Towards unstructured mortality prediction with free-text clinical notes.

Area of Science:

• Clinical informatics
• Biomedical data science
• Healthcare analytics

Background:

• Vast amounts of healthcare data are generated, but primarily structured data is used for clinical outcome prediction.
• Unstructured clinical notes contain valuable subjective information that is often underutilized for mortality prediction.
• Effective utilization of diverse healthcare data is crucial for advancing predictive modeling in medicine.

Purpose of the Study:

• To evaluate the performance enhancement in mortality prediction using multiple, minimally preprocessed unstructured clinical notes.
• To assess the utility of subjective information from clinical notes in improving predictive discrimination.
• To compare the efficacy of unstructured data-based prediction against traditional structured data approaches.

Main Methods:

• Development of a hierarchical model incorporating both convolutional and recurrent neural network layers.
• Concurrent modeling of multiple clinical notes within an individual hospital stay.
• Evaluation on the Medical Information Mart for Intensive Care III (MIMIC-III) dataset for in-hospital mortality prediction.

Main Results:

• The proposed model achieved higher predictive performance metrics compared to methods relying solely on structured data.
• The approach demonstrated effectiveness even with minimal preprocessing of unstructured clinical notes.
• Significant gains in discrimination were observed by incorporating information from multiple notes.

Conclusions:

• Minimally preprocessed unstructured clinical notes can substantially enhance in-hospital mortality prediction.
• Incorporating raw, unstructured healthcare data into clinical prediction models is essential for improved performance.
• This study underscores the potential of leveraging rich textual data for more accurate clinical outcome forecasting.