1. Introduction
Artificial intelligence (AI), particularly methods of machine learning (ML) and data-driven modeling, is becoming increasingly important in biomedical research and clinical care. In allergology, a medical discipline characterized by pronounced clinical heterogeneity, complex immunological networks, and a wide range of relevant influencing factors, AI offers the potential to substantially extend existing concepts of diagnostics, phenotyping, patient information, patient consent, and therapeutic decision-making. International professional societies and study groups have emphasized the need to integrate innovative digital technologies into care of atopic diseases in a structured, evidence-based, and responsible manner [1, 2, 3].
A growing body of evidence already suggests the feasibility of selected AI applications in allergology. ML-based approaches have been applied to the phenotyping and endotyping of asthma, allergic rhinitis, and atopic dermatitis, frequently using clustering algorithms and multivariate modelling strategies [4, 5, 6, 7]. Additional fields of application include AI-supported analysis of molecular allergy diagnostics, including diagnosis and allergen immunotherapy selection for seasonal allergic rhinitis, component-resolved diagnostics, with the aim of improving the clinical interpretation of complex sensitization patterns [5, 8, 9] and estimating the risk of severe systemic reactions, in Hymenoptera venom allergy [8] and β-lactam allergy [10, 11, 12] as well as the likelihood of reaction in oral food challenges [13, 14, 15].
Furthermore, predictive models for treatment response to allergen-specific immunotherapy [16] as well as for the selection of targeted biological therapies [17] have been described.
The potential added value of AI in allergological care can be seen particularly in the integration of heterogeneous data sources, improved diagnostic precision, and support for individualized therapeutic decision-making [4, 5, 6, 18]. AI-based systems can help to capture interindividual variability more effectively and may contribute to the standardization of complex clinical assessments [4, 5, 6, 17]. Importantly, AI is not intended to replace medical expertise but rather to augment clinical reasoning and support evidence-based decision-making, as emphasized in recent work on clinical AI [19, 20].
Despite these promising developments, relevant limitations remain. These include limited generalizability of many models, potential bias resulting from non-representative datasets, a lack of data transparency, explainability (“black-box” models), and unresolved ethical, legal, and regulatory issues. Standardized validation, robust clinical evidence, and clear governance structures are therefore essential for responsible implementation [4, 8, 21].
Furthermore, there is a risk that physicians who are inexperienced in allergology will invest less time and care in their own training and continuing education in the field of allergology, as they can obtain medical information for self-education more quickly and easily via freely accessible AI. This information may be incorrect or at least incomplete, leading to potentially dangerous misjudgments, especially when patients use AI to make self-diagnoses and treatment decisions [22].
The aim of this position paper is to assess the current state of AI in allergology, define clinically relevant fields of application, critically appraise opportunities and limitations, and formulate consensus-based recommendations for the evidence-based, safe, and ethically responsible implementation of AI in research and clinical practice.
General areas of application
Beyond medicine, AI technologies are widely used in fields such as image and speech recognition, natural language processing, autonomous systems, financial risk modelling, and environmental analysis. In these domains, ML and deep learning (DL) approaches have demonstrated the ability to extract relevant patterns from high-dimensional data, automate complex analytical processes, and achieve predictive performance exceeding that of traditional methods of data analysis [23].
In medicine, applications range from image- and signal-based diagnostics to risk prediction, clinical decision support, workflow optimization, and personalized treatment planning. AI-driven models have shown promising performance in integrating multimodal data sources, such as clinical records, imaging, laboratory results, and molecular profiles, with the aim of improving diagnostic accuracy and patient stratification. Rather than replacing clinicians, contemporary medical AI is increasingly framed as technology augmenting healthcare, designed to support clinical reasoning, reduce variability in complex assessments, and assist in individualized therapeutic decision-making.
Narrative and systematic reviews in medical AI consistently underline that clinical benefit depends not only on algorithmic performance but also on careful validation, interpretability, and seamless integration into real-world clinical workflows [24].
Conclusion 1
Rather than replacing clinicians, medical AI is increasingly conceptualized as an augmentative tool that supports clinical reasoning, reduces variability in complex assessments, and facilitates individualized treatment decisions. Evidence from narrative and systematic reviews indicates that clinical benefit depends not only on algorithmic performance, but also on robust validation, interpretability, and effective integration into clinical workflows.
2. Materials and methods
This Position Paper is based on a non-systematic narrative review of the literature combined with structured expert input. A targeted literature search was conducted from November 2025 to January 2026 in PubMed to identify relevant publications on the application of AI in allergology. After inclusion of additional relevant literature during drafting the manuscript, the review of the literature covered the period from November 2025 to March 2026, providing expert consensus on the content of each chapter of this Position Paper.
All chapters were then reviewed and discussed by all co-authors; in cases of dissent, iterative discussion and feedback was applied to achieve consented texts, with agreement of at least 70% of authors necessary for each debated statement. Used combinations of keywords included artificial intelligence, machine learning, allergy, allergic disease, atopic disease, clinical decision support, digital health, and patient engagement.
This review focuses on peer-reviewed articles addressing clinically relevant applications of AI in allergology, such as diagnostic support, risk prediction and phenotyping, treatment optimization, automation of clinical or laboratory procedures, patient-facing digital tools, and ethical, legal, and societal considerations. In addition to original research articles, gray literature, legal texts, narrative reviews, commentaries, and perspective papers were included when they provided conceptual insights, highlighted unmet needs, or addressed patient-centered and implementation-related aspects. Further relevant publications were identified through manual screening of reference lists of key articles.
Given the rapidly evolving nature of AI technologies and their translation into clinical practice, this review was designed as a focused, narrative synthesis rather than a systematic review. The objective was to identify, appraise, and integrate representative, influential, and clinically relevant evidence addressing key questions related to the application of AI in allergology patients. Evidence selection was guided by relevance to clinical decision-making, methodological rigor, and impact on practice.
This approach was chosen to inform practical guidance, while explicitly acknowledging the strengths and limitations inherent to a non-systematic evidence synthesis.
The development of this Position Paper was informed by the AGREE II reporting framework to enhance methodological transparency and reporting quality. Although the present manuscript does not represent a formal clinical practice guideline with graded recommendations, several core domains of the AGREE framework were considered during manuscript preparation. These include definition of scope and objectives, transparent description of the development group, expert consensus procedures, and disclosure of competing interests.
The AGREE principles were applied pragmatically to improve clarity, transparency, and reproducibility of the consensus process [25].
Conclusion 2
This Position Paper has been developed with reference to the AGREE II framework aimed to enhance transparency and reporting quality. Although not a formal guideline, key domains, such as scope, authorship, consensus methods, and conflict of interest disclosure, were addressed and AGREE principles were applied pragmatically to improve clarity and reproducibility.
3. AI taxonomy
3.1. Distinction between AI and methods of classical statistics
Classical statistical methods and AI methods may appear as completely separate entities. In practice, both aim to leverage data to solve particular tasks. Clinicians trained in epidemiology and biostatistics might be more familiar with classical statistical methods.
We aim to provide a brief introduction to AI, highlighting overlaps with classical statistics, while also drawing clear distinctions between AI methods and classical statistical and data analysis methods.
Together, these approaches are complementary: classical statistics offers interpretability and potential causal insights, while AI, in particular ML methods, provide powerful tools for prediction and pattern recognition of large data sets and complex data.
Conclusion 3
Classical statistical and AI methods may seem distinct, but both aim to leverage data to address specific tasks. Therefore, it is fundamental to understand both distinctions and overlap between both analytical approaches.
3.2. The different forms of AI
Machine learning
As opposed to most of the classical statistical methods mentioned before, ML encompasses data-driven methods that focus on optimizing predictive performance or on discovering hidden structures in the data rather than on inference and interpretability [26]. The main difference between these approaches and traditional statistics is not purely mathematical, but rather related to the workflow and objectives. Instead of using expert knowledge during model selection and to determine the assumptions made about the data-generating mechanism, ML algorithms – such as random forests, gradient boosting, or neural networks – learn patterns directly from data by iteratively optimizing model parameters.
As a result, these methods can represent complex relationships and interactions within the data without explicit model specifications [26]. Thus, ML places greater focus on evaluating how well a model performs on new, unseen data, than on the interpretability of the model itself.
Deep learning
A branch of ML known as DL uses neural networks with multiple layers that automatically learn hierarchical representations of data. These methods can be applied to most tasks where ML methods are suitable, and are particularly powerful dealing with high-dimensional unstructured data [27], such as medical images, ECG waveforms, and free-text clinical notes. DL models often operate as black boxes, making interpretability more challenging given their complexity [28].
Vision models
Vision models are AI systems designed for the analysis and interpretation of visual data, such as images or video frames. In medical applications, including allergy and immunology, these models are predominantly based on DL architectures, most notably convolutional neural networks (CNNs) and, more recently, vision transformers (ViTs). Their primary objective is to extract reproducible and quantitatively interpretable features from pixel-level data.
Convolutional neural networks
CNNs represent the most established vision architecture in medicine. They learn hierarchical image features, ranging from low-level patterns (edges, contrasts) to complex disease-associated structures. In allergology, CNN-based approaches have demonstrated high performance in image-driven tasks, particularly in dermatology.
CNN-based systems have been developed to support the assessment of urticaria activity, conjunctival allergen provocation tests, and skin prick test reactions, with evidence suggesting reduced interobserver variability compared with manual evaluation [29, 30, 31, 32].
Vision transformers
ViTs apply attention-based mechanisms, which enable context-dependent weighting of relevant information by the model itself to image patches and allow modeling of global spatial relationships [33]. While their use in allergology is still largely exploratory, they are increasingly relevant for multimodal systems that integrate visual data with clinical text. Their broader adoption is currently constrained by substantial data requirements and computational complexity [29].
Beyond clinical imaging, vision models are also applied in environmental allergy research. DL-based pollen classification from microscopy images, through vision models, has demonstrated high accuracy but remains challenged by particle overlap, limited depth information, and morphological similarity between taxa, emphasizing the importance of careful dataset curation and external validation [34, 35].
Large language models and transformers
LLMs and the related reasoning models are recent developments in DL. These have gained popularity because of their use in chatbots, such as ChatGPT, Claude, Google Gemini and the European LLM Mistral. These models are trained on incrementally large collections of text in order to recognize probabilistic associations between words and generate text, thus collectively being considered types of generative AI.
Their defining feature is the use of self-attention mechanisms, which allow contextual dependencies across long text sequences to be modeled efficiently. In medicine, LLMs enable analysis of unstructured data such as clinical documentation, guidelines, and patient-generated queries.
In allergology, LLMs have been evaluated for knowledge retrieval, clinical support, and patient education. Several systematic assessments indicate that responses to allergology-related questions are often coherent and largely factually correct. However, consistent limitations in reliability and transparency have been reported.
An EAACI-affiliated evaluation of ChatGPT-4 responses on allergen immunotherapy rated overall information quality as fair, while reliability scores were insufficient due to absent authorship, references, and disclosure [36]. Importantly, LLMs generate text based on probabilistic likelihood rather than explicit fact verification.
A systematic analysis of 120 interdisciplinary allergology questions identified not only minor inaccuracies but also a small number of clinically relevant errors with potential safety implications, particularly in pediatric and food allergy contexts [22]. These findings underscore that current LLMs cannot – or should not – be considered autonomous clinical decision tools, although they could be useful in a blended human–AI approach (human-in-the-loop and human-on-the-loop) [37].
Ongoing research therefore focuses on domain-specific adaptation strategies, including finetuning and retrieval-augmented generation, which explicitly ground model outputs in curated medical sources such as guidelines and peer-reviewed literature [38]. Within an AI typology framework, LLMs should be regarded as powerful language-processing systems with supportive roles in allergy care, requiring continuous expert oversight.
Conclusion 4
ML algorithms, such as random forests, gradient boosting, and neural networks, derive patterns from data through iterative parameter optimization. Consequently, emphasis is primarily placed on generalization performance to unseen data rather than on model interpretability.
4. Applications for AI in allergology
4.1. Environmental analytics
4.1.1. Pollen analysis
Pollen and allergies
AI has the potential to facilitate the integration of exposomics into allergy care. Exploring possibilities to do so via pollen monitoring seems promising as a use case detailing the high granularity of data in exposomics. As Europe is characterized by a highly heterogeneous distribution of allergenic plant species, resulting in substantial regional differences in pollen exposure [39, 40], the allergenic burden experienced by sensitized individuals strongly depends on geographic location.
In Northern and Central Europe, the predominant aeroallergens originate from pollen of the Betulaceae family, particularly birch (Betula), hazel (Corylus), and alder (Alnus) [41, 42]. In contrast, in Mediterranean regions, pollen from the Oleaceae family, including olive (Olea), ash (Fraxinus), and privet (Ligustrum), represents a major source of allergenic exposure [43, 44]. Across Europe, however, among other important allergens such as weed pollen, including mugwort (Artemisia) and ragweed (Ambrosia), pollen of the Poaceae (grass) family remains the most common aeroallergen and a leading cause of seasonal allergic rhinitis [45, 46, 47].
Pollen monitoring in the world
Individuals sensitized to pollen rely on information about local pollen concentrations to better manage symptoms and adapt daily activities to environmental exposure. To monitor airborne particles such as pollen, and frequently also allergenic fungal spores including Cladosporium and Alternaria spp., aerobiological monitoring stations have been established worldwide (Figure 1a).
Traditionally, pollen monitoring has relied on manual sampling devices, particularly Hirst-type traps, which require microscopic identification by trained experts and typically provide results with delays of several days. These approaches are labor-intensive and subject to methodological limitations [48, 49, 50, 51].
In recent years, automated monitoring systems integrating AI-based image recognition have emerged, and it is now possible to avoid some of the mistakes of the manual traps [52, 53, 54]. With automatic pollen monitoring, 1- to 3-hour values spread over the day are reported online with a 1- to 5-hour delay (the time depends on the instrument). Also, these rapid online data open new possibilities for next-day pollen predictions [55].
However, symptom severity correlates even more strongly with the concentration of allergens released from pollen grains than with pollen counts themselves [56], as demonstrated for Phl p 5 (grass pollen) in the UK [57]. Still, pollen counts are a good substitute parameter to explain allergic symptoms of allergic individuals as pollen are the only, although variable, source of these allergens. Because the direct measurement of airborne allergen concentrations is technically demanding and time-consuming, pollen counting remains a rapid, inexpensive, and practical alternative.
AI in pollen recognition
Currently, several automated pollen monitoring systems using AI are capable of sampling, analyzing, and identifying airborne pollen and spores in near real time or real time. A comparative performance study conducted in Central Europe (Munich, Germany) evaluated multiple instruments by comparing their measurements with the mean values obtained from 4 co-located Hirst-type traps. Averaging the 4 manual samplers helped reduce uncertainties inherent to the manual counting method.
From this comparison, 2 reliable automated devices emerged: the Hund BAA500 (Helmut Hund, Wetzlar, Germany), based on image recognition (Figure 1b), and the Swisens Poleno Jupiter, which combines holographic imaging with fluorescence spectral analysis (Figure 1c) [58].
For the Hund BAA500, pollen classification is performed using DL models based on CNNs [54, 59, 60, 61]. An important advantage of this system is the possibility of expert validation: all captured pollen images are stored online and can be re-examined individually by aerobiological experts, allowing human-in-the-loop verification of the automated classifications (https://validation.pollenscience.eu).
The Swisens Poleno Jupiter system identifies airborne particles using holographic imaging combined with fluorescence spectral analysis. Its classification algorithm likewise relies on DL CNN-based methods [62]. In contrast to the Hund BAA500, this system provides fully automated online reporting, enabling faster data availability but offering fewer options for manual validation of individual particle classifications.
Network with automatic devices
In Europe, there are only a few networks of automatic pollen monitors. The largest one is based on image recognition and will use about 28 Hund BAA500 devices and includes 8 automatic monitors (ePIN Network) in Bavaria, Germany, operational since 2019 [63]. The network has been expanding since 2025 under the leadership of the new German Weather Service into a Germany-wide network.
Another major network is based in Switzerland (SwissPollen network) with 15 automatic monitors using the Swisens Poleno [62].
Pollen forecast using AI
Once the airborne concentration of allergenic pollen has been measured and delivered online to the public, the next step in supporting individuals with pollen allergies is pollen forecasting. The most widely used forecasting approaches rely on numerical atmospheric dispersion models, such as SILAM [64] and ICON-ART [55]. These models enable pollen forecasts at the European scale, but currently include only a limited number of allergenic pollen taxa [65].
More recently, ML-based approaches have been developed to improve pollen forecasting accuracy. Such models have been applied to predict airborne concentrations of Ambrosia, birch, Olea, and grass pollen, as well as Quercus, Cupressaceae, Urticaceae, and Alternaria spores [66, 67, 68, 69, 70, 71, 72]. These approaches already provide relatively accurate predictions at the local scale, although scaling such models to larger geographic regions remains challenging.
Conclusion 5
Integrating real-time pollen data with AI-based meteorological and dispersion models is expected to improve forecast accuracy and exposure assessment. The distinct morphology of pollen enables reliable AI-based image recognition and automated analysis. Despite higher initial costs than Hirst-type traps, reduced manual labor may improve cost-effectiveness.
Ongoing technological advances are likely to support broader implementation, benefiting research, and clinical allergy management.
4.1.2. Hymenoptera monitoring and classification
AI, particularly DL-based image recognition systems, has substantially improved automated insect classification. Most published work to date has emerged from agricultural, ecological, or taxonomic research rather than from a clinical or allergological perspective [73, 74, 75, 76, 77, 78, 79, 80]. Nevertheless, these technological advances are directly relevant to allergology, particularly for the identification of medically important Hymenoptera species.
The dominant methodological approach involves CNNs trained on annotated image datasets [73, 74, 75, 76, 77, 78]. High species-level classification performance has been demonstrated across multiple taxa. For example, DL models achieved 91.6% accuracy for bumble bee species identification [79], while automated hornet detection systems reported precision–recall F1 scores ≥ 0.99 for invasive Vespa velutina detection [76], highlighting a favorable relation of precision and recall.
Real-time monitoring platforms integrating object detection and tracking demonstrated an average precision of 89% in field-based insect classification [77], and hybrid detection–tracking approaches achieved honeybee tracking rates of 86.6% in complex outdoor environments [73]. Embedded hive-monitoring systems based on YOLO (You Only Look Once), a fast architecture for object detection, reported F1 scores of 0.95 for honey bee detection and 96.28% tracking accuracy [75]. Beyond visual classification, complementary AI-supported approaches include morphometric and genomic methods.
Geometric morphometrics discriminated honeybee subspecies with 84.9% accuracy at the individual wing level [74], while image-based neural network classification of parasitoid wasps achieved an average genus-level accuracy of 96% [78]. At the molecular level, whole-genome resequencing enabled the delineation of a previously undescribed Apis mellifera subspecies, demonstrating the role of genomic data in high-resolution taxonomic validation [80].
Collectively, these studies demonstrate expert-level classification performance under controlled and semi-field conditions. However, their primary focus remains agricultural monitoring, biodiversity research, and invasion biology rather than clinical allergology. Future translational efforts integrating AI systems for point-of-care classification and epidemiological surveillance may enable more precise environmental risk stratification and aid diagnostics for venom-allergic individuals [73, 74, 75, 76, 77, 78, 79, 80].
Conclusion 6
AI, particularly DL-based image recognition methods such as CNNs, has significantly advanced automated insect classification. Emerging translational applications integrating point-of-care AI with epidemiological surveillance may enhance environmental risk stratification and support the diagnosis of venom allergy.
4.2. Clinical decision support systems
Clinical decision support systems (CDSS) are digital systems designed to support clinical tasks and assist decision-making processes [81] (see Chapter 2). CDSS are already established in several areas of clinical practice, particularly in image-based diagnostics such as melanoma detection in dermatology [82].
Increasingly, CDSS are being explored as potent tools in allergology [81, 83].
The translation of CDSS into routine clinical practice in allergology is currently evolving and novel applications illustrate their potential for different clinical tasks, supporting diagnosis, risk stratification, and therapeutic decision-making.
4.2.1. Medical history/data analysis
A potential field of application for CDSS lies in the partial automation of history-taking tasks.
By efficiently processing large volumes of data, CDSS may contribute to substantial time savings and facilitate the structured collection of patient information, including medical history and relevant environmental factors. Conceptually, CDSS consist of three core components: data input, a processing algorithm or underlying AI, and the reporting of results [84].
At the level of AI-based processing, however, interpretation of results may be challenging due to the opacity of complex models. Approaches within the field of explainable artificial intelligence aim to address this limitation [85, 86].
Even today, AI-enhanced history-taking tools can be used in clinical practice, for example to handle incoming patient calls or to assist in documenting and transcribing patient histories. However, despite their increasing sophistication, these technologies remain imperfect, and require human oversight [87].
4.2.2. Diagnostics/prospective information
Improvements in medical diagnostics are particularly evident when AI is combined with human clinical assessment, compared with purely human-based evaluation. In specific domains, such as melanoma detection, AI systems already achieve diagnostic performance that equals or even exceeds individual human expertise, especially when used in conjunction with physician judgment [88]. In this context, CDSS may provide information that supports earlier recognition, risk stratification, and diagnostic confidence.
CDSS can broadly support the interpretation of IgE microarrays for allergy diagnostics [83] and the etiological diagnosis of pollen allergy with AIT prescription [89]. A large-scale project for an AI-driven global atlas of IgE antibody responses to environmental antigens in humans has also been initiated [90], based on the use of IgE microarrays in patients with IgE-mediated allergies.
4.2.3. Therapeutic decision support
Therapeutic decision support represents a core clinical application area of CDSS. ML-based CDSS can identify patients who may benefit from therapy modification based on sensitization profiles, biomarkers, demographic characteristics, prior medication, or disease course. In addition, such systems may alert physicians to newly available therapeutic options or recent extensions of regulatory approvals that may not previously have been considered for an individual patient [91].
In allergology, CDSS have been developed to assist in the prescription of AIT to this end [89] as well as for day-to-day pharmacological management and short-term prevention of seasonal allergic rhinitis [92].
Further CDSS for health surveillance of pollen allergy have been proposed using a crowdsourced digital application based on a participatory and person-centered digital medicine approach utilizing volunteered geographic information, supporting personalized short-term prediction of allergic symptoms in patients with hay fever [93].
In current applications of CDSS in allergology, their supportive character remains central, as these systems are intended to complement clinical decision-making rather than replace medical professionals. However, the role of CDSS may evolve in the future of allergology, with AI increasingly discussed as a “collaborative partner” rather than merely a support tool [94].
4.2.4. Practical applications
CDSS can be used to structure and search large databases for specific information, including multimodal clinical data, assisting especially at bottlenecks in precision medicine [81], and more generally, in complex clinical tasks [95]. Potential fields of application in allergology are detailed in Table 1.
4.2.5. Outlook
Reliable use of CDSS in allergology requires models whose training reflects the linguistic and structural characteristics of medical texts in different languages as well as variations in national regulatory frameworks. At the same time, the systematic transfer of established ML approaches from other medical disciplines should be evaluated to support the development of allergology-specific, task-focused, and agentic CDSS.
In allergology, the development of guideline-based and retrieval-augmented generation-supported systems, together with companion tools endorsed by professional societies, could serve as a model for prognostic decision support in other medical fields. For broader clinical adoption, readily deployable out-of-the-box CDSS tools are required to enable rapid translation into routine allergological care.
A human-in-the-loop approach remains essential, particularly for complex and high-stakes clinical tasks. This perspective aligns with emerging approaches that deploy AI in agentic networks, in which specialized agents are optimized for specific tasks rather than relying on a single, less specific multipurpose system [29, 102].
Conclusion 7
CDSS are digital tools designed to support clinical tasks and decision-making. Their integration into routine allergological practice is evolving, with emerging applications demonstrating potential for diagnosis, risk stratification, and therapeutic decisions.
Reliable implementation requires models trained on linguistic and structural characteristics of medical texts in different languages and aligned with variations in national regulatory frameworks. A human-in-the-loop approach remains essential, particularly for complex, high-stakes tasks, consistent with emerging AI concepts based on task-specific, agentic networks rather than general-purpose systems.
4.3. Patient education and self-management support
Supportive AI tools for patients include chatbots, connective online platforms, mobile applications, wearables, and smart (sensor) technologies. In the past decade, a rising number of digital solutions has been developed and evaluated [103]. In the following, we focus exclusively on those implementing ML or other aspects of AI.
Conversational agents (chatbots) to support patient self-management
Despite the current diversity of available chatbots, ChatGPT (Open AI) gains outstanding attention in the public as well as in the scientific community. Most individuals in Germany made their first experience with generative AI in the form of an LLM, mainly ChatGPT [104, 105]. Consequently, this chatbot has been most frequently evaluated in the medical context of allergology.
A recent performance assessment of ChatGPT4 across 120 interdisciplinary questions around allergies resulted in a reasonably good accuracy of the model compared to expert opinion (mean accuracy 4.1/5, standard deviation: 0.78, range: 1 – 5). Yet, the chatbot delivered erroneous information related to a case of pediatric food allergy, which could have resulted in a life-threatening risk, suggesting a need for human (expert) supervision [22]. Studies on the performance of ChatGPT on queries regarding atopic dermatitis [106], allergic rhinitis, chronic rhinosinusitis [107], chronic urticaria [108], and oral immunotherapy [109] showed a good performance and underline the potential of LLM technology to deliver medical information independently of the availability of medical staff.
A specific performance evaluation of ChatGPT-4 regarding 24 questions on allergen immunotherapy also showed a solid correlation between the LLM and expert opinion [36], but in line with others [110], the authors underline a lack of transparency and readability, especially for users without previous knowledge on the topic or academic education. Particularly the risks of limited transparency and potentially life-threatening accuracy flaws have been addressed by Aulenbacher et al. [31], who created a chatbot for chronic spontaneous urticaria queries based on pre-selected sources from a systematic literature search. This approach led to a significantly higher average accuracy (94.6%) compared to ChatGPT-3.5 (42.6%) and ChatGPT-4o (85.7%).
An app-based intervention for patients with grass pollen allergy demonstrated that AI-based symptom forecasting features improved patient self-management and reduced symptom burden [92].
In summary, conversational agents have the potential to play a significant role in information delivery, patient education, and guidance [111, 112]. However, transparency, human oversight and careful source selection are needed to ensure patient safety, as also discussed in chapter 4.2.
Tackling treatment adherence with AI tools
As acknowledged by the WHO, “increasing the effectiveness of adherence interventions may have a far greater impact on the health of the population than any improvement in specific medical treatments” [113]. The logical consequence is the development and adaptation of AI tools also for this purpose. Digital tools in general have been shown to be effective measures to increase adherence [114, 115].
However, research on AI tools in this specific area is still scarce [116]; No AI-enabled medical devices have been approved by the FDA in the field of allergy so far [117]. To target patients at high risk of low adherence, several predictive ML models have been evaluated to enable timely, personalized interventions. Applications range from subcutaneous immunotherapy [118], over asthma medication [119], to adolescents receiving personalized motivational messages or gamified reminders if they forget to take medication or carry their adrenaline autoinjector [120].
Lv et al. [119] performed a randomized controlled study proving that a mobile application-assisted intervention could improve adherence to anti-asthmatic medication in 152 children (6 – 11.9 years old), using medication reminders, exacerbation alerts, providing treatment recommendations, and facilitating instant communication with healthcare providers. Another example is the specifically developed PollDi App [92] for grass pollen-allergic patients that was tested in a stratified, controlled intervention study. The working group was able to show a significant improvement of nasal and conjunctival allergic symptoms of participants with a full version of the smartphone app vs. limited access.
Especially the availability of near real-time pollen data and forecasts led to increased self-medication and thereby reduced symptom severity. A Chinese randomized controlled trial was able to show that a smartwatch-based medication detection and reminder tool [121] significantly improved medication intake and relieved allergic rhinitis symptoms in the study group. Further, AI-driven wheeze detectors [122] and digital inhalers [123] have been identified as potentially valuable tools to support the correct application of short-acting beta-agonist and anti-inflammatory asthma medication, both in terms of timing and inhalation technique.
Among patients suffering from atopic dermatitis and chronic urticaria, changes in skin condition require adaptations of treatment plans, making adherence challenging for a proportion of patients [124].
Future fields of AI application in the area of atopic dermatitis involve systems assessing eczema control, thus recognizing deterioration of skin status or flares and encouraging patients to seek advice from their health care provider. Assessment of itch intensity via electronic questionnaires can also add to a better recognition of disease severity via AI-driven systems [125].
Although applications improving adherence to therapy (both topical and systemic) would possibly show positive effects on eczema control and patients’ quality of life, there is no disease-specific scientific evidence on the role of AI technologies to improve treatment adherence. However, several publications report the feasibility of monitoring drug intake using smart pill bottles [126], an approach that could be expanded to topical skin treatment dispensers. However, it remains to be discussed if this approach is favorable for the patient–doctor/nurse relationship.
Conclusion 8
In patient education and self-management, only a few digital applications in the field of allergy include advanced AI models. Current evidence is limited by mixed quality and a lack of standardized evaluation methods for both AI tools and adherence. However, AI has potential to improve these patient empowerment tools, particularly among those in underserved areas, or those at high risk of non-adherence.
Authors’ contributions
Oliver Pfaar: writing, conceptualization, development of design and methodology, review and editing. Stephan Traidl: writing, conceptualization, development of design and methodology, review and editing. Friederike Bärhold: writing, review, and editing.
Sven Becker: writing, review, and editing. Christoph Bergmann: writing, conceptualization, review, and editing. Jeroen Buters: writing, visualization, review, and editing.
Stephanie Dramburg: writing, conceptualization, review, and editing. Wojciech Francuzik: writing, conceptualization, review, and editing. Michael Gerstlauer: conceptualization, review, and editing.
Eckard Hamelmann: writing, review, and editing. Camilo José Hernandez Toro: writing, conceptualization. Ludger Klimek: review and editing.
Lars Lange: writing, visualization, review, and editing. Paolo M. Matricardi: conceptualization, review, and editing.
Laura Pfeffer: writing, review, and editing. Mariel Suarez-Suarez: writing, visualization, review, and editing. Philipp Utz: writing, review, and editing.
Martin Wagenmann: drafting, writing, conceptualization, review, and editing. Margitta Worm: writing, review, and editing. Mehrak Yoosefi: writing, conceptualization, review, and editing.
Sebastian Seurig: writing, management and coordination (co-lead), conceptualization, development of design and methodology, review, and editing. Sonja Mathes: writing, management and coordination (co-lead), conceptualization, development of design and methodology, review, and editing.
Funding
There was no external funding for this study.
Conflict of interest
Oliver Pfaar reports grants and/or personal fees and/or travel support from AEDA, Alfried Krupp Krankenhaus, ALK-Abelló, Allergopharma, Almirall, Altamira Therapeutics, ASIT Biotech, AstraZeneca, Bencard Allergie GmbH/Allergy Therapeutics, Biologixpharma, Blueprint, Breazy Health, Cliantha, Deutsche AllergieLiga e.V., Cogitando GmbH/medcram, Deutsche Forschungsgemeinschaft, Dustri-Verlag, ECM Expro&Conference Management GmBH, Forum für Medizinische Fortbildung, Georg-Thieme-Verlag, GSK, HAL Allergy Holding B.V./HAL Allergie GmbH, Inmunotek, Ingress Health, Institut für Disease Management, IQVIA Commercial, Japanese Society of Allergology, Königlich Dänisches Generalkonsulat, Laboratorios LETI/LETI Pharma, Lilly, Lofarma, Medcram/Cogitando, Medizinische Hochschule Hannover, med update europe GmbH, Meinhardt Congress GmbH, Novartis, Paul-Ehrlich-Institut, Paul-Martini-Stiftung, PneumoLive, Pohl-Boskamp, Procter & Gamble, Red Maple Trials Inc., Regeneron, RG Aerztefortbildung, ROXALL Medizin, Sanofi Aventis, Sanofi Genzyme, Springer Publisher, Stallergenes Greer, streamedup! GmbH, Technical University Dresden, John Wiley & sons publishers, Wort & Bild Verlag, Verlag ME, Gemeinnütziger Verein zur Förderung der Allergiefortbildung (VFAF), all outside the submitted work. Oliver Pfaar is board member of the ENT section of the European Academy of Allergy and Clinical Immunology (EAACI) and member of the extended board of directors of the German Society of Allergy and Clinical Immunology (DGAKI); coordinator, main- or co-author of different position papers and guidelines in rhinology, allergology and allergen-immunotherapy; and he is Editor-in-Chief of Clinical Translational Allergy.
Stephan Traidl has received honoraria for lectures or scientific advice from AbbVie, ALK, Almirall, Galderma, Janssen/JNJ, LEO Pharma, Lilly, Novartis, and Regeneron/Sanofi.
Friederike Bärhold received speaker honoraria or advisory board fees from GSK, Bencard, AstraZeneca, Leti, and travel reimbursement from HAL Allergy. FB is involved in research sponsored by Astra Zeneca, Sanofi, GSK, all outside the here-submitted work.
Sven Becker has received lecture fees and participated in advisory boards of Allergopharma, Allergy-Therapeutics, ALK-Abelló, Ambu, AstraZeneca, Bencard Allergie, GSK, HAL Allergie, MSD, Sanofi, Smart Reporting, Stallergenes; and received research grants from Allergy-Therapeutics, BMBF, Akribion, Polyrizon, BRAIN AG, Pari, and Altamira AG.
Christoph Bergmann received honoraria for presentations or advisory from Allergy Therapeutics, Bencard, Astra Zeneca, GSK, HAL Allergy, and SCS.
Jeroen Buters has received honoraria for lectures from ALK, Allergopharma, Stallergenes, AEDA, and DGAKI. Other research funding was obtained from DFG, EU, and the Bavarian Government LGL.
Stephanie Dramburg reports research funding from Thermo Fisher Scientific, speaker´s fees from Allergopharma, LETI, Bencard, dbv Technologies, OMRON Healthcare, Stallergenes, ALK, Siemens Healthineers, and is member of an advisory board for Bencard, as well as member of a data safety monitoring board for Allergy Therapeutics.
Wojciech Francuzik declares no COI.
Michael Gerstlauer is involved in studies with Allergopharma, Bencard Allergie, Boehringer, InfectoPharm, Stallergenes-Greer, has served at consulting for and received speaker honoraria from ALK-Abelló, Allergopharma, Bencard Allergie, Leti Pharma, Meda Pharma, Sanofi, Stallergenes-Greer, is part of specialist societies AGPAS, GPAU, GPP, DGKJ, EAACI, ERS, and is involved in AWMF guidelines for Allergen Immunotherapy, Insect Venom Allergy, Allergy Prevention, Severe Bronchial Asthma In Childhood, Allergic Rhinoconjunctivitis, S2k Bronchial Asthma, National Disease Management Guideline Bronchial Asthma, and the EAACI guideline on Allergen Immunotherapy.
Eckard Hamelmann declares no COI.
Camilo José Hernandez Toro declares no COI.
Ludger Klimek has received research grants from Allergy Therapeutics/Bencard, United Kingdom/Germany; ALK-Abelló, Denmark; Allergopharma, Germany; ASIT Biotech, Belgium; AstraZeneca, Sweden; Bionorica, Germany; Biomay, Austria; Boehringer Ingelheim, Germany; Circassia, USA; Stallergenes, France; Cytos, Switzerland; Curalogic, Denmark; HAL, the Netherlands; Hartington, Spain; Lofarma, Italy; Viatris/Mylan, USA; Novartis, Switzerland; Leti, Spain; Roxall, Germany; GlaxoSmithKline (GSK), UK; and Sanofi, France, and/or has served as a speaker or worked as a consultant for the aforementioned pharmaceutical companies. LK is President of the Ärzteverband Deutscher Allergologen (AeDA), Vice President of the German Academy for Allergy and Environmental Medicine (DAAU), and a member of the Governance and NAIS Committees of the European Academy of Allergy and Clinical Immunology (EAACI); he is also editor-in-chief of *AllergoJournal* and *AllergoJournal International*.
Lars Lange has received lecture fees and participated in advisory boards from Stallergenes, Allergopharma, Infectopharm, HAL Allergy, Thermo Fisher Scientific, DBV technologies, Novartis, Nestle Nutritional Institute, Nutricia, Sanofi, all outside the submitted work.
Paolo M. Matricardi reports research support, speaker and/or consultancy fees from Allergopharma, Euroimmun, Hycor, MAD-X, Omron Healthcare, Stallergenes-Greer, Thermo Fisher Scientific, research support from TPS software production, speaker and/or consultancy fees from ALK-Abellò, Angelini, Bencard Allergie, HAL, all outside the submitted work.
Laura Pfeffer declares no COI.
Mariel Suarez-Suarez declares no COI.
Philipp Utz has received honoraria for lectures or scientific advice from Aimmune, Allergopharma, ALK-Abelló, Berlin-Chemie, DAAB, Nutricia, PARI, and Sanofi-Regeneron.
Martin Wagenmann has received lecture fees and participated in advisory boards of Allergopharma, ALK-Abelló, AstraZeneca, Celltrion, CSL Behring, Genzyme, GSK, HAL Allergie, MSD, NeilMed, Regeneron, Sanofi, Stallergenes; and received research grants from ALK-Abelló, AstraZeneca, EU, GSK, Novartis, Regeneron, Sanofi, Takeda.
Margitta Worm reports honoraria and/or consulting fees from AbbVie, ALK-Abelló, Allergopharma, Almirall, Amgen, Boehringer Ingelheim, DBV Technologies, Genzyme, Kymab, LEO Pharma, Eli Lilly, Mylan (Viatris), Novartis, Pfizer, Regeneron Pharmaceuticals Inc., Sanofi, Stallergenes Greer.
Mehrak Yoosefi declares no COI.
Sebastian Seurig reports personal financial interests: ALK-Abelló (consulting, advisory board); AstraZeneca (lectures/training); Allergopharma (lectures); Bencard Allergy (advisory board).
Sonja Mathes received speaker honoraria from Thermo Fisher, ALK-Abelló, AstraZeneca and Bencard and travel reimbursement by ALK-Abelló and Bencard.
EQUATOR AGREE II
Compliance with EQUATOR AGREE II statement.
Table 1.Emerging applications of clinical decision support systems in allergology. Clinical fieldEmerging CDSS applicationsApplicability and clinical benefit in allergologyPediatric practicePredominantly rule-based CDSS for triage and diagnostic support, with emerging AI-based CDSS for acute care risk stratification (e.g., pneumonia detection [96]). AI-enhanced assistance tools for communication and clinical documentation, including telephone triage and speech-to-text systems [97].Support of triage and prioritization in pediatric emergency and outpatient settings; assistance in asthma phenotyping and endotyping; structured assessment of symptom patterns and exacerbation risk; reduction of documentation burden and optimization of clinical workflows.Dermatology/allergological dermatologyEmerging CDSS for monitoring inflammatory skin diseases [98]; AI-based tools for precision and semi-automated skin-prick testing [32].Reduction of interobserver variability; prioritization of cases; adjunctive support for monitoring disease activity in atopic dermatitis and chronic urticaria.ENT allergology (rhinology/conjunctiva)Emerging CDSS for structured symptom assessment, interpretation of nasal and conjunctival provocation tests, integration of clinical, endoscopic, and laboratory data, etiological diagnosis of seasonal allergic rhinitis, allergen immunotherapy prescription, day-to-day pharmacological management, and short-term prevention of seasonal allergic rhinitis [89].Support for indication, standardization, and interpretation of nasal and conjunctival allergen provocation testing; improved reproducibility and safety; assistance in the etiological diagnosis of allergic rhinitis and differentiation from non-allergic rhinitis; facilitation of guideline-adherent diagnostic pathways.Cross-specialty clinical tasksCDSS for real-time data integration, structured anamnesis, and interpretation of complex laboratory data (e.g., IgE microarrays and molecular allergy diagnostics) [99].
AI-enhanced assistance tools for communication and clinical documentation, including telephone triage and speech-to-text systems [100].Support for standardization of allergological diagnostic work-ups; facilitation of interdisciplinary clinical decision-making; improved consistency in complex cases (e.g., polysensitization or multimorbidity); support for precision-oriented diagnostic and therapeutic strategies.Administrative and organizational processesAI-supported CDSS for diagnostic coding and billing [101].Optimization of practice organization; reduction of administrative workload; improved coding accuracy; indirect quality gains through reallocation of clinician time toward patient care.CDSS = clinical decision support systems; AI = artificial intelligence.