Dimensionality reduction and ensemble of LSTMs for antimicrobial resistance prediction

Area of Science:

• Medical Informatics
• Computational Biology
• Infectious Diseases

Background:

• Rising bacterial antibiotic resistance poses a significant global health threat.
• Nosocomial infections, particularly in Intensive Care Units (ICUs), are exacerbated by antibiotic-resistant pathogens like Pseudomonas aeruginosa.
• Effective prediction of antibiotic resistance is vital for timely and appropriate patient treatment.

Purpose of the Study:

• To develop a predictive model for antibiotic resistance in Pseudomonas aeruginosa nosocomial infections within an ICU setting.
• To identify key risk factors contributing to antibiotic resistance using a data-driven approach.
• To leverage Long Short-Term Memory (LSTM) artificial neural networks for accurate prediction.

Main Methods:

• Utilized Multivariate Time Series modeling on Electronic Health Records (EHR) data from 2004-2019.
• Developed a data-driven dimensionality reduction method incorporating feature importance techniques and LSTM sequential capabilities.
• Employed an ensemble of LSTMs to enhance prediction stability and reduce performance variance.

Main Results:

• Identified patient admission information, administered antibiotics, and prior antimicrobial resistance as critical risk factors.
• The proposed dimensionality reduction approach improved performance and reduced feature count compared to conventional methods.
• The LSTM-based framework demonstrated computational efficiency and promising predictive performance.

Conclusions:

• The developed LSTM framework offers a computationally efficient solution for predicting antibiotic resistance in ICU settings.
• The model effectively handles high dimensionality, data scarcity, and concept drift common in clinical data.
• This approach supports clinical decision-making by identifying key risk factors for Pseudomonas aeruginosa infections.