Mining clinical relationships from patient narratives

BMC Bioinformatics

Table 6 Performance by sentences.

		Number of sentence boundaries between arguments
		inter-	intra-	inter- and intra-sentential
Relation	Metric	1 ≤ n ≤ 5	n < 1	n ≤ 1	n ≤ 2	n ≤ 3	n ≤ 4	n ≤ 5
has_finding	P	24	68	65	62	60	61	61
	R	18	89	81	79	78	78	77
	F1	18	76	72	69	67	68	67
has_indication	P	18	49	42	42	36	32	30
	R	17	59	47	42	42	39	38
	F1	16	51	42	39	37	34	33
has_location	P	n/a	74	72	73	72	72	72
	R	n/a	83	81	81	81	82	82
	F1	n/a	77	75	76	75	76	76
has_target	P	3	64	62	59	60	59	58
	R	1	75	66	64	62	61	61
	F1	2	68	63	61	60	60	59
laterality_modifies	P	n/a	86	84	86	86	86	87
	R	n/a	89	88	88	88	87	88
	F1	n/a	85	84	85	86	85	86
negation_modifies	P	n/a	80	79	79	80	80	80
	R	n/a	94	93	91	93	93	93
	F1	n/a	86	85	84	85	86	85
sub_location_modifies	P	n/a	89	88	88	89	89	89
	R	n/a	95	95	95	95	95	95
	F1	n/a	91	91	91	91	91	91
Overall	P	22	69	65	64	62	61	60
	R	17	83	75	73	71	70	70
	F1	19	75	69	68	66	65	65

Variation in performance, by number of sentence boundaries (n) crossed by a relationship. For all cases, the cumulative feature set +event of Table 4 was used. For the inter-sentential-only classifier 1 ≤ n ≤ 5, the score fields for some relations are marked as n/a (not applicable). This is because some relations are either absent from the inter-sentential data (i.e. only ever appear intra-sententially), or are so rare that they do not appear in all training/test folds, and so a macro-average cannot be computed across the folds.

ISSN: 1471-2105