Improving rare disease classification using imperfect knowledge graph

Table 2 Rare disease classification performance on HaoDaiFu corpus

Percentage Bins	(0, 0.02%]		(0.02%, 0.05%]		(0.05%, 0.1%]		(0.1%, 0.5%]		(0.5%, 1%]
	89 diseases		277 diseases		205 diseases		194 diseases		32 diseases
	F₁	MRR	F₁	MRR	F₁	MRR	F₁	MRR	F₁	MRR
BOW	34.10	45.86	40.80	49.91	49.48	58.81	53.23	62.80	62.23	75.31
LSTM	0.00	0.41	0.01	1.07	0.38	5.91	12.29	27.23	40.07	53.04
UpSample	35.17 ^∗	47.10 ^∗	40.69	50.43 ^∗	47.63	57.63	49.85	59.75	58.6	68.95
χ²	34.04	46.75 ^∗	40.81 ^∗	50.66 ^∗	49.15	58.53	51.74	61.38	61.55	74.05
BOW+ χ²	34.56	47.25 ^∗	42.41	51.84 ^∗	50.03 ^∗	59.33 ^∗	53.15	62.34	62.10	73.97
KG₁	33.66	44.98	38.25	47.45	45.17	53.97	48.07	57.55	59.21	71.29
KG₁₂	33.51	44.92	39.08	48.07	45.23	54.55	48.66	58.00	59.2	71.43
BOW+KG\(^{\text {pseudo-doc}}_{1}\)	31.91	42.81	37.51	46.08	44.08	53.22	47.01	56.94	55.91	69.47
BOW+KG\(^{\text {pseudo-count}}_{1}\)	34.87 ^∗	46.14 ^∗	41.74 ^∗	50.14 ^∗	49.31	57.94	52.56	61.59	61.65	74.19
BOW+KG\(^{\text {late-fusion}}_{1}\)	33.33	45.42	38.41	48.68	47.15	56.39	51.13	60.18	61.42	73.30
BOW+KG\(^{\text {early-fusion}}_{1}\)	36.87	48.36 ^∗	43.11 ^∗	51.79 ^∗	50.06 ^∗	58.99	52.86	61.90	61.90	73.57
BOW+KG\(^{\text {early-fusion}}_{12}\)	36.94 ^∗	48.22 ^∗	42.63 ^∗	51.40 ^∗	49.66	58.62	52.60	61.51	61.47	73.23

The higher F₁ and MMR, the better. Each column’s highest number is shown in boldface, second highest number shown with underline. The left three percentage bins are rare disease bins; the right two bins are for comparison purposes. “ ^∗” denotes results significantly higher than BOW (randomization test, significance level α=0.05)

ISSN: 1472-6947