Table 2.
Clustering scores by embedding method
| Model | Embedding | Embedding dimension | NMI (200) | NMI (400) | AMI (200) | AMI (400) | Silhouette |
|---|---|---|---|---|---|---|---|
| Med2Vec | Sep | 200 | 0.2011 | 0.2400 | 0.0498 | 0.0480 | −0.5822 |
| CBOW | Sep | 10 | 0.4254 | 0.4774 | 0.1493 | 0.1408 | −0.4647 |
| CBOW | Sep | 50 | 0.4261 | 0.4635 | 0.1776 | 0.1679 | −0.3897 |
| CBOW | Sep | 100 | 0.4011 | 0.4335 | 0.1702 | 0.1622 | −0.3884 |
| SG | Sep | 10 | 0.5221 | 0.5737 | 0.2052 | 0.1837 | −0.3161 |
| SG | Sep | 50 | 0.5500 | 0.5905 | 0.2754 | 0.2572 | −0.1981 |
| SG | Sep | 100 | 0.5288 | 0.5751 | 0.2852 | 0.2797 | −0.2001 |
| CBOW | Co | 10 | 0.4313 | 0.4773 | 0.1590 | 0.1429 | −0.4639 |
| CBOW | Co | 50 | 0.4576 | 0.4935 | 0.2287 | 0.2133 | −0.3549 |
| CBOW | Co | 100 | 0.4478 | 0.4825 | 0.2323 | 0.2154 | −0.3448 |
| SG | Co | 10 | 0.5220 | 0.5798 | 0.2035 | 0.1913 | −0.3197 |
| SG | Co | 50 | 0.5726 | 0.6144 | 0.2979 | 0.2864 | −0.1648 |
| SG | Co | 100 | 0.5605 | 0.6134 | 0.2963 | 0.3107 | −0.1615 |
| Med2Vec* | N/A | 200 | 0.2755 | 0.3472 | 0.0524 | 0.0437 | −0.5001 |
| SG* | Co | 100 | 0.5843 | 0.6448 | 0.2559 | 0.2598 | −0.1722 |
*
Models trained on a subsample of codes which occurred in the translated Med2Vec comparison.
Note that the “Co” designation in the embedding column indicates a model which trained category and code embeddings jointly, whereas a “Sep” designation indicates that these embeddings were trained separately.