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

Dimension reduction is considered as a necessary technique in Electronic Healthcare Records (EHR) data processing. However, no existing work addresses both of the two points: 1) generating low-dimensional representations for each patient visit; and 2) taking advantage of the well-organized medical concept structure as the domain knowledge. Hence, we propose a new framework to generate low-dimensional representations for medical data records by combining the concept-structure based distance with manifold learning. To demonstrate the efficacy, we generated low-dimensional representations for hospital visits of heart failure patients, which was further used for a 30-day readmission prediction. The experiments showed a great potential of the proposed representations (AUC = 60.7%) that has comparative predictive power of the state-of-the-art methods, including one hot encoding representations (AUC = 60.1%) and PCA representations (AUC = 58.3%), with much less training time (improved by 99%). The proposed framework can also be generalized to various healthcare-related prediction tasks, such as mortality prediction.

Share

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

MD-Manifold: A Medical Distance Based Manifold Learning Approach for Heart Failure Readmission Prediction

Online

Dimension reduction is considered as a necessary technique in Electronic Healthcare Records (EHR) data processing. However, no existing work addresses both of the two points: 1) generating low-dimensional representations for each patient visit; and 2) taking advantage of the well-organized medical concept structure as the domain knowledge. Hence, we propose a new framework to generate low-dimensional representations for medical data records by combining the concept-structure based distance with manifold learning. To demonstrate the efficacy, we generated low-dimensional representations for hospital visits of heart failure patients, which was further used for a 30-day readmission prediction. The experiments showed a great potential of the proposed representations (AUC = 60.7%) that has comparative predictive power of the state-of-the-art methods, including one hot encoding representations (AUC = 60.1%) and PCA representations (AUC = 58.3%), with much less training time (improved by 99%). The proposed framework can also be generalized to various healthcare-related prediction tasks, such as mortality prediction.

https://aisel.aisnet.org/hicss-54/ks/augmented_intelligence/3