A comprehensive literature search was performed in August 2024 using the PubMed database. The search was conducted without date restrictions to capture the full scope of research in this emerging field. This broad approach ensured that all relevant studies, regardless of publication date, were included, providing a more thorough evaluation of AI applications in patch testing and the observance of any trends over time. Literature searches were conducted using combinations of keywords, such as “artificial intelligence,” “machine learning,” “patch testing,” and “contact dermatitis” or “skin” (see Multimedia Appendix 1 for the full search term list). These terms were chosen to ensure a wide net was cast, incorporating both general AI terms and specific patch testing and dermatology-related concepts.

Criteria for inclusion and exclusion were defined prior to screening to reduce potential biases. Studies were included if they met the following criteria: the population consisted of human patient populations undergoing patch testing; the study design involved AI (which includes ML and DL); and outcomes reported on the performance of these algorithms. All publication types, including journal articles, conference abstracts, and preprints, were considered. Studies were excluded if they were not written in English or if they were not original research, such as review papers or perspectives.

The PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) approach was selected to ensure transparency and replicability in the selection process, providing a clear pathway from initial search to final inclusion [17,18]. Each article was independently reviewed by 2 authors. In cases of disagreement, a third author resolved the discrepency. From the included studies, the following data elements were then extracted: study design; sample size; skin types included; length of study for each participant; location of study; materials used (such as types of allergen panels and imaging equipment); type of AI algorithm and its performance in the study; limitations and challenges of the study; and future directions. To ensure that AI models were properly evaluated, each study relied on a clearly defined ground truth as the reference standard for their data. This ground truth was established by dermatologists’ manual interpretation of patch test reactions, typically following standardized grading criteria such as the International Contact Dermatitis Research Group (ICDRG) scale, with some studies following European Society of Contact Dermatitis guidelines or similar clinical severity scales [19]. The findings were then synthesized to highlight trends, gaps, and potential areas for future research in the application of AI in patch testing. This synthesis serves as a foundation for guiding future research efforts, with the goal of synthesizing both technical and clinical factors of the clinical patch testing procedure, analysis, data capturing, image capturing and storage, AI algorithms, and diagnostic accuracy comprehensively, contributing to the current gaps in the current practice of AI integration within dermatological patch testing diagnostics.

A total of 94 records were ultimately screened and evaluated for eligibility, as shown in the PRISMA flow diagram (Figure 1) [17]. Of the 94 articles, our literature review identified 10 relevant studies that employed various AI techniques in the context of patch testing and skin sensitization prediction, as shown in Table 1. These studies encompassed a wide range of approaches, from image analysis of patch test results to molecular profiling and risk prediction models.

Geographically, the studies were conducted across various countries and continents, with the United States (3 studies) and Greece (3 studies) being the most represented. The remaining studies were distributed across other European countries, including the United Kingdom, Finland, and a multi-country study spanning Germany, Switzerland, and Austria. Sample sizes also varied considerably between the studies, ranging from 47 patients in the molecular signature study by Lefevre et al [25] to 105,325 patients in the large-scale analysis by Adler et al [29]. In total, 9 studies had dermatologists as authors, with some contributions including patient recruitment, clinical assessment, or patch test evaluation [20-28]. The materials used for patch testing varied, with many using standard European baseline series allergens. However, some studies, such as Lefevre et al [25] and Fortino et al [28], used specific sets of allergens and irritants for their molecular profiling approaches. Most studies classified reactions on a scale ranging from negative or irritant to +++ for strongly positive reactions, though the specific scoring systems and timepoints for evaluation varied between studies.

Of the 10 studies reviewed, 4 analyzed images, 4 analyzed clinical and demographic data, and 2 analyzed biological mechanisms of biopsies for patch testing. In total, 4 studies analyzed photographic images of patch test sites, which were captured using a range of imaging modalities [21,22,24,26]. Of these 4 studies, 3 used digital camera or smartphone camera images, while Vezakis et al [24] used an advanced multi-modal imaging device, the Antera 3D® camera, which captured 6 image modalities—color, redness, texture, fine lines, and volumes (see Multimedia Appendix 2 for expanded information on the 4 image datasets). The detailed information on skin topography and chromophore concentration, captured by the Antera 3D® camera independent of lighting, provides a standardization that improves accuracy and the need for additional standard dermoscopic image preprocessing techniques [24]. In total, 4 studies analyzed clinical and patient demographic data as predictive features for ML models, which included anatomical sites, age, gender, and sex [20,23,27,29]. Additionally, 3 studies included additional clinical parameters, such as occupation and atopic history [20,23,27]. Other clinical data included the patch test ICDRG evaluations, MOAHLFA (Male-Occupational-Atopic-Hand-Leg-Face-Age) Index, and skin characteristics [20,23]. Lastly, 2 studies analyzed genomic and molecular profiling of patch test biopsies [25,28]. Seven studies reported age groups in their studies with median or mean patient age ranges from 39 to 61 years [20-23,25,27,29]. Regarding skin types, only 3 out of the 10 studies reported on the distribution of skin tones in their datasets [21,22,26]. Chan et al [26] included Fitzpatrick skin types (FST) I-V, with the majority falling into FST II-III. Hall et al [22] reported that 82% of their patient population was White, with lighter skin tones typically ranging from FST I-III. Ravishankar et al [21] showed a significant imbalance, with 88% of images representing lighter skin tones from FST I-II.

Overall, convolutional neural networks (CNN) were the most commonly used algorithms for image analysis of patch test reactions, employed in 4 of the 10 studies [21,22,24,26]. These CNN-based models demonstrated high accuracy in identifying and classifying patch test reactions. Hall et al [22] reported an accuracy of 90.9% with an area under the curve (AUC) of 0.885, while Chan et al [26] achieved an even higher accuracy of 99.5% with an AUC of 0.915. Similarly, Ravishankar et al [21] and Vezakis et al [24] reported accuracies of 90.1% and 90%, respectively, further supporting the potential of CNN use in patch test interpretation. Other approaches, such as random forest (RF), gradient boosting (GB), and logistic regression (LR), were employed in studies focusing on risk prediction and biomarker discovery [25,27-29]. Notably, Cunningham et al [27] compared multiple algorithms and found that GB outperformed other predictive methods, including LR, RF, and AdaBoost, with AUCs of 0.69, 0.65, 0.60, and 0.58, respectively, for predicting cutaneous allergy risk. In total, 2 studies used multiple correspondence analysis to investigate patterns and relationships in patch test data, particularly in the context of occupational dermatitis and population-specific sensitization profiles [20,23]. While these studies did not provide specific accuracy metrics, they demonstrated the utility of AI techniques in uncovering complex associations within patch test data.

This review of 10 studies exploring the application of AI techniques in patch testing reveals promising advancements along with numerous challenges and limitations. The diverse range of approaches, from image analysis to molecular profiling and risk prediction, demonstrates the versatility of AI in addressing various aspects of contact dermatitis diagnosis and patch testing in general.

The high accuracy achieved by CNN-based models in analyzing patch test images is particularly significant. With accuracies ranging from 90.1% to 99.5%, these models show great potential for automating and standardizing patch test interpretation, as some studies have shown interrater variability in diagnosing patch test reactions [4]. This could lead to more consistent diagnoses across different clinical settings, reduce dermatologists’ workload, and help expand access to patch testing. One key barrier is the need for standardized imaging protocols [8]. The variability in the quality of images, as well as the inconsistency in how and when these images are captured, introduces a significant source of error in AI models. Standardized, high-quality image capture and storage protocols are essential for ensuring that AI systems can be effectively trained and applied across different clinical settings [22].

Moreover, our review underscores the necessity for large, diverse, and representative image databases to train AI models [22,24]. Specific areas researchers should focus on include the inclusion of patients across all Fitzpatrick skin types to address potential performance gaps in darker skin tones, which are often underrepresented in dermatologic datasets [30]. The development of datasets such as the Diverse Dermatology Images (DDI) dataset underscores this need. The DDI dataset includes representation across FST I-VI for biopsy-proven correlates of benign and malignant lesions, common dermatological conditions, and ambiguous lesions [30]. Their DDI research highlighted worsened performance in the ability of certain state-of-the-art dermatology AI algorithms to accurately diagnose skin conditions in darker skin tones of FST V-VI. Their research also found that fine-tuning on diverse image sets such as DDI could overcome the gap in performance of the AI algorithms when comparing FST I-II and V-VI. Ensuring geographic and demographic diversity by collaborating with institutions in varied regions globally can help capture heterogeneity in environmental exposures, allergen profiles, and clinical practices. Initiatives such as federated learning offer a multi-institutional collaborative effort to train AI algorithms while preserving institutional data privacy through a consensus model [31]. Federated learning-trained AI models have been shown to outperform models trained on single-institutional data [32]. Additionally, datasets should aim to balance age, gender, and clinical variations in skin reactions, such as weak versus strong positive reactions, to improve model robustness [33,34]. Synthetic data offer an increasingly used solution to build larger, more robust training datasets; however, special attention is needed to ensure the inclusion of diverse synthetic input to mitigate sample selection bias [35,36]. The creation of a global patch testing image repository would not only improve AI model performance but could potentially also accelerate the discovery of new dermatological insights, enable the continuous refinement of diagnostic algorithms, and increase diagnosis assistance for complex cases, especially in lower resource settings.

The application of other techniques such as RF, GB, and LR in risk prediction and biomarker discovery is also promising. The study by Cunningham et al [27], which found GB to outperform other methods in predicting patch testing results, suggests that more complex, nonlinear approaches may be necessary to capture the intricacies of skin sensitization mechanisms. This highlights the potential of ML in discerning subtle patterns that may not be apparent through traditional statistical analyses. A more widespread and diverse dataset would not only enhance the performance of AI but also address concerns around bias, ensuring that AI-driven diagnostic tools are equitable and effective for all patients, regardless of demographic factors [30].

Despite the promising results, several limitations were identified across the reviewed studies. First, most studies had relatively small sample sizes, with 8 out of 10 studies including fewer than 250 participants, and only 2 studies including more than 1000 patients [27,29]. This limits the generalizability of findings and may lead to overfitting in ML models, as many of the studies noted [24,27]. Second, there was a lack of diversity in skin types reported across studies, with 7 of our studies not specifying the range of Fitzpatrick skin types included. This is particularly important given that skin reactions can present differently across various skin types, potentially affecting the performance of image-based AI models [37,38]. Additionally, the lack of standardization in methodology across studies makes direct comparisons challenging. Some studies used standard European baseline series allergens, while others used specific sets of allergens, making it difficult to assess and compare the robustness of the models across different allergen panels.

The ethical implications of using AI in clinical practice and industry engagement in this space also warrant attention. As AI tools become more integrated into dermatology, it is crucial to maintain transparency and interpretability in AI models [39]. For successful implementation in patch-testing diagnostics, the AI system should provide clinicians with transparent, mechanism-based explanations of its predictions, including which clinical features or biomarkers are driving its decision-making process and model confidence [40]. Efforts to increase AI literacy among health care professionals, as well as to develop user-friendly AI interfaces, will be essential in fostering the integration of these technologies into routine clinical workflows.

Furthermore, the regulatory landscape for AI in dermatology, and health care more broadly, is still evolving. While AI tools show promise, rigorous validation and regulatory approval are needed before they can be fully integrated into clinical practice [41]. Dermatologists, health care institutions, and national and international policymakers must collaborate to develop clear guidelines for the safe and effective use of AI in patch testing and other dermatological applications.

Overall, AI holds immense potential to revolutionize the diagnosis of contact dermatitis through more accurate and standardized patch testing methods. However, to realize this potential, further research is needed to address the challenges of standardization, data diversity, model transparency, and regulatory oversight. With concerted efforts, AI can serve as a powerful tool in dermatology, enhancing diagnostic capabilities, improving patient outcomes, advancing precision dermatology, and ultimately contributing to more equitable health care delivery [42].

This narrative review underscores the significant potential of AI to revolutionize patch testing by enhancing diagnostic accuracy, reducing inter-provider variability, and providing a more standardized, scalable system for interpreting digital patch test results. The high accuracies achieved by CNN models in patch test image analysis are particularly noteworthy, suggesting a possible path towards more standardized and objective patch test interpretation internationally. Our analysis also highlights a need for the development and adoption of standardized protocols for capturing patch test images. Establishing these protocols is crucial for facilitating accurate diagnostics across diverse patient populations, supporting quality improvement efforts, and promoting AI-driven advancements and analyses. The creation of expansive patch testing databases and standardized protocols will enable increased application of AI systems to deliver more accurate, equitable, and scalable care in the management of ACD.