Chun-Nan Hsu

Background: Early prediction of in-hospital mortality in critically ill patients can aid clinicians in optimizing treatment. The objective was to develop a multimodal deep-learning model, using structured and unstructured clinical data, to predict in-hospital mortality risk among critically ill patients after their initial 24 h intensive care unit (ICU) admission. Methods: This study used data from Medical Information Mart for Intensive Care III (MIMIC-III), MIMIC-IV, the Electronic Intensive Care Unit Collaborative Research Database (eICU), and the high time-resolution ICU data set (HiRID). A multimodal model was developed on the MIMIC data sets, featuring time series components occurring within the first 24 h of ICU admission and predicting risk of subsequent inpatient mortality. Inputs included time-invariant variables, time-variant variables, clinical notes, and chest x-ray images. External validation occurred in a temporally separated MIMIC population, the high time-resolution ICU data set, and the eICU data set. Area under the receiver operating characteristics (AUROC), area under the precision–recall curves (AUPRC), and Brier scores were reported. Results: A total of 203,434 ICU admissions from more than 200 hospitals between 2001 to 2022 were included, in which mortality rate ranged from 5.2 to 7.9% across the four data sets. The model integrating structured data points had AUROC, AUPRC, and Brier scores of 0.92 (95% CI, 0.90 to 0.93), 0.53 (95% CI, 0.49 to 0.57), and 0.19 (95% CI, 0.18 to 0.20), respectively. The model was externally validated on eight different institutions within the eICU data set, demonstrating AUROCs ranging from 0.84 to 0.92. When including only patients with available clinical notes and imaging data, the AUROC, AUPRC, and Brier score improved from 0.87 (95% CI, 0.85 to 0.89) to 0.89 (95% CI, 0.87 to 0.91), 0.43 [(95% CI, 0.36 to 0.51) to 0.48 (95% CI, 0.40 to 0.56), and 0.37 (95% CI, 0.36 to 0.39) to 0.17 (95% CI, 0.16 to 0.19), respectively. The difference in AUROC was statistically significant based on a DeLong test ( P = 0.0167). Conclusions: These findings highlight the importance of incorporating multiple sources of patient information for mortality prediction and the importance of external validation.

High-quality metadata annotations for data hosted in large public repositories are essential for research reproducibility and for conducting fast, powerful and scalable meta-analyses. Currently, a majority of sequencing samples in the National Center for Biotechnology Information’s Sequence Read Archive (SRA) are missing metadata across several categories. In an effort to improve the metadata coverage of these samples, we leveraged almost 44 million attribute–value pairs from SRA BioSample to train a scalable, recurrent neural network that predicts missing metadata via named entity recognition (NER). The network was first trained to classify short text phrases according to 11 metadata categories and achieved an overall accuracy and area under the receiver operating characteristic curve of 85.2% and 0.977, respectively. We then applied our classifier to predict 11 metadata categories from the longer TITLE attribute of samples, evaluating performance on a set of samples withheld from model training. Prediction accuracies were high when extracting sample Genus/Species (94.85%), Condition/Disease (95.65%) and Strain (82.03%) from TITLEs, with lower accuracies and lack of predictions for other categories highlighting multiple issues with the current metadata annotations in BioSample. These results indicate the utility of recurrent neural networks for NER-based metadata prediction and the potential for models such as the one presented here to increase metadata coverage in BioSample while minimizing the need for manual curation. Database URL: https://github.com/cartercompbio/PredictMEE