Table 3.

Classification task performance of the information extraction models: simple model, bagging model, and two partitioned bagging models. We adopted both micro- and macro-averaged F1 scores as a performance metric; micro F1 weighs on individual cases and macro F1 considers a balance of the classes.

	Site		Subsite		Laterality
	Micro F1	Macro F1	Micro F1	Macro F1	Micro F1	Macro F1
Single model
MT-CNN	0.9082	0.6312	0.6523	0.2483	0.8923	0.5087
MT-HCAN	0.9138	0.6533	0.6632	0.2614	0.8983	0.5006
Bagging model
MT-CNN	0.9127	0.6402	0.6681	0.2612	0.8980	0.5172
MT-HCAN	0.9168	0.6618	0.6724	0.2716	0.9015	0.5089
Combo	0.9177	0.6660	0.6774	0.2701	0.9022	0.5176
Partitioned bagging model A - abstention classifiers
MT-CNN	0.8992	0.6104	0.6690	0.2976	0.8982	0.5223
MT-HCAN	0.8992	0.6048	0.6595	0.2565	0.8967	0.5119
Combo	0.9034	0.6117	0.6708	0.2814	0.8992	0.5211
Partitioned bagging model B - additive preclassification
MT-CNN	0.9042	0.6301	0.6750	0.3098	0.9004	0.5349
MT-HCAN	0.9036	0.6385	0.6656	0.2677	0.8997	0.5239
Combo	0.9047	0.6317	0.6745	0.2946	0.9007	0.5326
	Histology		Behavior		Grade
Single model
MT-CNN	0.7645	0.2601	0.9753	0.8660	0.7599	0.6274
MT-HCAN	0.7684	0.2824	0.9765	0.8713	0.7637	0.6637
Bagging model
MT-CNN	0.7754	0.2724	0.9776	0.8684	0.7727	0.6390
MT-HCAN	0.7748	0.2934	0.9779	0.8692	0.7717	0.6726
Combo	0.7815	0.2925	0.9791	0.8807	0.7787	0.6677
Partitioned bagging model A - abstention classifiers
MT-CNN	0.7793	0.3651	0.9807	0.9096	0.7857	0.6384
MT-HCAN	0.7686	0.2935	0.9777	0.8872	0.7752	0.6491
Combo	0.7821	0.3474	0.9806	0.9129	0.7869	0.6488
Partitioned bagging model B - additive preclassification
MT-CNN	0.7817	0.3664	0.9809	0.9130	0.7864	0.6482
MT-HCAN	0.7704	0.2989	0.9777	0.8852	0.7752	0.6305
Combo	0.7828	0.3488	0.9806	0.9127	0.7870	0.6419