discrepancies in data extraction were resolved through consensus.

Data items

For each included study, we extracted information on study characteristics (authors, year of publication, country, and setting), study population (sample size, age range, clinical condition), methodological design, AI approach (machine learning or deep learning algorithm used), type of dataset employed (public or clinical), and main outcomes. Outcomes of interest included accuracy, sensitivity, specificity, F1-score, area under the receiver operating characteristic curve (AUROC), and other reported performance metrics. Additional variables collected were funding sources, validation strategy (e.g., cross-validation, external validation), and whether the study involved real-world clinical implementation.

Study risk of bias assessment

Two reviewers (MRM, V Cascella) independently rated the methodological quality of all included studies using the Revised Cochrane risk-of-bias tool for randomized trials (RoB 2) [12]. As recommended by the guidelines, the risk of bias was described and assessed for all primary outcomes evaluated in each study. Any disagreements about the methodological quality were resolved through consensus.

Effect measures

The primary effect measures considered were accuracy, sensitivity, specificity, F1-score, AUROC, and other key AI metrics, as reported in the original studies.

Synthesis methods

Given the methodological heterogeneity across studies (e.g., different AI algorithms, datasets, and outcomes), a narrative synthesis was conducted. Studies were grouped by type of dataset (clinical vs. benchmark/public datasets) and by AI method applied (machine learning vs. deep learning). Results were tabulated to allow for structured comparison of study characteristics and main findings. No quantitative meta-analysis was conducted due to variability in outcome measures and study designs.

Reporting bias assessment

Asymmetry or reporting bias was not formally assessed through statistical tests, since no meta-analysis was performed. However, we noted whether studies selectively reported only favorable performance metrics or failed to provide confidence intervals.

Certainty assessment

Given the methodological nature of the review and the lack of homogeneity across outcomes, the certainty of the body of evidence was not graded using GRADE. Instead, emphasis was placed on highlighting recurring strengths and limitations of the included studies, as well as identifying areas of consistency versus heterogeneity.

Study selection

The literature search generated 79 references (Scopus = 16, PubMed = 31, Web of Science = 4, Cochrane Library = 2, IEEE = 26), of which 25 studies met our inclusion criteria. Numbers and reasons for exclusion at each stage are reported in Figure 1. The full search strategy for each database is reported in Table S1.

Display full size

Figure 1. PRISMA flow diagram. Adapted from [10]. © 2019 The Authors. Licensed under a Creative Commons Attribution (CC BY 4.0).

During the full-text review, four studies were excluded despite initially appearing to meet the inclusion criteria. The study by Chan et al. [13] was excluded because it was only an abstract (conference proceeding) and did not provide sufficient methodological details for inclusion. Other manuscripts were excluded [4, 14–16]. Specifically, they failed to focus on facial expression-based pain assessment [4, 14, 16], or we found a lack of AI strategies [15].

Study characteristics

In this systematic review, a total of 25 studies published between 2015 and 2025 were included [17–41]. Most of this scientific output concerned experimental or feasibility studies. The number of participants varied widely, ranging from small experimental samples of fewer than 30 healthy volunteers exposed to controlled pain stimuli (e.g., cold pressor or heat pain tests) to large clinical datasets exceeding 500 postoperative patients.

The population covered different contexts, such as healthy adults in laboratory conditions, surgical and perioperative patients, intensive care unit (ICU) patients, older adults with dementia, and oncology patients. While the common focus of these studies was the use of AI applied to facial expressions for the automatic detection or quantification of pain, some studies also integrated multimodal data (e.g., speech, audio, physiological signals), though facial analysis remained the primary input modality.

About the AI methods, researchers employed different AI modalities. They ranged from classical machine learning algorithms (e.g., random forests and support vector machines) to deep learning approaches. Artificial neural network models included convolutional neural networks (CNNs), often combined with recurrent neural networks (RNNs) or long short-term memory (LSTM) units [18, 27, 29, 35, 37]. More recent contributions introduced vision transformers (ViTs). These architectures leverage self-attention mechanisms to process images as sequences of patches, thereby improving robustness in different challenging conditions such as ICU scenarios [17, 31]. In parallel, attention-based deep learning architectures, such as enhanced residual attention-based subject-specific network (ErAS-Net), were specifically designed to enhance feature selection and improve classification performance across datasets [33]. A distinct subgroup of studies focused on multimodal approaches, integrating facial data with additional modalities such as speech, audio, or physiological signals [20, 22, 28, 38]. These systems consistently outperformed unimodal facial analysis, particularly in challenging populations such as ICU patients with partial facial occlusion [40] or older adults with dementia [36], although at the expense of increased computational and methodological complexity. Other innovative approaches included the application of transfer entropy to landmark time-series analysis [25], surface electromyographic (sEMG) signals to capture subtle muscle activation [32], and binary classifiers trained on facial AUs derived from the FACS [21, 26].

The reported outcomes were generally expressed as accuracy, area under the curve (AUC), sensitivity, specificity, and F1-score, with many studies achieving high performance on benchmark datasets, such as Delaware [42], University of Northern British Columbia Pain Expression dataset (UNBC)-McMaster dataset [43], or on locally collected clinical datasets.

The summary of the individual study characteristics, including setting and number of patients included, population, approach, and inclusion criteria, AI method, and reported outcomes, is presented in Table 1.

Risk of bias in studies

Since most included studies were observational, experimental, or methodological (not RCTs), the RoB 2 tool was adapted to the specific study designs. Specifically, the category “Low risk” was assigned when the methods, datasets, and validation were clearly reported; “Some concerns” was used when limitations such as small samples, lack of external validation, or simulation-only data were present (Table 2).

Overall, we assigned low risk of bias to studies with robust datasets, transparent methodology, external validation, or those conducted within the framework of a registered clinical trial. Studies judged as having some concerns were typically characterized by small sample sizes, absence of external validation, or approaches tested only in laboratory-controlled settings, which can limit their generalizability. None of the studies included clearly fell into the category of high risk of bias.

Results of individual studies

The included studies reported heterogeneous outcomes reflecting the performance of AI applied to facial expression recognition for pain assessment. Most studies evaluated models on either publicly available datasets [e.g., UNBC-McMaster Shoulder Pain, BioVid Heat Pain, Multimodal International Pain dataset (MIntPAIN)] or on original clinical cohorts (perioperative, postoperative, oncology, or ICU patients). Outcomes were mainly expressed as accuracy, AUROC, sensitivity, specificity, recall, and F1-scores. Confidence intervals were rarely reported.

Experimental and dataset-based studies demonstrated very high performance. Bargshady et al. [18] proposed an ensemble deep learning framework integrating CNN and RNN, reaching > 89% accuracy and AUC 0.93 on MIntPAIN and UNBC. Chen et al. [26] introduced a weakly supervised approach based on combinations of AUs, achieving 87% accuracy with AUC 0.94 on UNBC. Morsali and Ghaffari [33] developed ErAS-Net, an attention-based deep learning model, achieving 98.7% accuracy for binary and 94.2% for four-class pain classification, with cross-dataset validation on BioVid still showing robust results (78.1%). Rodriguez et al. [37] combined CNNs and LSTMs for temporal analysis, outperforming state-of-the-art methods on UNBC. Pikulkaew et al. [35] applied deep learning to 2D motion and expressions, achieving > 95% accuracy across three pain classes.

Transformer-based methods showed promise. Bargshady et al. [17] employed video ViTs (ViViTs) on AI4PAIN and BioVid datasets, with accuracies of 66.9% and 79.9%, outperforming CNN baselines. Mao et al. [31] refined Conv-Transformer architectures with multi-task learning, improving estimation of continuous pain intensities on UNBC.

Clinical studies confirmed feasibility in real-world settings. Zhang et al. [41] applied VGG16 to > 3,000 images from 503 postoperative patients, reporting AUROC 0.898 for severe pain detection and consistent F1-scores. Heintz et al. [30] validated CNN-based nociception recognition perioperatively, reporting robust AUROC and calibration (Brier score) across multicenter datasets. Park et al. [34] compared models in 155 gastrectomy patients, showing that facial-expression machine learning achieved an AUROC 0.93 for severe postoperative pain, outperforming analgesia nociception index (ANI) and vital signs. Bellal et al. [19] tested the NEVVA^© device in ICU patients, showing the feasibility of automated detection in non-communicative patients.

Other methodological innovations included electromyography of facial muscles [32], showing modest predictive capacity (c-index 0.64), and transfer entropy applied to facial landmarks [25], which demonstrated robustness to uncertainty. Earlier, Casti et al. [24] benchmarked multi-expert calibration, confirming reproducibility of automated systems. Moreover, Dutta and M [27] provided proof-of-concept evidence for real-time video analysis.

Special populations were investigated. Rezaei et al. [36] validated a deep learning system in older adults with dementia, showing reliable detection where self-report is not possible. Yuan et al. [40] developed an AU-guided CNN for ICU patients with occluded faces (ventilation), reporting strong performance across binary, multiclass, and regression tasks. On the other hand, oncology-related contributions are limited. In 2023, Cascella et al. [23] developed a binary classifier on cancer patients’ videos, achieving ~94% accuracy and an AUROC of 0.98. Later, the same authors integrated facial and vocal features in oncologic pain monitoring, confirming feasibility in clinical use [21] and subsequently also tested the You Only Look Once version 8 (YOLOv8) architecture for real-time pain detection, reporting strong detection performance [22].

Several multimodal approaches combined facial expression with additional modalities. Benavent-Lledo et al. [20] used computer vision and multimodal signals, achieving > 96% accuracy and > 94% F1-score on UNBC and BioVid. Ghosh et al. [28] proposed an IoT-enabled system integrating facial and audio data, reaching near-perfect accuracy (> 99%) in binary and three-class classification. Semwal and Londhe [38] designed a spatio-temporal behavioral system with multimodal integration, confirming the added value of combining sources.

In summary, across the 25 included studies [17–41], AI-based facial expression recognition for pain consistently demonstrated high performance in controlled datasets and increasing feasibility in clinical populations (postoperative, perioperative, oncology, ICU, dementia). The overall result ranged from modest (with an accuracy c-index of 0.64) [32] to excellent (with an accuracy > 98%) [33]. Figure 2 illustrates performance (i.e., accuracy) ranges across different AI methods.

Display full size

Figure 2. AI methods used for facial expression-based pain assessment (2015–2025) and their reported performance ranges (i.e., accuracies). Traditional machine learning (ML) approaches showed moderate performance (65–85%). Convolutional neural networks (CNNs) and RNN/hybrid (recurrent neural network) models achieved higher accuracy (70–95% and 75–93%, respectively). Transformer-based models reached accuracies ranging from 66% to 98%. Multimodal approaches (facial + audio and/or physiological signals) consistently outperformed unimodal systems, achieving accuracies up to 99.5%.