Identifying infected patients using semi-supervised and transfer learning.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Decision Support

Background:

• Early infection identification is crucial for improving patient outcomes.
• Manual chart review for infection status is labor-intensive and limits sample size for model development.
• Developing automated models for early infection detection is essential for large-scale clinical application.

Purpose of the Study:

• To compare semi-supervised and transfer learning algorithms against traditional manual chart review for identifying infections in hospitalized patients.
• To evaluate the performance of different machine learning approaches in early infection detection.
• To assess the impact of transfer learning and semi-supervised methods on model discrimination and calibration.

Main Methods:

• A multicenter retrospective study involving admissions from 6 hospitals.
• Utilized "gold-standard" infection labels from manual chart review and "silver-standard" labels from Sepsis-3 criteria.
• Developed deep learning and non-deep learning models using transfer learning and semi-supervised methods with patient data from the first 24 hours of admission.

Main Results:

• Deep learning and non-deep learning models demonstrated comparable discrimination (AUC 0.82).
• Semi-supervised and transfer learning did not enhance discrimination compared to models using only silver- or gold-standard data.
• Transfer learning achieved superior calibration, indicated by a high P value for the unreliability index and a low Brier score.

Conclusions:

• Machine learning models, including deep and non-deep learning approaches, perform similarly for infection identification.
• Semi-supervised and transfer learning models offer comparable discrimination to baseline methods like XGBoost.
• Transfer learning shows promise in improving the calibration of infection prediction models in hospitalized patients.