AI IN severe asthma showed poor agreement with multidisciplinary teams when selecting biologic therapy for uncontrolled disease.
Researchers compared multidisciplinary team decisions with recommendations from five large language models in 114 adults with biologic-naive, severe uncontrolled asthma. All patients were considered for treatment with benralizumab, dupilumab, or tezepelumab after structured specialist assessment.
Agreement between the multidisciplinary team and individual AI models was poor, with Cohen κ values ranging from 0.143 to 0.213. Agreement among the AI models themselves was also limited, ranging from κ 0.348 to 0.586.
The multidisciplinary team selected benralizumab for 41 patients, dupilumab for 28, and tezepelumab for 45. Across the AI models, tezepelumab was consistently recommended less often than the other biologics.
The investigators identified three recurring types of AI error: lack of logical consistency, decisions unsupported by current evidence, and nonsensical output. Unsupported recommendations were the most common.
Performance was particularly variable among patients with high levels across all three Type 2 biomarkers. In this subgroup, first choice agreement with the multidisciplinary team ranged from 18.2% to 29.5% across the five models.
The models also sometimes handled biomarkers and comorbidities inconsistently. In some cases, current blood eosinophil counts were prioritized over historical peaks, potentially changing phenotype classification. Two recommendations also fell outside the three biologic options explicitly provided in the prompt.
Biologic selection in severe asthma extends beyond inflammatory biomarkers. The study emphasized that multidisciplinary assessment incorporates comorbidities, prior treatments, patient preferences, adherence, inhaler technique, functional status, and psychosocial factors that may affect the suitability of a therapy.
Among 83 patients who had received their multidisciplinary team selected biologic for more than 4 months, 52 had undergone formal clinical review. Responses were rated excellent in 20 patients, good in 13, partial in 10, while 9 discontinued or failed treatment because of inadequate response or adverse events.
The findings do not establish that every AI recommendation was clinically incorrect, because the study measured agreement with real world multidisciplinary decisions rather than comparative treatment outcomes. However, the observed discordance, hallucinations, and inconsistent reasoning indicate that current AI in severe asthma is not suitable as a standalone or adjunctive decision maker for biologic therapy.
The authors concluded that multidisciplinary team assessment remains the reference standard, while future AI systems may warrant reevaluation as models and clinical validation improve.
Reference
Suter P et al. Comparison of artificial intelligence and multidisciplinary team assessment for evaluating biologic decisions in severe uncontrolled asthma. J Allergy Clin Immunol Pract. 2026. doi:10.1016/j.jaip.2026.07.049.
Featured Image: fizkes on Adobe Stock.
Featured Image: fizkes on Adobe Stock.
Each article is made available under the terms of the Creative Commons Attribution-Non Commercial 4.0 License.
