Table 4 Retrieval performance metrics across modalities and fusion methods
From: HONeYBEE: enabling scalable multimodal AI in oncology through foundation model-driven embeddings
Modality | Samples | P@5 | P@10 | P@20 | Clustering | Retrieval | Failure% |
|---|---|---|---|---|---|---|---|
| Â | Â | Â | Â | Â | AMI | AMI@10 | Â |
Individual modalities | |||||||
Clinical | 10,857 | 0.976 | 0.964 | 0.947 | 0.702 | 0.868 | 3.6 |
Molecular | 13,804 | 0.392 | 0.350 | 0.310 | 0.249 | 0.232 | 65.0 |
Radiology | 1149 | 0.373 | 0.343 | 0.310 | 0.228 | 0.272 | 66.1 |
Pathology reports | 10,857 | 0.746 | 0.703 | 0.651 | 0.341 | 0.585 | 29.7 |
Whole slide images | 8060 | 0.155 | 0.143 | 0.132 | 0.056 | 0.041 | 85.7 |
Multimodal fusion | |||||||
Concatenation | 11,341 | 0.504 | 0.461 | 0.418 | 0.347 | 0.408 | 53.9 |
Mean pooling | 11,341 | 0.492 | 0.446 | 0.403 | 0.336 | 0.382 | 55.4 |
Kronecker product | 11,341 | 0.284 | 0.269 | 0.253 | 0.320 | 0.320 | 73.1 |