ICD Coding on MIMIC-III 50 labels (test)
64.1F1 MicroHLAN
Evaluation Results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| HLANLabel embedding initialization=false2020.10 | 64.1 | 88.4 | 91.9 | 57.1 | 62.5 | — | — | 65 | 51 | 72.9 | 57.3 | |
| HLANLabel embedding initialization=true2020.10 | 64 | 88.4 | 91.9 | 56.8 | 62.4 | — | — | 65.5 | 50.2 | 73.2 | 56.9 | |
| CNNLabel embedding initialization=true2020.10 | 63.4 | 88.3 | 91.3 | 59.1 | 62.1 | — | — | 53 | 66.7 | 57.7 | 70.4 | |
| CNNLabel embedding initialization=false2020.10 | 62.4 | 88.1 | 90.9 | 58.4 | 61.8 | — | — | 51.5 | 67.4 | 55.6 | 71.2 | |
| HLANSentence splitting=true2020.10 | 61.4 | 86.9 | 90.4 | 54.5 | 60.2 | — | — | 63.6 | 47.8 | 71.5 | 53.8 | |
| CNN+attLabel embedding initialization=false2020.10 | 60.7 | 88.1 | 91.1 | 54.8 | 60.8 | — | — | 63.1 | 48.4 | 70.9 | 53.1 | |
| CNN+attLabel embedding initialization=true2020.10 | 60.6 | 88.3 | 91.3 | 53.6 | 61.6 | — | — | 64.3 | 46 | 71.6 | 52.5 | |
| HANLabel embedding initialization=true2020.10 | 60.1 | 87.3 | 90.3 | 53 | 59.9 | — | — | 61.3 | 46.9 | 67.9 | 54.2 | |
| HANLabel embedding initialization=false2020.10 | 59.4 | 87 | 90.1 | 52.8 | 59.5 | — | — | 61.7 | 46.3 | 68.2 | 52.9 | |
| HA-GRULabel embedding initialization=true2020.10 | 59.1 | 86.4 | 90.1 | 51.7 | 59.5 | — | — | 62.1 | 44.3 | 71.1 | 50.7 | |
| HA-GRULabel embedding initialization=false2020.10 | 57.2 | 85.3 | 89.2 | 49.9 | 57.9 | — | — | 59.3 | 43.1 | 69.5 | 48.7 | |
| Bi-GRULabel embedding initialization=true2020.10 | 52.5 | 80.9 | 85.8 | 42.8 | 52.1 | — | — | 47.3 | 39.2 | 57.5 | 48.4 | |
| Bi-GRULabel embedding initialization=false2020.10 | 51.2 | 80.6 | 85.5 | 41.2 | 51.3 | — | — | 47.2 | 36.7 | 58.1 | 45.8 | |
| GRPOBackbone=Qwen3-4B, Paradigm=LLM (Post-training)2026.06 | 0.736 | — | — | 0.681 | — | — | — | 72.2 | 66.5 | 75.6 | 71.7 | |
| GKI-ICD2025.05 | 0.735 | 0.933 | 0.952 | 0.692 | 0.681 | — | — | — | — | — | — | |
| PHIBackbone=Qwen3-4B, Paradigm=LLM (Post-training)2026.06 | 0.735 | — | — | 0.686 | — | — | — | 70.4 | 68.7 | 73.3 | 73.7 | |
| AKILextra_annotations=DRG codes, CPT codes and medications2025.05 | 0.734 | 0.928 | 0.95 | 0.692 | 0.683 | — | — | — | — | — | — | |
| MRRextra_annotations=DRG codes, CPT codes and medications2025.05 | 0.732 | 0.927 | 0.947 | 0.687 | 0.685 | — | — | — | — | — | — | |
| CoRelation2025.05 | 0.731 | 0.933 | 0.951 | 0.693 | 0.683 | — | — | — | — | — | — | |
| KEPTLongformerBest epoch=42022.10 | 0.7285 | 0.9263 | 0.9476 | 0.6891 | 0.6726 | — | — | — | — | — | — | |
| KEPTLongformer2025.05 | 0.728 | 0.926 | 0.947 | 0.689 | 0.672 | — | — | — | — | — | — | |
| SFTBackbone=Qwen3-4B, Paradigm=LLM (Post-training)2026.06 | 0.726 | — | — | 0.683 | — | — | — | 67.9 | 70.3 | 70.1 | 75.2 | |
| MSMN2022.03 | 0.725 | 0.928 | 0.947 | 0.683 | 0.68 | — | — | — | — | — | — | |
| MSMN2025.05 | 0.725 | 0.928 | 0.947 | 0.683 | 0.68 | — | — | — | — | — | — | |
| KEPTLongformer (w/o HSAP)Best epoch=5, variant=without HSAP2022.10 | 0.7192 | 0.9233 | 0.9431 | 0.6795 | 0.6718 | — | — | — | — | — | — | |
| PLM-ICD2025.05 | 0.719 | 0.91 | 0.934 | 0.663 | 0.66 | — | — | — | — | — | — | |
| MSMNBest epoch=152022.10 | 0.7178 | 0.925 | 0.9439 | 0.6764 | 0.6723 | — | — | — | — | — | — | |
| JointLAAT2020.07 | 0.716 | 0.925 | 0.946 | 0.661 | 0.671 | 54.6 | 35.7 | — | — | — | — | |
| JointLAAT2022.03 | 0.716 | 0.925 | 0.946 | 0.661 | 0.671 | — | — | — | — | — | — | |
| LAAT2020.07 | 0.715 | 0.925 | 0.946 | 0.666 | 0.675 | 54.7 | 35.7 | — | — | — | — | |
| LAAT2022.03 | 0.715 | 0.925 | 0.946 | 0.666 | 0.675 | — | — | — | — | — | — | |
| PLM-CA2025.05 | 0.71 | 0.916 | 0.936 | 0.671 | 0.664 | — | — | — | — | — | — | |
| JointLAATBest epoch=102022.10 | 0.7084 | 0.9236 | 0.9424 | 0.6695 | 0.6636 | — | — | — | — | — | — | |
| GRPOBackbone=Qwen2.5-1.5B, Paradigm=LLM (Post-training)2026.06 | 0.707 | — | — | 0.646 | — | — | — | 67.8 | 64.1 | 71.9 | 69.6 | |
| PHIBackbone=Qwen2.5-1.5B, Paradigm=LLM (Post-training)2026.06 | 0.706 | — | — | 0.653 | — | — | — | 66.3 | 67 | 68.9 | 72.3 | |
| SFTBackbone=Qwen2.5-1.5B, Paradigm=LLM (Post-training)2026.06 | 0.694 | — | — | 0.652 | — | — | — | 64.1 | 68.8 | 65.4 | 73.8 | |
| MSATT-KG2020.07 | 0.684 | 0.914 | 0.936 | 0.638 | 0.644 | — | — | — | — | — | — | |
| MSATT-KG2022.03 | 0.684 | 0.914 | 0.936 | 0.638 | 0.644 | — | — | — | — | — | — | |
| MSATT-KGsource=collected from original paper2022.10 | 0.684 | 0.914 | 0.936 | 0.638 | 0.644 | — | — | — | — | — | — | |
| MSATT-KG2025.05 | 0.684 | 0.914 | 0.936 | 0.638 | 0.644 | — | — | — | — | — | — | |
| PLM-ICDBackbone=RoBERTa, Paradigm=PLM Baselines2026.06 | 0.681 | — | — | 0.635 | — | — | — | 63.7 | 63.3 | 68 | 68.1 | |
| KEPTLongformer (w/o HSAP & Prompt)Best epoch=17, variant=without HSAP & Prompt2022.10 | 0.6722 | 0.9054 | 0.9318 | 0.5861 | 0.6438 | — | — | — | — | — | — | |
| MultiResCNN2019.11 | 0.67 | 0.899 | 0.928 | 0.606 | 0.641 | — | — | — | — | — | — | |
| MultiResCNN2020.07 | 0.67 | 0.899 | 0.928 | 0.606 | 0.641 | — | — | — | — | — | — | |
| MultiResCNN2022.03 | 0.67 | 0.899 | 0.928 | 0.606 | 0.641 | — | — | — | — | — | — | |
| HyperCore2022.03 | 0.663 | 0.895 | 0.929 | 0.609 | 0.632 | — | — | — | — | — | — | |
| MultiResCNNBest epoch=182022.10 | 0.6624 | 0.893 | 0.9204 | 0.5929 | 0.6156 | — | — | — | — | — | — | |
| LAATBackbone=Bi-LSTM, Paradigm=PLM Baselines2026.06 | 0.65 | — | — | 0.56 | — | — | — | 64.3 | 53 | 71.4 | 59.7 | |
| MultiResCNNBackbone=ResNet, Paradigm=PLM Baselines2026.06 | 0.639 | — | — | 0.562 | — | — | — | 57.7 | 58.1 | 65.6 | 62.3 | |
| DR-CAML2018.02 | 0.633 | 88.4 | 0.916 | 0.576 | 61.8 | — | — | — | — | — | — | |
| DR-CAML2019.11 | 0.633 | 0.884 | 0.916 | 0.576 | 0.618 | — | — | — | — | — | — | |
| DR-CAML2020.07 | 0.633 | 0.884 | 0.916 | 0.576 | 0.618 | — | — | — | — | — | — | |
| CNN2018.02 | 0.625 | 87.6 | 0.907 | 0.576 | 62 | — | — | — | — | — | — | |
| CNN2020.07 | 0.625 | 0.876 | 0.907 | 0.576 | 0.62 | — | — | — | — | — | — | |
| CNNBackbone=CNN, Paradigm=PLM Baselines2026.06 | 0.625 | — | — | 0.526 | — | — | — | 57.4 | 52.4 | 65.6 | 59.7 | |
| LEAM2020.07 | 0.619 | 0.881 | 0.912 | 0.54 | 0.612 | — | — | — | — | — | — | |
| CAML2018.02 | 0.614 | 87.5 | 0.909 | 0.532 | 60.9 | — | — | — | — | — | — | |
| CAML2019.11 | 0.614 | 0.875 | 0.909 | 0.532 | 0.609 | — | — | — | — | — | — | |
| CAML2020.07 | 0.614 | 0.875 | 0.909 | 0.532 | 0.609 | — | — | — | — | — | — | |
| CAML2022.03 | 0.614 | 0.875 | 0.909 | 0.532 | 0.609 | — | — | — | — | — | — | |
| CAML2025.05 | 0.614 | 0.875 | 0.909 | 0.532 | 0.609 | — | — | — | — | — | — | |
| Bi-GRU2018.02 | 0.549 | 82.8 | 0.868 | 0.484 | 59.1 | — | — | — | — | — | — | |
| BIGRU2020.07 | 0.549 | 0.828 | 0.868 | 0.484 | 0.591 | — | — | — | — | — | — | |
| CAMLBackbone=CNN, Paradigm=PLM Baselines2026.06 | 0.545 | — | — | 0.444 | — | — | — | 56.8 | 38.9 | 69 | 45.1 | |
| Logistic Regression2018.02 | 0.533 | 82.9 | 0.864 | 0.477 | 54.6 | — | — | — | — | — | — | |
| LR2020.07 | 0.533 | 0.829 | 0.864 | 0.477 | 0.546 | — | — | — | — | — | — | |
| Shi et al.2018.02 | 0.532 | — | 0.9 | — | — | — | — | — | — | — | — | |
| C-LSTM-Att2019.11 | 0.532 | — | 0.9 | — | — | — | — | — | — | — | — | |
| C-LSTM-Att2020.07 | 0.532 | — | 0.9 | — | — | — | — | — | — | — | — | |
| ClinicalBERTBest epoch=15, max token limit=5122022.10 | 0.5162 | 0.8194 | 0.8565 | 0.4361 | 0.5259 | — | — | — | — | — | — | |
| GRUBackbone=Bi-GRU, Paradigm=PLM Baselines2026.06 | 0.441 | — | — | 0.335 | — | — | — | 44.2 | 30 | 56.1 | 36.3 | |
| C-MemNN2020.07 | 0.42 | 0.833 | — | — | — | — | — | — | — | — | — | |
| KEPTLongformerSetting (trained from)=Finetuned, # Train Param=119.4M2022.10 | 0.326 | 0.827 | 0.833 | 0.304 | — | — | — | — | — | — | — | |
| KEPTLongformerSetting (trained from)=Pretrained, # Train Param=119.4M2022.10 | 0.309 | 0.814 | 0.823 | 0.258 | — | — | — | — | — | — | — | |
| KEPTLongformerSetting (trained from)=w/o HSAP, # Train Param=119.4M2022.10 | 0.299 | 0.802 | 0.822 | 0.243 | — | — | — | — | — | — | — | |
| KEPTLongformerSetting (trained from)=w/ LM & First, # Train Param=9.4M2022.10 | 0.296 | 0.79 | 0.815 | 0.235 | — | — | — | — | — | — | — | |
| KEPTLongformerSetting (trained from)=w/ LM & Last, # Train Param=9.4M2022.10 | 0.234 | 0.776 | 0.784 | 0.173 | — | — | — | — | — | — | — | |
| Few-shot+CoTBackbone=Qwen3-4B, Paradigm=LLM (Prompting)2026.06 | 0.182 | — | — | 0.078 | — | — | — | 18.2 | 5.9 | 52.8 | 11 | |
| MSMNSetting (trained from)=Pretrained, # Train Param=16.4M2022.10 | 0.172 | 0.753 | 0.762 | 0.171 | — | — | — | — | — | — | — | |
| KEPTLongformerSetting (trained from)=w/ LM only, # Train Param=0.6M2022.10 | 0.169 | 0.75 | 0.769 | 0.152 | — | — | — | — | — | — | — | |
| KEPTLongformerSetting (trained from)=Zero shot, # Train Param=02022.10 | 0.167 | 0.749 | 0.765 | 0.152 | — | — | — | — | — | — | — | |
| Zero-shot+CoTBackbone=Qwen3-4B, Paradigm=LLM (Prompting)2026.06 | 0.106 | — | — | 0.042 | — | — | — | 18 | 2.8 | 61.5 | 5.8 | |
| Zero-shotBackbone=Qwen3-4B, Paradigm=LLM (Prompting)2026.06 | 0.1 | — | — | 0.04 | — | — | — | 11.7 | 2.7 | 52.7 | 5.5 | |
| Few-shotBackbone=Qwen3-4B, Paradigm=LLM (Prompting)2026.06 | 0.08 | — | — | 0.039 | — | — | — | 12.1 | 2.8 | 33.4 | 4.6 | |
| Few-shot+CoTBackbone=Qwen2.5-1.5B, Paradigm=LLM (Prompting)2026.06 | 0.08 | — | — | 0.028 | — | — | — | 4.6 | 3.3 | 17.8 | 5.1 | |
| MSMNSetting (trained from)=Finetuned, # Train Param=16.4M2022.10 | 0.042 | 0.582 | 0.44 | 0.033 | — | — | — | — | — | — | — | |
| MSMNSetting (trained from)=Zero shot, # Train Param=02022.10 | 0.04 | 0.523 | 0.489 | 0.035 | — | — | — | — | — | — | — | |
| Few-shotBackbone=Qwen2.5-1.5B, Paradigm=LLM (Prompting)2026.06 | 0.039 | — | — | 0.019 | — | — | — | 5.6 | 1.4 | 16.1 | 2.2 | |
| Zero-shotBackbone=Qwen2.5-1.5B, Paradigm=LLM (Prompting)2026.06 | 0.016 | — | — | 0.007 | — | — | — | 4.7 | 0.4 | 29 | 0.8 | |
| Zero-shot+CoTBackbone=Qwen2.5-1.5B, Paradigm=LLM (Prompting)2026.06 | 0.015 | — | — | 0.006 | — | — | — | 6.9 | 0.3 | 28 | 0.8 | |
| C-MemNN2018.02 | — | 83.3 | — | — | 42 | — | — | — | — | — | — | |
| C-MemNN2019.11 | — | 0.833 | — | — | 0.42 | — | — | — | — | — | — |