A study of the effectiveness of transfer learning in individualized asthma risk prediction

Deep Learning classifiers require a vast amount of data to train models that generalize well and perform effectively on unseen data. However, small sizes of training data, especially in the medical domain, make this a challenging task. Transfer Learning (TL) can help overcome a scarcity of data by focusing on fine tuning a pre-trained model with a small amount of specialized training data. In the last few years, several studies have been performed on TL with medical images, and they point towards significant gains available with this method. However, to date no such studies have been performed in the area of individualized asthma prediction with limited training data for each patient. In this paper, we conduct a systematic study of transfer learning in this domain in the context of neural networks. Our TL approach trains the source model with population data of 25 asthma patients and then retrains the target model with a target patient's data. Our results show that transfer learning yields promising results in improving model performance on an individual basis. Further research directions that are worth investigating based on our results are pointed out as future work directions.