Comparison of Multimodal Deep Learning Approaches for Predicting Clinical Deterioration in Ward Patients:...

Area of Science:

• Clinical informatics
• Biomedical data science
• Healthcare artificial intelligence

Background:

• Machine learning models can decrease patient morbidity and mortality by identifying clinical deterioration.
• Current models often have high false positive rates and rely solely on structured data.

Purpose of the Study:

• To compare the predictive performance of models using structured data alone versus those incorporating unstructured data from clinical notes.
• To evaluate different methods for parameterizing concept unique identifiers (CUIs) from clinical notes.

Main Methods:

• Adult patients from two cohorts (University of Chicago and University of Wisconsin-Madison) were analyzed.
• Predictors included structured variables and unstructured variables from clinical notes, parameterized as CUIs using various methods (ST, ICDR-T, ICDR-BV, SE, CC).
• Deep recurrent neural networks were used to compare models for predicting intensive care unit transfer or death within 24 hours.

Main Results:

• The SapBERT embedding (SE) model showed the highest area under the precision-recall curve (0.208).
• The concept unique identifier clustering (CC) and structured-only models achieved the highest area under the receiver operating characteristic (0.870).
• Models demonstrated good calibration, and key CUIs like 'NPO - nothing by mouth' and 'chemotherapy' were highly predictive of deterioration.

Conclusions:

• While incorporating CUIs from clinical notes did not substantially enhance predictive performance for clinical deterioration, it offered valuable clinical context.
• Multimodal models combining structured data with SapBERT embeddings showed promise.
• Models utilizing CUIs can augment clinical decision-making by providing additional relevant information.