Multimodal temporal-clinical note network for mortality prediction.

Area of Science:

• Artificial Intelligence in Medicine
• Healthcare Informatics
• Machine Learning for Healthcare

Background:

• Mortality prediction is crucial for intensive care unit (ICU) management and personalized patient care.
• Electronic medical records (EMRs) contain rich clinical notes, often underutilized in mortality prediction models.
• Existing methods primarily focus on temporal events, neglecting valuable historical and report information within clinical notes.

Purpose of the Study:

• To develop a multimodal deep learning network integrating temporal data and clinical notes for improved mortality prediction.
• To enhance the utilization of diverse medical information present in clinical notes, including patient history and reports.
• To differentiate chronic illness patients based on historical clinical note information.

Main Methods:

• A multimodal temporal-clinical note network was proposed, utilizing Long Short-Term Memory (LSTM) networks for time series embedding.
• Clinical notes were processed using a label-aware convolutional neural network (CNN) for text embedding.
• A scoring function was developed to assess the importance of different clinical note sections.

Main Results:

• The proposed multimodal network achieved superior performance in mortality prediction, indicated by improved AUCPR and AUCROC scores compared to existing methods.
• Visualization techniques highlighted the importance of specific words within clinical notes, demonstrating improved model interpretability.
• The study confirmed that incorporating medical history from clinical notes significantly enhances mortality prediction accuracy.

Conclusions:

• The multimodal temporal-clinical note network effectively integrates diverse data sources for superior mortality prediction.
• The inclusion of medical history and reports from clinical notes is vital for improving prediction performance.
• Label-aware CNNs offer a promising approach for extracting relevant information from clinical text, further boosting prediction accuracy.