Location

Online

Event Website

https://hicss.hawaii.edu/

Start Date

4-1-2021 12:00 AM

End Date

9-1-2021 12:00 AM

Description

Warfarin is a widely used oral anticoagulant worldwide. However, due to the complex relationship between individual factors, it is challenging to estimate the optimal warfarin dose to give full play to its ideal efficacy. Currently, there are plenty of studies using machine learning or deep learning techniques to help with the optimal warfarin dose selection. But few of them can resolve missing values and high-dimensional data naturally, that are two main concerns when analyzing clinical real world data. In this work, we propose to regard each patient’s record as a set of observed individual factors, and represent them in an embedding space, that enables our method can learn from the incomplete date directly and avoid the negative impact from the high-dimensional feature set. Then, a novel neural network is proposed to combine the set of embedded vectors non-linearly, that are capable of capturing their correlations and locating the informative ones for prediction. After comparing with the baseline models on the open source data from International Warfarin Pharmacogenetics Consortium, the experimental results demonstrate that our proposed method outperform others by a significant margin. After further analyzing the model performance in different dosing subgroups, we can conclude that the proposed method has the high application value in clinical, especially for the patients in high-dose and medium-dose subgroups.

Share

COinS
 
Jan 4th, 12:00 AM Jan 9th, 12:00 AM

Warfarin Dose Estimation on High-dimensional and Incomplete Data

Online

Warfarin is a widely used oral anticoagulant worldwide. However, due to the complex relationship between individual factors, it is challenging to estimate the optimal warfarin dose to give full play to its ideal efficacy. Currently, there are plenty of studies using machine learning or deep learning techniques to help with the optimal warfarin dose selection. But few of them can resolve missing values and high-dimensional data naturally, that are two main concerns when analyzing clinical real world data. In this work, we propose to regard each patient’s record as a set of observed individual factors, and represent them in an embedding space, that enables our method can learn from the incomplete date directly and avoid the negative impact from the high-dimensional feature set. Then, a novel neural network is proposed to combine the set of embedded vectors non-linearly, that are capable of capturing their correlations and locating the informative ones for prediction. After comparing with the baseline models on the open source data from International Warfarin Pharmacogenetics Consortium, the experimental results demonstrate that our proposed method outperform others by a significant margin. After further analyzing the model performance in different dosing subgroups, we can conclude that the proposed method has the high application value in clinical, especially for the patients in high-dose and medium-dose subgroups.

https://aisel.aisnet.org/hicss-54/hc/big_data_on_healthcare_app/9