Fig 1.
Proposed framework for short-term, mid-term and long-term mortality prediction tasks.
Fig 2.
Cohort selection.
Fig 3.
Problem formulation using history window and prediction window.
TA, TP, and TF represent the time of admission, end of history window, and end of prediction window, respectively. The folded-corner boxes within the history window represent clinical notes from different categories such as ECG, Echo, Radiology, etc. Each patient has a variable number of clinical notes within the history window. The dark grey band within the prediction window represents the label of the patient. Label 1 means the patient is still alive, and label 0 means the patient has passed away.
Fig 4.
Distribution of sequence length of 33,740 concatenated clinical notes after tokenization using PubMedBERT model.
The notes were taken within 24 hours after admission.
Fig 5.
Illustration of feature extraction from one note using the PubMedBERT model.
Here B refers to the number of blocks, ti refers to the i-th token, and T refers to the total number of tokens.
Table 1.
Class distribution for short/mid/long-term prediction window.
Table 2.
Hyperparameter optimization for classification models.
Table 3.
Test AU-ROC scores by four models trained with features extracted using TF-IDF.
Table 4.
Test AU-ROC scores for four models trained with features extracted using FastText.
Table 5.
Test AU-ROC scores for four models trained with features extracted using PubMedBERT.
Fig 6.
Test ROC curve, AU-ROC score, sensitivity and specificity scores for logistic regression model with TF-IDF, FastText, and PubMedBERT.
The optimum threshold value for sensitivity and specificity scores has been calculated using Youden’s Index. The x-axis represents the prediction window. The grey boxes show the number of deceased and alive patients with respect to the prediction windows.
Fig 7.
Top 10 most important features that are predictive of the mortality prediction outcome—Alive (green bars), deceased (red bars), by logistic regression model with TF-IDF.
Fig 8.
Importance of the features that remained significant over short/mid/long-term prediction windows.
These features are from the logistic regression with TF-IDF.