Multitask learning from clinical text and acute physiological conditions differentially improve the prediction of mortality and diagnosis at the ICU

This is the authors' abstract. We don't add key points for this paper.

The prediction of mortality of critically ill patients has stimulated the development of many severity scoring algorithms. The majority of the models use physiological measurements obtained during the first hours of admission (i.e., heart rate, arterial blood pressure, or respiratory rate). In this study, we propose to improve the performance of current scoring system by including free text from patient’s medical history. Although the primary outcome was in-hospital mortality, we chose a model architecture to provide simultaneous assessment of ICD-9 codes and groupings. We hypothesized that including patients’ medical history with a multitask learning approach would improve model performance. We compared the predictive performance obtained with our approach to the best models previously proposed in the literature (baseline models). We used the MIMIC publicly available database which includes > 60,000 ICU admissions between 2001 and 2012. The patients’ condition at admission was accounted for by the preliminary diagnosis at admission and the medical history extracted from the discharge summaries notes. Unstructured data was processed through a Gated Recurrent Units layer with pre-trained word embeddings, and the hidden states were concatenated to the remaining structured-tabular data. Baseline models achieved similar results than in previously published work, but our artificial neural networks models showed significant improvement towards classification of mortality (AUC-ROC = 0.90). Including the medical history improved all tasks but relatively more the ICD-9 codes prediction than the mortality. The clinical prediction model presented here could be used to identify patients’ risk groups, which would improve the quality of ICU care, and further help to efficiently allocate hospital resources.