Antimicrobial resistance (AMR) is one of the leading causes of death worldwide. In 2019, bacterial AMR was found to be associated with approximately 4.95 million of deaths across 204 countries with 1.2 million of them directly attributed to it. Misuse and overuse of antibiotics are significant drivers in the development of AMR. To deal with AMR, a multi-modal approach is needed including antimicrobial stewardship to preserve the effectiveness of currently available agents. Recent research has focused on using electronic health records (EHRs) for infection diagnoses and antibiotic therapy selection through machine learning (ML), but little work has focused on antimicrobial stewardship. Inadequate, ineffective or incomplete antibiotic treatment often leads to antibiotic retreatment (antibiotic readmission), causing unnecessary and excessive use of antibiotics.
This project applied electronic health records of 2,189 intensive care unit (ICU) patients from the MIMIC-IV database to develop ML-based decision support models using deep learning approach to reliably predict whether ICU patients will be retreated with antibiotic if current antibiotic treatment is discontinued. This project is formulated and addressed as both classification and regression tasks. Accurately predicting antibiotic readmission (antibiotic retreatment) for ICU patients could be extremely helpful in optimising antibiotic treatment to combat AMR by providing individualised predictions to support the decision on continuation or cessation of antibiotic treatment.
Key Contributions of FYP
- Constructed data preprocessing pipeline to generate dense time-series representation as input for deep learning models
- Identified a range of routinely collected EHR clinical variables that are meaningful for antibiotic readmission prediction
- Explored different advanced deep learning architectures like LSTM and CNN for antibiotic readmission prediction as well as their performance in both classification and regression tasks
- Proposed a new deep learning architecture with very promising performance in antibiotic readmission prediction - both binary and regression tasks
- Explored joint learning method to improve the prediction performance by training the model with both binary and continuous labels
- Provided critical evaluation of the results and interpreted the logic behind the prediction