Sepsis prediction, early detection, and identification using clinical text for machine learning: a systematic review

Area of Science:

• Medical Informatics
• Computational Biology
• Clinical Data Science

Background:

• Sepsis remains a critical challenge in healthcare, necessitating advanced methods for early detection and prediction.
• Machine learning (ML) offers potential for analyzing complex patient data to improve sepsis outcomes.
• Unstructured clinical text contains valuable information often missed by traditional structured data analysis.

Purpose of the Study:

• To evaluate the impact of incorporating unstructured clinical text into ML models for sepsis prediction.
• To compare the performance of ML models using combined text and structured data versus structured data alone.
• To identify effective ML and Natural Language Processing (NLP) techniques for sepsis detection.

Main Methods:

• A systematic literature search was conducted across major scientific databases (PubMed, Scopus, ACM DL, dblp, IEEE Xplore).
• Studies using ML or NLP on clinical text for sepsis detection, identification, or prediction were included.
• Extracted data included sepsis definitions, datasets, data types, ML models, NLP techniques, and evaluation metrics.

Main Results:

• Machine learning models combining unstructured clinical text (e.g., narrative notes) with structured data (demographics, vitals, labs) demonstrated superior performance.
• The combined approach led to earlier and more accurate sepsis prediction compared to models using only structured data.
• Area Under the Curve (AUC) analysis indicated improved predictive accuracy with the integration of text data.

Conclusions:

• Utilizing both unstructured clinical text and structured data in ML models enhances sepsis identification and early detection.
• While approaches are heterogeneous, the synergistic effect of multi-modal data is evident.
• Further research is needed to address limitations such as data heterogeneity and transferability of models across different clinical settings.