Background
Air pollution has become an important health issue in Taiwan. Air pollutants can cause exacerbations of pre-existing asthma of other allergic diseases. Identifying of the relationship between asthma attack and air pollution is an important issue for improving the quality of the diseases care. Electronic health record (EHR) has been successfully used for predicting occurrences of a variety of diseases. Very few studies were based on the patient’s health factors, such as data from EHR, with air pollution database.

Methods
In our recent work, we have successfully applied data augmentation with active-learning selection strategy in using EHR for young stroke prediction. This method can achieve higher performance for prediction of the disease occurrence than non-data-augmented approaches. In this study, machine learning (ML) and deep learning (DL) were used on large clinical databases to predict the acute attacks of asthma.

Results
For asthma acute attack prediction, models using SL (simple learning approach) and DA2 (data augmentation with non-attack controls from similar disease cohort) can get better performance than those using other data augmentation method (DA1 or DA3) in ML models. For DL models, models using DA2 can also get better performance than those using SL, DA1 or DA3. The performance of DNN1 (DL with 1 fully connected hidden layer) and DNN2 (DL with 2 fully connected hidden layers) is better than that of DNN3 (DL with 3 fully connected hidden layers). Logistic regression (LR) and gradient boosting decision tree (GBDT) get the highest AUROC values (0.72) for acute attack prediction in asthma cohort.

Conclusions
In conclusion, algorithms based on transfer learning can outperform the commonly used approaches. Further research is necessary to determine whether the use of these algorithms could lead to an improved care quality for asthma patients.