Fine-grained CI discrimination on LEADS 10-fold cross validation (in-domain)
93.8AUCMLP + Qwen3-Omni-30B
Evaluation results
| Method | Links | |
|---|---|---|
| MLP + Qwen3-Omni-30BAdaptor=MLP, Base Model=Qwen3-Omni-30B, Post Training=None2026.06 | 93.8 | |
| MLP + Clinical BERTAdaptor=MLP, Base Model=Clinical BERT, Post Training=SFT2026.06 | 90.2 | |
| MLP + BERT+WhisperAdaptor=MLP, Base Model=BERT+Whisper, Post Training=SFT2026.06 | 89.1 | |
| MLP + BERTAdaptor=MLP, Base Model=BERT, Post Training=SFT2026.06 | 88.7 | |
| DeTAiL (self)Adaptor=DeTAiL (self), Base Model=Qwen2.5-Omni-7B, Post Training=Distilled GRPO2026.06 | 86.3 | |
| DeTAiL (teacher)Adaptor=DeTAiL (teacher), Base Model=Qwen2.5-Omni-7B, Post Training=Distilled GRPO2026.06 | 84.4 | |
| MLP + BERT (GT text)+WhisperAdaptor=MLP, Base Model=BERT (GT text)+Whisper, Post Training=SFT2026.06 | 84.3 | |
| MLP + Qwen2.5-Omni-7BAdaptor=MLP, Base Model=Qwen2.5-Omni-7B, Post Training=None2026.06 | 83.9 | |
| MLP + BERT (GT text)Adaptor=MLP, Base Model=BERT (GT text), Post Training=SFT2026.06 | 83 | |
| Qwen2.5-Omni-7B (Zero-shot)Adaptor=None, Base Model=Qwen2.5-Omni-7B, Post Training=None2026.06 | 71.9 | |
| LoRA + Qwen2.5-Omni-7B (Distilled GRPO)Adaptor=LoRA, Base Model=Qwen2.5-Omni-7B, Post Training=Distilled GRPO2026.06 | 60.9 | |
| LoRA + Qwen2.5-Omni-7B (GRPO)Adaptor=LoRA, Base Model=Qwen2.5-Omni-7B, Post Training=GRPO2026.06 | 59.9 | |
| LoRA + Qwen2.5-Omni-7B (SFT)Adaptor=LoRA, Base Model=Qwen2.5-Omni-7B, Post Training=SFT2026.06 | 53.7 |