Clinical decision support and workflow optimisation

AI-driven clinical decision support systems (CDSS), tools that assist clinicians in making evidence-based decisions at the point of care, have significant potential to transform allergy care [15]. For example, systems like HELIOT have demonstrated near-perfect accuracy in managing drug allergy alerts, highlighting AI’s capacity to improve clinical safety and efficiency [15, 16]. Integrated into EHRs, these tools provide real-time decision support during consultations and ward rounds, reducing alert fatigue, enhancing prescription safety, and improving adherence to evidence-based guidelines [16].

Beyond drug allergy management, CDSS in allergy settings can evolve to offer real-time asthma control recommendations, immunotherapy initiation criteria, and diagnostic guidance such as specific IgE or component-resolved testing. For example, asthma-focused CDSS that integrate patient data into the existing Global Initiative for Asthma (GINA) framework or National Asthma Education and Prevention Program (NAEPP) guidelines have shown improvements in standardised severity scoring and treatment step recommendations [17]. However, the study revealed suboptimal utilisation of the CDSS and limited adherence to the provided recommendations [17].

Advanced CDSS capabilities include dynamic risk stratification [18]. In food allergy clinics, for instance, systems can estimate allergen thresholds by analysing prior reactions and co-factors like infections or exertion, informing decisions about food challenges or home introduction, especially critical in paediatric settings.

CDSS also supports interdisciplinary care [18]. For example, an AI tool might flag PEG-asparaginase allergy before scheduling contrast-enhanced imaging. In dermatology and gastroenterology, CDSS can pre-emptively alert clinicians to risks before initiating biologics or immunomodulators.

With embedded NLP, CDSS can analyse patient histories to extract data on patterns of allergic reactions, including potential triggers and precipitants. For example, recurrent idiopathic flares in chronic urticaria may trigger recommendations for omalizumab evaluation.

Ultimately, as CDSS platforms evolve with ML, they can learn from outcome data, refining recommendations based on prior intervention success. This adaptive capability is amongst the promising strengths for CDSS and AI more broadly, training itself on patient data to fine-tune the foundation model in readiness for future patient interactions.

Virtual assistants and remote monitoring

Conversational AI tools, including chatbots, mobile applications, and AI-enabled medical devices, offer significant potential for supporting asynchronous patient engagement and disease monitoring in asthma and allergic rhinitis. While distinct in their design and purpose, these technologies can incorporate advanced AI functionalities to enhance usability and clinical relevance. Studies have demonstrated that telemedicine approaches improve treatment adherence and reduce exacerbations [19], outcomes that AI tools may help replicate or amplify through intelligent automation and real-time feedback [20]. These tools are particularly valuable in paediatric care and multilingual environments, where continuous, tailored communication is critical to patient outcomes.

NLP for documentation and clinical letters

NLP algorithms enable transcription, summarisation, and automated clinical letter generation, offering significant value in allergy clinics where thorough histories and consistent follow-up are essential [21, 22]. NLP tools can extract critical medical information, such as allergen triggers, symptom patterns, and past immunotherapy responses, from unstructured clinical notes [23]. When paired with ambient voice transcription, clinicians can dictate during or after consultations, with the NLP engine transcribing and formatting the content into structured clinical letters or discharge summaries. This reduces documentation time and enables immediate information sharing with referring physicians, enhancing communication and care coordination. More broadly, such tools can support clinician engagement with patients, particularly in specialties like paediatrics, where frequent in-person interactions benefit from streamlined, AI-assisted workflows.

Advanced NLP models trained on allergy-specific data can identify subtle distinctions, for example, differentiating IgE- from non-IgE-mediated reactions or recognising contextual indicators of anaphylaxis risk [23, 24]. In research, NLP supports high-throughput screening of EHRs to identify patients eligible for clinical trials, using allergy diagnosis codes, medication histories, and lab results [23, 24]. Clinically, NLP enables real-time decision support by detecting missing elements in documentation, such as incomplete immunotherapy schedules or omitted rescue medications. Some systems now offer live prompts during note-taking to ensure accuracy and completeness.

In allergy clinics, the implementation of AI-driven documentation tools such as ambient scribing systems has been shown to significantly enhance clinical workflow efficiency. For example, one study found that ambient scribe tools reduced clinicians’ time in notes per appointment by about 20.4% (from 10.3 to 8.2 minutes), increased same-day appointment closure by 9.3%, and decreased after-hours documentation by 30% [25]. Clinicians also reported lower mental burden and greater engagement with patients during consultations.

Critical allergy alert identification

AI systems can extract structured and unstructured data from EHRs to detect historical allergy events and flag potential triggers [26]. These tools use NLP to identify past reactions recorded in clinician notes, prescription histories, or scanned documents that may not be immediately visible in allergy fields. For instance, a prior reaction to an antibiotic may be recorded in a discharge summary or old correspondence but not coded in the structured allergy list. AI can flag such cases and prompt clinician review, potentially preventing serious medication errors. However, over-reliance on automated alerts may contribute to alarm fatigue, highlighting the need for carefully calibrated systems that prioritize clinical relevance.

Beyond static alerts, AI platforms are evolving to offer contextual and dynamic allergy alerts. These systems consider factors like current medications, lab results (e.g., elevated eosinophil counts), or recent diagnoses (e.g., new-onset asthma in the context of pollen season) to infer possible allergenic risk [24]. In paediatrics, where caregiver reports and evolving symptom profiles are key, AI can reconcile inconsistent entries and build a longitudinal allergy profile that updates in real time.

Hospitals are beginning to deploy AI tools for preoperative assessment to automatically flag patients with perioperative anaphylaxis risk based on past exposure to dyes, latex, or anaesthetic agents [27]. These alerts can help streamline perioperative planning and incorporate allergist input into complex perioperative plans.

Predicting OFC outcomes

OFCs remain the gold standard for confirming food allergies, but they are resource-intensive, costly, and anxiety-inducing—particularly in paediatric patients—and carry a significant risk of anaphylaxis. To mitigate these concerns, AI models such as logistic regression, support vector machines, and ensemble learning have been trained on combinations of skin prick test wheal sizes, specific IgE levels, component-resolved diagnostics (CRD), and clinical history to predict OFC outcomes. Support vector machines, which are algorithms that separate patients into groups based on patterns in test results, and ensemble learning approaches, which combine multiple models to improve accuracy, have been used to predict OFC outcomes [9]. These models consistently report area under the curve (AUC) between 0.85 and 0.96 for peanut, egg, and milk allergy prediction [9].

Biomarker discovery and interpretation

Multi-omics data, ranging from genomics and transcriptomics to metabolomics, can now be processed by unsupervised clustering and DL models to discover molecular signatures [29]. For example, latent class analysis and neural networks have helped define distinct inflammatory phenotypes in eosinophilic esophagitis and severe asthma. CRD, which analyse sensitisation to individual allergen proteins, can be interpreted using AI to differentiate primary from cross-reactive sensitisation [29].

Integrating imaging and phenotyping

ML also plays a role in phenotype clustering for conditions like asthma and chronic rhinosinusitis with nasal polyposis [30]. Using data from spirometry, FeNO levels, eosinophil counts, and symptom scores, unsupervised learning algorithms (e.g., k-means clustering, hierarchical clustering) can identify endotypes that may respond differently to biologics or environmental interventions [30].

Personalisation of treatment protocols

Traditionally, immunotherapy schedules have been developed based on population-level data and generalised protocols. AI-driven systems now allow for the stratification of patients using ML algorithms trained on a combination of clinical phenotypes, specific IgE patterns, CRD, comorbidities, environmental exposures, and even genomic information [31, 32]. These models can predict which patients are likely to benefit from sublingual vs. subcutaneous immunotherapy, determine optimal allergen selection, and guide the timing and duration of treatment.

Predictive analytics further identifies patients at risk of poor response or adverse events. One study used historical adherence and immunologic markers to forecast non-adherence in subcutaneous immunotherapy with 60–84% accuracy via sequential models [33].

For example, combining AI with pharmacogenomics opens new possibilities for tailoring biologic therapy in autoimmune diseases. TNF-α polymorphisms can influence response to etanercept in psoriatic arthritis, highlighting the role of genomic profiling in treatment choice [34]. Given the variable infection risks linked to TNF-α inhibitors, integrating genomic, clinical, and AI-driven insights may enable more precise decisions that maximize efficacy while reducing adverse events.

Monitoring during immunotherapy

AI has also made real-time monitoring during immunotherapy feasible and scalable. Wearable biosensors that track heart rate, respiratory rate, oxygen saturation, and skin conductance can feed data to AI systems that continuously evaluate for early signs of systemic reactions [35]. Integration with patient-facing apps allows for two-way communication: patients can log symptoms or side effects, and AI algorithms can automatically flag concerning trends for clinician review.

These platforms are particularly useful in home-based sublingual immunotherapy, where timely feedback from patients may be limited. In such settings, AI tools can act as virtual safety nets, enhancing patient autonomy while maintaining clinical oversight, but they also require robust regulatory frameworks to ensure patient safety remains paramount.

One proof-of-concept study utilized AI techniques—such as artificial neural networks and genetic algorithms—to develop a decision support system for allergen-specific immunotherapy [36]. This system achieved a 72% success rate in predicting symptom progression and therapy outcomes, illustrating how AI can meaningfully inform personalized treatment decisions [36].

Adaptive dosing and reinforcement learning

Reinforcement learning and Bayesian optimisation are increasingly adopted to dynamically adjust dosing based on immunologic feedback (e.g., IgG4 or basophil activation) and clinical response. While much work has focused on cancer immunotherapy [37], the framework is potentially extensible to allergen dosing for venom or food immunotherapy.

Digital coaching, adherence, and engagement

Digital tools supported by AI, such as virtual nurse assistants, daily adherence prompts, and behaviour prediction models, can improve patient engagement [38]. These systems not only remind patients to take doses but also track patterns in missed doses and side effects, enabling clinicians to intervene early. Some platforms use gamification or psychological behaviour modelling to enhance motivation, particularly in adolescents and young adults.