D2.515 - Symptom forecasting for birch and grass pollen allergy: the #berlinbreathing project

Symptom forecasting for seasonal allergic rhinitis (SAR) has been recently achieved in a German study cohort. However, it is unclear whether these results can be replicated in a similar yet independent population. Objective: To test the replicability of a previous SAR symptom forecast model (Holzmann et al. 2025) with data from an independent cohort of pollen allergic patients and healthy controls.

The #berlinbreathing cohort is a single-center longitudinal cohort of patients with diagnosed SAR and healthy controls from Berlin, Germany. Clinical, demographic, and multiplex IgE data were collected upon inclusion. During the pollen season 2025, participants filled a daily e-Diary (Pollenius^TM, TPS, Rome, Italy) on organ-specific SAR-symptoms, medication use, and time spent outdoors. Further, meteorological variables and daily pollen counts for birch and grasses were included in the forecasting model. A fraction of the participants (75%) was used to train XGboost classification models predicting nose, eye, and lung symptoms (each on a 0-2 severity scale) for the next day, respectively. Model performance was estimated on the test data set (25%). 

The 162 participants of #berlinbreathing (136 with SAR, 28 healthy controls) recorded 16.479 diary entries, of which 13.637 had complete data for the 4 consecutive days needed for the models (training: n=10.096; test: n=3.541). Next-day organ-specific symptoms were predicted on the test set with an average precision of 0.78, 0.69, and 0.80 for nose, eye, and lung symptoms, respectively.  Average sensitivity (0.77, 0.69, 0.77) and overall accuracy (0.81, 0.78, 0.90) on the test set were better for nose and lung symptoms than for eye symptoms. Among all included variables, lagged symptom reports had the highest feature importance.

The organ-specific symptom classification models had a comparable performance to those reported by Holzmann et al. (2025). These findings provide an independent replication of a modelling framework for the prediction of daily pollen allergy symptoms.