This study develops an advanced multimodal AI framework to strengthen early risk assessment in critical care and support resilient healthcare delivery. Utilizing the MIMIC-III database, this research extracted structured variables and clinical notes from 26,829 adult patients. A text mining approach based on the BERTopic model was employed to generate topic embeddings from unstructured notes, which were subsequently integrated with 16 quantitative variables. Six machine learning models: Adaboost, Gradient Boosting, Support Vector Classification (SVC), Bagging, Logistic Regression, and MLP Classifier were trained to predict short-term and long-term mortality outcomes. Model performance was evaluated through AUROC, accuracy, recall, precision, and F1-score metrics. The results demonstrate that integrating topic embeddings with structured data significantly improved short-term risk prediction. The SVC model, in particular, achieved an AUROC of 0.9137 for predicting 2-day mortality. Critical predictors identified included the Glasgow Coma Scale, White Blood Cell Count, and text-derived topics related to cardiovascular and neurological conditions. The study is based on a single-center dataset, limiting generalizability. Additionally, only a subset of textual data sources was analyzed, and improvements in long-term risk prediction were relatively modest. These findings demonstrate how multimodal AI can significantly improve early risk assessment and enhance resilience in critical care decision-making. This research pioneers the integration of BERTopic-based text mining with machine learning models for clinical risk prediction, highlighting the value of multimodal data fusion in improving predictive accuracy and enriching medical informatics.
Wu et al. (Thu,) studied this question.