Publicação
Clinical trial outcome prediction using a multimodal mixture-of-experts approach - an interpretative framework for predicting and explaining clinical trial failures with a retrieval-augmented chatbot
| Resumo: | This work presents MMCTO, a multimodal framework predicting clinical trial outcomes by integrating molecular, disease, and eligibility data. Based on the LIFTED architecture, it employs natural language transformation and a Mixture-of-Experts mechanism to unify heterogeneous inputs. It demonstrates superior predictive performance across trial phases on HINT and CTOD datasets. Ablation studies confirm the importance of LLM-generated features and conditioned gating. A RAG pipeline contextualizes results, while SHAP explanations provide transparency. The approach optimizes resources and streamlines processes, potentially avoiding costly failures and accelerating drug development timelines. |
|---|---|
| Autores principais: | Aparício, Carolina |
| Assunto: | Clinical trial outcomes Large Language Models Mixture-of-Experts LIFTED Retrieval-Augmented-Generation SHAP |
| Ano: | 2025 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | This work presents MMCTO, a multimodal framework predicting clinical trial outcomes by integrating molecular, disease, and eligibility data. Based on the LIFTED architecture, it employs natural language transformation and a Mixture-of-Experts mechanism to unify heterogeneous inputs. It demonstrates superior predictive performance across trial phases on HINT and CTOD datasets. Ablation studies confirm the importance of LLM-generated features and conditioned gating. A RAG pipeline contextualizes results, while SHAP explanations provide transparency. The approach optimizes resources and streamlines processes, potentially avoiding costly failures and accelerating drug development timelines. |
|---|