Clinical Text-Based Early Sepsis Risk Prediction: A Comparative Study of TF-IDF and ClinicalBERT

This study is conducted by the Master of Data Science program at Toronto Metropolitan University. Using clinical texts from the MIMIC-III intensive care database, it compares the performance of the traditional TF-IDF method and the pre-trained ClinicalBERT model in early sepsis prediction. Sepsis is one of the leading causes of death among ICU patients, and early identification is crucial. However, the information contained in clinical texts has not been fully utilized by traditional methods, so this study aims to explore a more effective prediction path.

Sepsis is a life-threatening organ dysfunction caused by infection and is the leading cause of death among ICU patients. Early identification can significantly improve prognosis, but traditional methods relying on physiological indicators and laboratory tests have lag issues. The massive unstructured clinical texts in ICUs (such as progress notes and nursing records) contain rich patient status information, yet they are difficult to be directly utilized by traditional scoring systems. Therefore, exploring clinical text-based prediction methods has important clinical value.

The study uses the public MIMIC-III intensive care database. The inclusion criteria are adult patients admitted to the ICU for the first time. Clinical text records within 24 hours after admission are extracted, and a sepsis case cohort is constructed based on ICD-9 diagnostic codes. This design ensures the clinical relevance of the study and simulates the risk assessment needs of doctors in the real scenario within the first 24 hours of a patient's ICU admission.

Traditional Machine Learning Baseline: TF-IDF Method Text features are extracted via TF-IDF, and models like XGBoost are trained as baselines. Its advantages are strong interpretability and high computational efficiency, but it cannot capture semantic relationships and context, making it difficult to handle synonyms, abbreviations, etc., in clinical texts.

Cutting-Edge Deep Learning: ClinicalBERT Fine-Tuning Based on the ClinicalBERT model pre-trained on massive clinical texts, fine-tuning helps understand professional terms, capture long-distance semantic dependencies, identify subtle differences in symptoms, and automatically learn relevant features.

Complete Workflow: Data inspection and cleaning → Cohort construction → Text extraction → Preprocessing (tokenization, stopword removal, etc.) → TF-IDF model training → ClinicalBERT fine-tuning → Model evaluation → Interpretability analysis.

Tech Stack: Python ecosystem tools, including Pandas (data processing), Scikit-learn (TF-IDF and traditional ML), XGBoost (gradient boosting), PyTorch (deep learning), Hugging Face Transformers (ClinicalBERT loading), and Jupyter Notebook (interactive development).

Clinical Value: If the model's accuracy is sufficient, it can be integrated into electronic medical record systems to analyze ICU patient records in real time, assist doctors in decision-making, optimize resource allocation, and support research on the pathogenesis of sepsis.

Future Outlook: From research prototype to clinical deployment, external validation, regulatory approval, ethical review, and other steps are required, but this study lays a methodological foundation for subsequent work.

This master's research project demonstrates the application of cutting-edge NLP technology in clinical problems. Through rigorous workflows, systematic model comparisons, and clinically relevant evaluations, it contributes solid work to the field of medical AI and serves as a complete project example worth referencing for medical AI researchers.