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.

Metric	Random Forest	Deep Learning	p-value
Image-level results
Accuracy	0.871± 0.01	0.946 ± 0.01	< 0.001
Sensitivity	0.883 ± 0.02	0.968 ± 0.02	0.01
Specificity	0.860 ± 0.01	0.927 ± 0.01	0.01
Positive predictive value	0.847 ± 0.01	0.921 ± 0.01	< 0.001
AUC	0.935 ± 0.001	0.977 ± 0.01	< 0.001
Patient-level results
Accuracy	0.917 ± 0.01	0.962 ± 0.01	0.002
Sensitivity	0.932 ± 0.03	0.993 ± 0.01	0.033
Specificity	0.905 ± 0.03	0.935 ± 0.01	n.s.
Positive predictive value	0.896 ± 0.02	0.930 ± 0.01	n.s.
AUC	0.960 ± 0.01	0.989 ± 0.01	0.002

The results are presented as the Mean ± SD of three models. Each model was trained on ~770 images from ~70 patients. These models were evaluated on the held-out test set of 105 patients. The patient-level diagnosis is the majority vote over all the images from a single patient. Statistics were determined by an unpaired two-sample t-test with an N of three folds.