Fig. 8: Mortality prediction using MIMIC-III dataset.
From: Inferring multimodal latent topics from electronic health records
Each unsupervised embedding method was trained on the patients with only one admission in the MIMIC-III data. The trained model was then applied to embed the second-last admission from patients with at least two admissions that are within 6 months apart. An elastic net (EN) classifier was trained to predict the mortality outcome in the last admission. This was performed in a fivefold cross-validation setting. a Precision–recall curve (PRC) were generated, and the area under of the PRC (AUPRC) were displayed in the legend for each embedding method. EN represents the performance of elastic net using the raw EHR features. b Linear coefficients for the topics from MixEHR and LDA. The top three and bottom three topics are highlighted. c Topic clinical features for the top three most positively predictive topics and three most negatively predictive topics based on the elastic net coefficients for the 75 latent disease topics from MixEHR. d Same as c, but for the top predictive topics from LDA.
