Projects Projects

PROJECTS

HOME
Finished Projects
以深度學習建構智慧多模式臨床治療成效預測系統
01
JAN
2019
Dec
2019
Chen-Ying Hung
Background
Cancer is a leading cause of death in Taiwan. Targeted therapy is most often used for advanced cancers. Previous works mostly focus on prediction the efficiency of targeted therapy by using genetic mutation of cancer cell. Very few studies were based on the patient’s health factors, such as data from electronic health records (EHR). EHR has been successfully used for predicting occurrences of a variety of other diseases. However, inadequate predictive performances have been observed in cases of rare occurrences due to both insufficient training samples and highly imbalanced class distribution.

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, deep learning was used on large clinical databases to predict mortality and target therapy related adverse event of cancer patients.

Results
For mortality prediction, the area under the receiver operating characteristic (AUROC) for deep learning model (0.814) is better than other machine learning approaches (0.720 for random forest [RF], 0.745 for gradient boosting decision tree [GBDT], 0.747 for logistic regression [LR]). For overall target therapy related adverse event prediction, the AUROC for deep learning model (0.747) is better than other machine learning approaches (0.716 for RF, 0.731 for GBDT, 0.729 LR). For Gefitinib related adverse event prediction, the performance of deep learning with transfer learning (AUROC 0.715) is better than deep learning models directly training with Gefitinib dataset (AUROC 0.708) or datasets of all target drugs (AUROC 0.695).

Conclusions
In conclusion, deep learning can achieve better performance than other machine learning approaches for predicting mortality and target therapy related adverse event of cancer patients. Applying transfer learning can improve performance of Gefitinib related adverse event prediction.
PARTNER