How can artificial intelligence (AI) and machine learning (ML) be applied to detect neonatal sepsis?

AI and Machine Learning in Neonatal Sepsis Detection

AI and machine learning should be used as clinical decision-support tools to supplement—not replace—clinician judgment in detecting neonatal sepsis, with models integrating real-time vital signs, laboratory biomarkers, and perinatal risk factors to enable early prediction 6-24 hours before clinical deterioration. 1

How AI/ML Models Work for Sepsis Detection

Machine learning models analyze patterns in clinical data that human clinicians may miss. Sepsis has been a major focus of ML evaluations because it carries high mortality and early detection improves outcomes 1. The models require:

• Real-time access to vital sign data (heart rate, respiratory rate, oxygen saturation)
• Laboratory values (C-reactive protein, white blood cell counts, procalcitonin, lactate)
• Perinatal and demographic risk factors
• Rapid alert systems to contact clinicians when sepsis risk increases 1

Specific Model Performance

For early-onset sepsis (EOS), ensemble tree-based classifiers (CatBoost, XGBoost) achieve AUROCs exceeding 0.95 in retrospective studies 2. A validated model from low-middle income countries demonstrated 90.3% sensitivity and could reduce antibiotic usage by 19.1% while maintaining 100% sensitivity in asymptomatic neonates with only perinatal risk factors 3.

For late-onset sepsis (LOS), continuous physiological monitoring models achieve AUROCs of 0.81-0.90, with risk increasing 150-fold up to 10 hours before clinical suspicion 4. The HeRO (Heart Rate Observation) monitor produced a 22% relative reduction in mortality in a prospective randomized trial 2.

Biomarkers (CRP, white blood cells) are superior to clinical signs and risk factors alone for prediction, with C-reactive protein being most important 5. Combining multiple vital signs improves performance compared to heart rate characteristics alone 4.

Critical Implementation Requirements

External Validation is Essential

The most critical need is external validation—models developed in one institution reflect that site's idiosyncrasies and perform worse in new settings 1. Currently, most models lack this validation 1. Performance must be evaluated:

• Across different hospitals and patient populations
• Within patient subgroups (models performing well on average may fail in important subpopulations) 1
• Over time as practices change (algorithm performance degrades when clinical workflows shift) 1

Common Pitfalls to Avoid

Data leakage is a critical error that artificially inflates performance 6:

• Illegitimate features: Epic's sepsis tool inappropriately used "antibiotics" as a feature to predict sepsis—but this information isn't available prospectively 6
• Temporal leakage: A lactate order (rather than the value) may predict sepsis only because it's ordered when sepsis is suspected; if ordering practices change, the model fails 1
• Train-test contamination: Samples from the same patient appearing in both training and test sets leads to overoptimistic results 6

Uncertainty Communication

Models must communicate when predictions are uncertain and suppress alerts when confidence is low 1. This increases perceived reliability and trustworthiness. Bayesian inference provides uncertainty estimates to distinguish:

• Model uncertainty (insufficient exposure to similar cases)
• Data uncertainty (noisy target outcomes) 1

Current Limitations and Gaps

Most existing models are retrospective, single-center studies from high-income countries, limiting generalizability to low-middle income countries where sepsis burden is highest 7, 2. Additional barriers include:

• Lack of uniform sepsis definitions across studies
• Variations in disease and pathogen profiles between settings
• Limited availability of electronic health data in resource-limited settings
• Insufficient prospective validation trials 7, 2

The marketplace is full of companies promising results that are difficult to assess due to lack of validation 1.

Practical Clinical Algorithm

1. Use AI models as screening tools in conjunction with clinical assessment—not as standalone diagnostic tools 1

2. Implement a two-step approach:

   • Screen all at-risk neonates with clinical risk stratification
   • Apply ML models to preselected high-risk populations to refine antibiotic decisions 5

3. Prioritize models that integrate:

   • Continuous vital sign monitoring (heart rate variability, respiratory patterns, oxygen saturation)
   • Laboratory biomarkers (CRP, WBC as minimum)
   • Perinatal risk factors 3, 4, 5

4. Require external validation data before institutional adoption—demand evidence the model performs in settings similar to yours 1

5. Monitor for algorithm drift: Establish protocols to reassess performance as your clinical practices evolve 1

6. Ensure explainability: Understand which features drive predictions to identify potential biases or illegitimate associations 6, 7