FRINK Linked Data Fragments server

Matches in SemOpenAlex for { <https://semopenalex.org/work/W4377012603> ?p ?o ?g. }

• W4377012603 endingPage "e399" @default.
• W4377012603 startingPage "e398" @default.
• W4377012603 abstract "High performance computing and increasingly complex computational neural networks are changing the way in which health care is provided. However, increasing the complexity of models might not necessarily be the solution, with data used and the question being asked important determinants of clinical applicability. In The Lancet Digital Health, Amirhossein Azhie and colleagues1Azhie A Sharma D Sheth P et al.A deep learning framework for personalised dynamic diagnosis of graft fibrosis after liver transplantation: a retrospective, single Canadian centre, longitudinal study.Lancet Digit Health. 2023; (published online May 18.)https://doi.org/10.1016/S2589-7500(23)00068-7Google Scholar describe the development of a deep learning model using a weighted Long Short-Term Memory (LSTM) neural network2Hochreiter S Schmidhuber J Long Short-Term Memory.Neural Comput. 1997; 9: 1735-1780Crossref PubMed Scopus (51923) Google Scholar to predict the risk of F2 or worse fibrosis in the transplanted liver, which can be associated with poorer survival in the recipient and can drive the need for retransplantation. The authors accessed adult liver transplantation data from the University Health Network, Toronto, Canada, from 1987 to 2019. 1893 people who received a liver transplant and at least one liver biopsy after transplantation were identified; if multiple biopsies over time had been done, the last in date sequence was used to define the fibrotic status of the transplanted liver. Clinically indicated rather than protocol biopsies were included. The dataset was split into a training set (1221 [70%] individuals) and test set (523 [30%] individuals). An additional model test was done on 149 patients who underwent transient elastography within 1 year of their liver biopsy. Initial features (n=26), a combination of demographic, clinical and longitudinal donor and recipient variables were used in training and testing of deep learning models. Features selected were based on local clinical practice and utility (ie, easily available blood tests and no need for additional specialised investigations). In total 167 091 longitudinal data points were used for training and testing of the machine learning models. In the weighted LSTM neural network, rather than imputing missing data Azhie and colleagues1Azhie A Sharma D Sheth P et al.A deep learning framework for personalised dynamic diagnosis of graft fibrosis after liver transplantation: a retrospective, single Canadian centre, longitudinal study.Lancet Digit Health. 2023; (published online May 18.)https://doi.org/10.1016/S2589-7500(23)00068-7Google Scholar used a padding and masking approach to handle the variability in length of data for each patient at various timepoints. However, a weighted approach was used to reduce imbalance for unbiased LSTM learning because there was an insufficient number of cases for training. The contribution of a variable to model prediction was measured using Integrated Gradient methodology, identifying ten features that were the most predictive of F2 or worse fibrosis. Tanh was used as the activation function within the weighted LSTM neural network and softmax applied to the output layer. Azhie and colleagues1Azhie A Sharma D Sheth P et al.A deep learning framework for personalised dynamic diagnosis of graft fibrosis after liver transplantation: a retrospective, single Canadian centre, longitudinal study.Lancet Digit Health. 2023; (published online May 18.)https://doi.org/10.1016/S2589-7500(23)00068-7Google Scholar tested the performance of their weighted LSTM neural network against a number of cross-sectional machine learning models (random forest, support vector machine, logistic regression, lasso regression, and ridge regression) to confirm that single timepoint information is less predictive than longitudinal data. Additionally, the authors also showed that their weighted LSTM outperformed other established deep learning models (Recurrent Neural and Temporal Convolutional Networks) used in timed series prediction. Model performance at predicting fibrosis was compared using aspartate aminotransferase-to-platelet ratio index (APRI), Fibrosis-4 index (FIB-4), and transient elastography. This work is the first description of deep learning being used to predict the event of fibrosis in a transplanted liver; and the authors concluded that the weighted LSTM model consistently performed better at predicting the risk of F2 or worse fibrosis. A large number of machine learning models with varying performance in predicting the binary outcome of F2 or worse fibrosis or not have been built and tested by Azhie and colleagues,1Azhie A Sharma D Sheth P et al.A deep learning framework for personalised dynamic diagnosis of graft fibrosis after liver transplantation: a retrospective, single Canadian centre, longitudinal study.Lancet Digit Health. 2023; (published online May 18.)https://doi.org/10.1016/S2589-7500(23)00068-7Google Scholar with the aim of the algorithm to flag up patients at risk of developing fibrosis to trigger a review in clinical management. Initiating an earlier change in management, the authors suggest, might then reduce the need for liver biopsy and also improve long-term outcomes. Based on the work presented, temporal variability, and the different number of follow-up intervals for people who had liver transplantation was well captured by weighted LSTM compared with other recognised longitudinal machine learning frameworks. Subgroup analysis of the test data showed that the weighted LSTM continues to perform well, irrespective of the cause of fibrosis or change in transplant era. This suggests that there is no need for retraining of their deep learning model for a specific disease or fibrosis aetiology. The weighted LSTM neural network appears to predict the risk of F2 or worse fibrosis in the transplanted liver better than FIB-4 and APRI, and is equivalent to transient elastography. However, there are a number of caveats, mainly related to the absence of any validated technique to measure fibrosis in the transplanted liver. Additionally, the ground truth and gold standard reference that the authors use for classifying fibrosis is based on a histological assessment of a liver biopsy using the METAVIR score, a scoring system derived from non-transplant Hepatitis C liver biopsies.3Bedossa P Poynard T An algorithm for the grading of activity in chronic hepatitis C. The METAVIR Cooperative Study Group.Hepatol. 1996; 24: 289-293Crossref PubMed Google Scholar Likewise, histological abnormalities in post liver transplant biopsies do not always equate with liver function. Neither do they always need to drive a change in clinical care. Many of these issues are also raised by Azhie and colleagues;1Azhie A Sharma D Sheth P et al.A deep learning framework for personalised dynamic diagnosis of graft fibrosis after liver transplantation: a retrospective, single Canadian centre, longitudinal study.Lancet Digit Health. 2023; (published online May 18.)https://doi.org/10.1016/S2589-7500(23)00068-7Google Scholar these are issues that will influence how their deep learning model has learnt to predict significant fibrosis and will also affect how useful their weighted LSTM will be in clinical practice. In the meantime, the team are planning to develop a clinical dashboard to integrate with their present electronic health records (Epic Systems, Verona, WI, USA), and it will be of interest to see how their neural network evolves and whether it will improve outcome in their patient population. The code is open source4Fibrosis_LSTM2021.https://github.com/divya031090/Fibrosis_LSTMDate accessed: April 28, 2023Google Scholar and remains to be validated. I declare no competing interests. A deep learning framework for personalised dynamic diagnosis of graft fibrosis after liver transplantation: a retrospective, single Canadian centre, longitudinal studyDeep learning algorithms, particularly weighted LSTM, outperform other routinely used non-invasive modalities and could help with the earlier diagnosis of graft fibrosis using longitudinal clinical and laboratory variables. The list of most important predictive variables for the development of fibrosis will enable clinicians to modify their management accordingly to prevent onset of graft cirrhosis. Full-Text PDF Open Access" @default.
• W4377012603 created "2023-05-19" @default.
• W4377012603 creator A5077466283 @default.
• W4377012603 date "2023-07-01" @default.
• W4377012603 modified "2023-09-26" @default.
• W4377012603 title "Will deep learning change outcomes in liver transplant?" @default.
• W4377012603 cites W2064675550 @default.
• W4377012603 cites W2094174828 @default.
• W4377012603 cites W4377012520 @default.
• W4377012603 doi "https://doi.org/10.1016/s2589-7500(23)00089-4" @default.
• W4377012603 hasPubMedId "https://pubmed.ncbi.nlm.nih.gov/37210230" @default.
• W4377012603 hasPublicationYear "2023" @default.
• W4377012603 type Work @default.
• W4377012603 citedByCount "0" @default.
• W4377012603 crossrefType "journal-article" @default.
• W4377012603 hasAuthorship W4377012603A5077466283 @default.
• W4377012603 hasBestOaLocation W43770126031 @default.
• W4377012603 hasConcept C108583219 @default.
• W4377012603 hasConcept C154945302 @default.
• W4377012603 hasConcept C15744967 @default.
• W4377012603 hasConcept C41008148 @default.
• W4377012603 hasConcept C71924100 @default.
• W4377012603 hasConceptScore W4377012603C108583219 @default.
• W4377012603 hasConceptScore W4377012603C154945302 @default.
• W4377012603 hasConceptScore W4377012603C15744967 @default.
• W4377012603 hasConceptScore W4377012603C41008148 @default.
• W4377012603 hasConceptScore W4377012603C71924100 @default.
• W4377012603 hasIssue "7" @default.
• W4377012603 hasLocation W43770126031 @default.
• W4377012603 hasLocation W43770126032 @default.
• W4377012603 hasOpenAccess W4377012603 @default.
• W4377012603 hasPrimaryLocation W43770126031 @default.
• W4377012603 hasRelatedWork W2731899572 @default.
• W4377012603 hasRelatedWork W2748952813 @default.
• W4377012603 hasRelatedWork W2899084033 @default.
• W4377012603 hasRelatedWork W2939353110 @default.
• W4377012603 hasRelatedWork W3009238340 @default.
• W4377012603 hasRelatedWork W3215138031 @default.
• W4377012603 hasRelatedWork W4312962853 @default.
• W4377012603 hasRelatedWork W4321369474 @default.
• W4377012603 hasRelatedWork W4327774331 @default.
• W4377012603 hasRelatedWork W4360585206 @default.
• W4377012603 hasVolume "5" @default.
• W4377012603 isParatext "false" @default.
• W4377012603 isRetracted "false" @default.
• W4377012603 workType "article" @default.