A nursing note-aware deep neural network for predicting mortality risk after hospital discharge

Area of Science:

• Utilizes natural language processing (NLP) and deep learning for clinical data analysis.
• Focuses on extracting valuable insights from unstructured electronic nursing records.

Background:

• Post-intensive care unit (ICU) readmissions and mortality present significant healthcare challenges.
• Existing research often overlooks unstructured nursing data, limiting predictive accuracy.
• Natural language processing (NLP) offers a method to derive structured information from clinical text.

Purpose of the Study:

• To develop the Crucial Nursing Description Extractor (CNDE) for analyzing electronic nursing records.
• To predict post-ICU discharge mortality and identify patients at high risk for readmission.
• To implement a deep neural network (NurnaNet) combined with BioClinicalBERT for risk prediction.

Main Methods:

• Employed a cohort and system development design using the MIMIC-III database (2001-2012).
• Excluded patients under 18 or over 89, and those who died in-hospital; analyzed 16,973 nursing records.
• Developed CNDE with logarithmic likelihood ratio for keyword extraction, integrated with BioClinicalBERT for survival prediction (6 months, 2 years).

Main Results:

• NurnaNet achieved notable F1-scores of 0.67030 (6 months) and 0.70874 (2 years).
• Demonstrated performance improvements of 2.05% (6 months) and 1.08% (2 years) over BioClinicalBERT alone.
• Model performance was evaluated using precision, recall, F1-score, ROC curve, AUC, and PR curve.

Conclusions:

• The Crucial Nursing Description Extractor (CNDE) effectively processes long nursing records, extracting key information.
• NurnaNet demonstrates strong performance in analyzing nursing records to predict patient mortality risk post-discharge.
• Findings support timely adjustments to follow-up and treatment plans for high-risk patients.