Publicação
TrialMatch: A Transformer Architecture to Match Patients to Clinical Trials
| Resumo: | Around 80% of clinical trials fail to meet the patient recruitment requirements, which not only hinders the market growth but also delays patients’ access to new and effec- tive treatments. A possible approach is to use Electronic Health Records (EHRs) to help match patients to clinical trials. Past attempts at achieving this exact goal took place, but due to a lack of data, they were unsuccessful. In 2021 Text REtrieval Conference (TREC) introduced the Clinical Trials Track, where participants were challenged with retrieving relevant clinical trials given the patient’s descriptions simulating admission notes. Utilizing the track results as a baseline, we tackled the challenge, for this, we re- sort to Information Retrieval (IR), implementing a pipeline for document ranking where we explore the different retrieval methods, how to filter the clinical trials based on the criteria, and reranking with Transformer based models. To tackle the problem, we ex- plored models pre-trained on the biomedical domain, how to deal with long queries and documents through query expansion and passage selection, and how to distinguish an eligible clinical trial from an excluded clinical trial, using techniques such as Named Entity Recognition (NER) and Clinical Assertion. Our results let to the finding that the current state-of-the-art Bidirectional Encoder Representations from Transformers (BERT) bi-encoders outperform the cross-encoders in the mentioned task, whilst proving that sparse retrieval methods are capable of obtaining competitive outcomes, and to finalize we showed that the use of the demographic information available can be used to improve the final result. |
|---|---|
| Autores principais: | Cardoso, Bruno Dantas |
| Assunto: | TREC Clinical Trial Electronic Health Record Transformer BERT T5 |
| Ano: | 2022 |
| 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: | Around 80% of clinical trials fail to meet the patient recruitment requirements, which not only hinders the market growth but also delays patients’ access to new and effec- tive treatments. A possible approach is to use Electronic Health Records (EHRs) to help match patients to clinical trials. Past attempts at achieving this exact goal took place, but due to a lack of data, they were unsuccessful. In 2021 Text REtrieval Conference (TREC) introduced the Clinical Trials Track, where participants were challenged with retrieving relevant clinical trials given the patient’s descriptions simulating admission notes. Utilizing the track results as a baseline, we tackled the challenge, for this, we re- sort to Information Retrieval (IR), implementing a pipeline for document ranking where we explore the different retrieval methods, how to filter the clinical trials based on the criteria, and reranking with Transformer based models. To tackle the problem, we ex- plored models pre-trained on the biomedical domain, how to deal with long queries and documents through query expansion and passage selection, and how to distinguish an eligible clinical trial from an excluded clinical trial, using techniques such as Named Entity Recognition (NER) and Clinical Assertion. Our results let to the finding that the current state-of-the-art Bidirectional Encoder Representations from Transformers (BERT) bi-encoders outperform the cross-encoders in the mentioned task, whilst proving that sparse retrieval methods are capable of obtaining competitive outcomes, and to finalize we showed that the use of the demographic information available can be used to improve the final result. |
|---|