Table 1.
Distribution of relation classes in the training and test datasets. Both actual numbers and percentages are shown. “PP None” indicates None relation between medical problems. “Tep None” indicates None relation between tests and medical problems. “TrP None” indicates None relation between treatments and medical problems.
| Relation Type | Training | Training % | Test | Test % | Effective Training* |
|---|---|---|---|---|---|
| PIP | 1239 | 38.4% | 1986 | 61.6% | 1123 |
| PP None | 7349 | 39.64% | 11190 | 60.36% | 4453 |
| TeCP | 303 | 34.0% | 588 | 66.0% | 271 |
| TeRP | 1734 | 36.4% | 3033 | 63.6% | 1564 |
| TeP None | 1535 | 38.50% | 2452 | 61.50% | 1379 |
| TrAP | 1423 | 36.4% | 2487 | 63.6% | 1284 |
| TrCP | 296 | 40.0% | 444 | 60.0% | 270 |
| TrIP | 107 | 35.1% | 198 | 64.9% | 100 |
| TrNAP | 106 | 35.7% | 191 | 64.3% | 101 |
| TrWP | 56 | 28.1% | 143 | 71.9% | 48 |
| TrP None | 2329 | 40.05% | 3486 | 59.95% | 2081 |
*
Effective Training denotes the number of samples used to train each class. It is less than the number of samples in the training dataset due to random allocation of 10% training dataset as validation set for all relations, and down-sampling for PP relations. We refer the reader to Experiments and Results section for more detail.