Two-step deep-learning candidemia prediction model using two large time-sequence electronic health datasets

Area of Science:

• Medical Informatics
• Computational Biology
• Infectious Diseases

Background:

• Candidemia is a life-threatening bloodstream infection with poor predictive accuracy using current methods.
• Delayed empiric antifungal therapy is common, even in high-risk individuals.

Purpose of the Study:

• To develop and validate a deep learning model for predicting 7-day candidemia risk.
• To implement a two-step framework integrating candidemia and mortality risk to guide antifungal therapy decisions.

Main Methods:

• A deep learning model (PyTorch_EHR) was developed using electronic health record data from two large cohorts (HMHS and MIMIC-IV).
• Model performance was compared against logistic regression, LightGBM, and existing candidemia scores.
• A two-step prediction framework combined candidemia and 30-day mortality risk models.

Main Results:

• The deep learning model outperformed other methods in predicting candidemia across both cohorts.
• The two-step framework identified additional candidemia cases, improving coverage for high-risk patients.
• A significant proportion of identified high-risk patients with high mortality did not receive empiric antifungal therapy.

Conclusions:

• A two-step deep learning framework can enhance early identification of patients at high risk for candidemia.
• This approach may facilitate more timely initiation of empiric antifungal therapy.
• Further prospective studies are needed to validate these findings.