Publicação
Prognostic prediction models using Self-Attention for ICU patients developing acute kidney injury
| Resumo: | The general growth and improved accessibility to electronic health records demands an identical level of progress in terms of the research community regarding clinical models. The usage of machine learning techniques is key to this development, and so they are increasingly being used in large medical databases with the purpose of creating solutions that work for specified patients, no matter the task or the disease. Acute kidney injury (AKI) is a broad disease defined by abrupt changes in renal function. AKI has a high morbidity and mortality, with an increased focus on critically ill patients. The main goal of this thesis is to study the development of AKI within a patient’s stay in the intensive care unit (ICU). Data from the MIMIC-III database was used to collect information regarding the patients. After a detailed exclusion criteria, those were evaluated in terms of AKI stages, with the purpose of predicting the next value of AKI stage one hour after the sequence of information fed to the model. This can suggest the capacity of the model at predicting the aggravation of a patient’s AKI condition. The sequences used have hourly information for every feature, and were used sequences of 6h, 12h and 24h length. Self-attention mechanisms were used to make the predictions, using an adaptation for multi-variate time series built from the successfully used models on natural language processing (NLP) tasks. The predictions on this work were made for two variations of the KDIGO classification system: one where only the serum creatinine (SCr) criteria was taken into account to determine the patient’s AKI stage, and other where both SCr and urine output (UO) were considered. While most works addressing AKI only tend to use SCr values to determine the patient’s AKI condition, the results were compared using both approaches and were better when using both SCr and UO. For those experiments, the model achieved up to 68.05% accuracy predicting an episode of AKI, compared to the 66.67% accuracy achieved using only SCr values, which outperformed state-of-the-art results for both cases. Feature importance was also used for each dataset associated with the two variations of KDIGO classification system to identify what were the most important features. Furthermore, final results were compared when using all features versus only using the most 10 important ones. |
|---|---|
| Autores principais: | Domingues, Pedro Miguel Pereira |
| Assunto: | Insuficiência Renal Aguda Prognóstico MIMIC-III Self-Attention Feature Importance Teses de mestrado - 2022 |
| Ano: | 2022 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório da Universidade de Lisboa |
| _version_ | 1866811096032083968 |
|---|---|
| author | Domingues, Pedro Miguel Pereira |
| author_facet | Domingues, Pedro Miguel Pereira |
| author_role | author |
| contributor_name_str_mv | Garcia, Nuno Ricardo da Cruz Madeira, Sara Alexandra Cordeiro Repositório Científico de Acesso Aberto da ULisboa |
| country_str | PT |
| creators_json_txt | [{\"Person.name\":\"Domingues, Pedro Miguel Pereira\"}] |
| datacite.contributors.contributor.contributorName.fl_str_mv | Garcia, Nuno Ricardo da Cruz Madeira, Sara Alexandra Cordeiro Repositório Científico de Acesso Aberto da ULisboa |
| datacite.creators.creator.creatorName.fl_str_mv | Domingues, Pedro Miguel Pereira |
| datacite.date.Accepted.fl_str_mv | 2022-01-01T00:00:00Z |
| datacite.date.available.fl_str_mv | 2022-07-20T09:35:46Z |
| datacite.date.embargoed.fl_str_mv | 2022-07-20T09:35:46Z |
| datacite.rights.fl_str_mv | http://purl.org/coar/access_right/c_abf2 |
| datacite.subjects.subject.fl_str_mv | Insuficiência Renal Aguda Prognóstico MIMIC-III Self-Attention Feature Importance Teses de mestrado - 2022 |
| datacite.titles.title.fl_str_mv | Prognostic prediction models using Self-Attention for ICU patients developing acute kidney injury |
| dc.contributor.none.fl_str_mv | Garcia, Nuno Ricardo da Cruz Madeira, Sara Alexandra Cordeiro Repositório Científico de Acesso Aberto da ULisboa |
| dc.creator.none.fl_str_mv | Domingues, Pedro Miguel Pereira |
| dc.date.Accepted.fl_str_mv | 2022-01-01T00:00:00Z |
| dc.date.available.fl_str_mv | 2022-07-20T09:35:46Z |
| dc.date.embargoed.fl_str_mv | 2022-07-20T09:35:46Z |
| dc.format.none.fl_str_mv | application/pdf |
| dc.identifier.none.fl_str_mv | http://hdl.handle.net/10451/53871 |
| dc.language.none.fl_str_mv | eng |
| dc.rights.none.fl_str_mv | http://purl.org/coar/access_right/c_abf2 |
| dc.subject.none.fl_str_mv | Insuficiência Renal Aguda Prognóstico MIMIC-III Self-Attention Feature Importance Teses de mestrado - 2022 |
| dc.title.fl_str_mv | Prognostic prediction models using Self-Attention for ICU patients developing acute kidney injury |
| dc.type.none.fl_str_mv | http://purl.org/coar/resource_type/c_bdcc |
| description | The general growth and improved accessibility to electronic health records demands an identical level of progress in terms of the research community regarding clinical models. The usage of machine learning techniques is key to this development, and so they are increasingly being used in large medical databases with the purpose of creating solutions that work for specified patients, no matter the task or the disease. Acute kidney injury (AKI) is a broad disease defined by abrupt changes in renal function. AKI has a high morbidity and mortality, with an increased focus on critically ill patients. The main goal of this thesis is to study the development of AKI within a patient’s stay in the intensive care unit (ICU). Data from the MIMIC-III database was used to collect information regarding the patients. After a detailed exclusion criteria, those were evaluated in terms of AKI stages, with the purpose of predicting the next value of AKI stage one hour after the sequence of information fed to the model. This can suggest the capacity of the model at predicting the aggravation of a patient’s AKI condition. The sequences used have hourly information for every feature, and were used sequences of 6h, 12h and 24h length. Self-attention mechanisms were used to make the predictions, using an adaptation for multi-variate time series built from the successfully used models on natural language processing (NLP) tasks. The predictions on this work were made for two variations of the KDIGO classification system: one where only the serum creatinine (SCr) criteria was taken into account to determine the patient’s AKI stage, and other where both SCr and urine output (UO) were considered. While most works addressing AKI only tend to use SCr values to determine the patient’s AKI condition, the results were compared using both approaches and were better when using both SCr and UO. For those experiments, the model achieved up to 68.05% accuracy predicting an episode of AKI, compared to the 66.67% accuracy achieved using only SCr values, which outperformed state-of-the-art results for both cases. Feature importance was also used for each dataset associated with the two variations of KDIGO classification system to identify what were the most important features. Furthermore, final results were compared when using all features versus only using the most 10 important ones. |
| dirty | 0 |
| eu_rights_str_mv | openAccess |
| format | masterThesis |
| fulltext.url.fl_str_mv | https://repositorio.ulisboa.pt/bitstreams/ab47f68d-bff8-40d2-acbb-1ff42d5f5ed5/download |
| id | ul_b89e1568c22e2f7ed5e2babc6ffab2de |
| identifier.url.fl_str_mv | http://hdl.handle.net/10451/53871 |
| instacron_str | ul |
| institution | Universidade de Lisboa |
| instname_str | Universidade de Lisboa |
| language | eng |
| network_acronym_str | ul |
| network_name_str | Repositório da Universidade de Lisboa |
| oai_identifier_str | oai:repositorio.ulisboa.pt:10451/53871 |
| organization_str_mv | urn:organizationAcronym:ul |
| person_str_mv | Domingues, Pedro Miguel Pereira |
| publishDate | 2022 |
| reponame_str | Repositório da Universidade de Lisboa |
| repository_id_str | urn:repositoryAcronym:ul |
| service_str_mv | urn:repositoryAcronym:ul |
| spelling | engpt_PTThe general growth and improved accessibility to electronic health records demands an identical level of progress in terms of the research community regarding clinical models. The usage of machine learning techniques is key to this development, and so they are increasingly being used in large medical databases with the purpose of creating solutions that work for specified patients, no matter the task or the disease. Acute kidney injury (AKI) is a broad disease defined by abrupt changes in renal function. AKI has a high morbidity and mortality, with an increased focus on critically ill patients. The main goal of this thesis is to study the development of AKI within a patient’s stay in the intensive care unit (ICU). Data from the MIMIC-III database was used to collect information regarding the patients. After a detailed exclusion criteria, those were evaluated in terms of AKI stages, with the purpose of predicting the next value of AKI stage one hour after the sequence of information fed to the model. This can suggest the capacity of the model at predicting the aggravation of a patient’s AKI condition. The sequences used have hourly information for every feature, and were used sequences of 6h, 12h and 24h length. Self-attention mechanisms were used to make the predictions, using an adaptation for multi-variate time series built from the successfully used models on natural language processing (NLP) tasks. The predictions on this work were made for two variations of the KDIGO classification system: one where only the serum creatinine (SCr) criteria was taken into account to determine the patient’s AKI stage, and other where both SCr and urine output (UO) were considered. While most works addressing AKI only tend to use SCr values to determine the patient’s AKI condition, the results were compared using both approaches and were better when using both SCr and UO. For those experiments, the model achieved up to 68.05% accuracy predicting an episode of AKI, compared to the 66.67% accuracy achieved using only SCr values, which outperformed state-of-the-art results for both cases. Feature importance was also used for each dataset associated with the two variations of KDIGO classification system to identify what were the most important features. Furthermore, final results were compared when using all features versus only using the most 10 important ones.application/pdfpt_PTPrognostic prediction models using Self-Attention for ICU patients developing acute kidney injuryDomingues, Pedro Miguel PereiraGarcia, Nuno Ricardo da CruzMadeira, Sara Alexandra CordeiroHostingInstitutionOrganizationalRepositório Científico de Acesso Aberto da ULisboae-mailmailto:repositorio@reitoria.ulisboa.ptrepositorio@reitoria.ulisboa.ptURNurn:tid:2032177302022-07-20T09:35:46Z202220222022-01-01T00:00:00ZHandlehttp://hdl.handle.net/10451/53871http://purl.org/coar/access_right/c_abf2open accessInsuficiência Renal AgudaPrognósticoMIMIC-IIISelf-AttentionFeature ImportanceTeses de mestrado - 20222811323 bytesliteraturehttp://purl.org/coar/resource_type/c_bdccmaster thesishttp://purl.org/coar/access_right/c_abf2application/pdffulltexthttps://repositorio.ulisboa.pt/bitstreams/ab47f68d-bff8-40d2-acbb-1ff42d5f5ed5/download |
| spellingShingle | Prognostic prediction models using Self-Attention for ICU patients developing acute kidney injury Domingues, Pedro Miguel Pereira Insuficiência Renal Aguda Prognóstico MIMIC-III Self-Attention Feature Importance Teses de mestrado - 2022 |
| status | SINGLETON |
| subject.fl_str_mv | Insuficiência Renal Aguda Prognóstico MIMIC-III Self-Attention Feature Importance Teses de mestrado - 2022 |
| title | Prognostic prediction models using Self-Attention for ICU patients developing acute kidney injury |
| title_full | Prognostic prediction models using Self-Attention for ICU patients developing acute kidney injury |
| title_fullStr | Prognostic prediction models using Self-Attention for ICU patients developing acute kidney injury |
| title_full_unstemmed | Prognostic prediction models using Self-Attention for ICU patients developing acute kidney injury |
| title_short | Prognostic prediction models using Self-Attention for ICU patients developing acute kidney injury |
| title_sort | Prognostic prediction models using Self-Attention for ICU patients developing acute kidney injury |
| topic | Insuficiência Renal Aguda Prognóstico MIMIC-III Self-Attention Feature Importance Teses de mestrado - 2022 |
| topic_facet | Insuficiência Renal Aguda Prognóstico MIMIC-III Self-Attention Feature Importance Teses de mestrado - 2022 |
| url | http://hdl.handle.net/10451/53871 |
| visible | 1 |