Article Figures & Data
Figures
- FIG 1.
Frequencies of common findings. Bar lengths and numeric labels represent the frequency of each individual common finding across the labeled radiology reports. Each report may describe multiple findings or no findings at all.
- FIG 2.
Model coefficient heatmap. Normalized model coefficients corresponding to selected unigrams (horizontal axis labels) are shown for each model (vertical axis labels). Red values indicate a positive association between unigram frequency and the probability of the corresponding finding. One unigram was chosen from a list of 5 n-grams with the largest coefficients in each model. Coefficient labels are ordered by model of origin (“vasogenic” is the selected unigram for the edema model, “skull” is the selected unigram for the fracture model, and so forth). Preference was given to n-grams that were closely related to the finding of interest. N-grams that contained the finding word or words were not chosen for display, though these were typically the n-grams with the highest coefficients in each model. Coefficients were normalized by row and column averages for better comparison across models because coefficients are strongly affected by the n-gram frequency and model regularization terms.
- FIG 3.
Visualizing model discrimination by filtering uncertain predictions. Changes in the error rate are shown for higher-performing models (left) and lower-performing models (right) when data near the decision boundary are removed. The error rate is calculated as 1 minus the model accuracy.
Tables
Average model performance by classifier typea
Model LR SVM RF MLP AUROC 0.936 (0.910–0.959) 0.932 (0.907–0.954) 0.950 (0.931–0.967) 0.887 (0.848–0.922) AUPRC 0.695 (0.610–0.779) 0.650 (0.565–0.740) 0.710 (0.626–0.792) 0.613 (0.525–0.704) Sensitivity 0.748 (0.662–0.831) 0.679 (0.589–0.768) 0.728 (0.641–0.812) 0.660 (0.573–0.747) Specificity 0.950 (0.935–0.964) 0.944 (0.929–0.959) 0.955 (0.940–0.968) 0.938 (0.921–0.952) Accuracy 0.932 (0.916–0.947) 0.923 (0.906–0.938) 0.936 (0.921–0.951) 0.914 (0.896–0.930) Note:—LR indicates logistic regression; SVM, support vector machine; RF, random forest; MLP, multilayer perceptron.
↵a Average AUROC, AUPRC, sensitivity, specificity, and accuracy are shown for each model type across all findings. Numbers in parentheses are the range of values corresponding to the 95% confidence interval
Jump to section
Related Articles
Cited By...
- No citing articles found.