Table 2.

(a) Tile-classifier metrics for the task of classifying foci of interest on H&E slides, and slide-level classifier metrics for the IHC request prediction task, for the three cross-validation splits of the dataset. (b) Slide-level classifier results for the task of IHC requesting evaluated on the validation set, for the three annotating pathologists. (c) Estimation of time savings and extra costs from introduction of the model (n = 380 IHC-requested cases).

(a)
Split number	Tile classification accuracy	Tile classification AUC	IHC order accuracy	IHC order AUC
0	0.86	0.91	0.99	0.99
1	0.85	0.94	0.99	0.99
2	0.91	0.93	0.99	0.99

(b)
Pathologist	Validation accuracy	Validation AUC
1	0.979	0.977
2	0.764	0.737
3	0.674	0.681

(c)
Operating point (n = 380)	Specificity	Average false positive rate	Turnaround time savings (days)	Reporting time savings (hours)	Extra cost of unnecessary IHC orders (£)
Reflex testing	NA	NA	1170	70	4180
1	0.6	0.15	703	42	627
2	0.75	0.33	879	52	1379
3	0.9	0.48	1054	63	2006

The false positive rate is averaged over the three pathologists. The minimum expected savings of 3 days 2 h for turnaround time and 11 min for reporting times were used in the calculations. We assume an IHC order cost of £11 per slide.