Table 1.
Patient demographics for the heart failure data set.
Fig 1.
Schematic overview of digital pathology workflow to detect heart failure.
(a) Patients were divided into a training and test dataset. WSI were scanned and regions of interest (ROI) were extracted for image analysis. All ROIs from the same patient were given the same label, which was determined by whether the patient had clinical or pathological evidence of heart disease. (b) Three-fold cross validation was used to train heart failure classifiers using a deep learning model or engineered features in WND-CHARM + a random decision forest classifier. (c) Trained models were evaluated at the image and patient-level on a held-out test dataset.
Table 2.
Image and patient-level performance evaluation for predicting clinical heart failure from H&E stained whole-slide images for validation folds of the training data set.
Fig 2.
Review of classified images and unsupervised clustering of patients.
Representative images from ROIs that were: correctly classified by both algorithms (a); errors unique to the Random Decision Forest (b); and errors unique to Deep Learning (c). The ground truth is shown in white text in the upper left corner of the image whereas the algorithm prediction is shown in the upper right and color coded with green = correct and yellow = incorrect. (d) ROIs from two patients without clinical heart failure that were classified with clinically “Failing” patients by both algorithms. These patients were later found to have evidence of tissue-level pathology by two independent pathologists. (e) Consensus clustering using WND-CHARM feature vector reveals evidence for three clusters in the data. The consensus clustering dendrogram and class results are shown above the clustergram. Some patients form a small third cluster between the two larger groups, marked by an arrowhead, which was found to contain a patient with tissue pathology but no clinical heart disease.
Table 3.
Patient-level performance evaluation for predicting clinical heart failure or severe tissue pathology from H&E stained whole-slide images for validation folds of the training data set.
Table 4.
Patient-level performance evaluation for predicting clinical heart failure or severe tissue pathology from H&E stained whole-slide images for the held-out test set.
Fig 3.
Receiver Operator Characteristic (ROC) curve for detection of clinical heart failure or severe tissue pathology.
(a) ROC curve for image-level detection on the training dataset (DL vs. RF, p < 0.0001, two-sample Kolmogorov-Smirnov (KS) test). (b) ROC curve for patient-level detection on the training dataset (DL vs. RF, ns, KS test). (c) ROC curve for image-level detection on the held-out test dataset (DL vs. RF, p < 0.0001, KS test). (d) ROC curve for patient-level detection on the held-out test dataset (DL vs. RF, ns, KS test).