Entry number:14
SUBJECT
AKI Helper: we can help Real time Predict Acute Kidney Injury
DEPARTMENT
Department of Computer Science and Information Engineering
PROFESSOR AND STUDENT
Professor:Jung-Hsien Chiang
Student:Ling-Xuan, Ying
ABSTRACT
According to Environmental Protection Administration statistics, there were no less than 92 thousands people there undergoing dialysis in 2019. The medical costs of 2019 are about 53.3 billion. The problem of kidney disease is so series that we can’t ignore this anymore.
We set a target of predicting Acute Kidney Injury ( AKI ) sequentially. We used MIT MIMIC Data and eICU Collaborative Research Database train our model. For feature selection, we used multiple simple models to train the data and analyze the results. After consulting with the doctor and referring to the analyses, we selected 18 features for training data. For the sequential characteristics in the common hospital data, we use LSTM as our final model, and splitting data per four hours into one group, then concatenating 6 groups as one training data with a single ground truth which is 0 or 1. The model will output probability between 0-1 per four hours, in order to generate sequential results to be alert to the AKI probability increasing with time. Finally, we deployed our LSTM model to the web, so that doctors can use our research easily.
We built on a foundation of predicting the AKI probability to gain time, reduce the cost and improve the quality in medical domain. In the future, we can use the foundation to predict the recovery ratio, the dialysis rate and so on after AKI.