Predicting SARS-CoV-2 infection among hemodialysis patients using deep neural network methods

Area of Science:

• Infectious Diseases
• Artificial Intelligence in Healthcare
• Nephrology

Background:

• Dialysis patients face higher COVID-19 morbidity and mortality risks.
• Early identification of SARS-CoV-2 infections is crucial for implementing control measures in dialysis centers.
• Existing predictive models often require complex feature engineering and show limitations in performance.

Purpose of the Study:

• To develop and evaluate deep learning models for predicting severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infections in dialysis patients during the incubation period.
• To compare the predictive accuracy of deep learning models against traditional machine learning approaches.
• To identify key predictive features for early COVID-19 diagnosis in this vulnerable population.

Main Methods:

• Collected diverse data: demographics, clinical, treatment, laboratory, vaccination, socioeconomic status, and COVID-19 surveillance.
• Developed and implemented deep learning models, specifically Recurrent Neural Networks (RNN), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM).
• Compared deep learning models against traditional methods like logistic regression, Support Vector Machines (SVM), and XGBoost, focusing on accuracy and feature engineering requirements.

Main Results:

• Deep learning models, particularly LSTM and CNN, demonstrated superior accuracy in predicting SARS-CoV-2 infections compared to traditional models.
• The LSTM model achieved an Area Under the Curve (AUC) exceeding 0.80, peaking at 0.91, significantly outperforming XGBoost.
• At a 20% false positive rate, LSTM and CNN identified 66% and 64% of positive cases, respectively, compared to 42% for XGBoost.

Conclusions:

• Deep neural networks, especially LSTM and CNN, offer a more accurate and efficient approach to predicting COVID-19 in dialysis patients during incubation.
• Minimal feature engineering is required for deep learning models, simplifying their implementation.
• Early detection through these advanced models can lead to timely interventions, potentially reducing severe outcomes for dialysis patients.