Abstract
Discussion Forum (0)
ePoster
Topic: End-of-Life and Supportive Care
Sauthier, Nicolas, MD BEng MSc c.1;Bouchakri, Rima, PhD1; Carrier, François-Martin, MD MSc1; Chassé, Michaël, MD PhD1
1 Centre de Recherche du Centre Hospitalier de l'Université de Montréal (CRCHUM), Montréal, Canada
CRCHUM
Introduction:
Despite a slow improvement in total organ donors over the last 20 years1, organ donation is still not meeting the demand. Organ transplantation depends on potential organ donor identification which is still a major challenge. Retrospective studies show that we may miss between 30% and 50 % of potential organ donors2–5. By increasing the total number of potential donors referred to an Organ Donation Organization (ODO), the number of organ donors could substantially increase.
Objectives:
We aimed to develop a deep learning6 model using only routinely collected data that could help with the screening of potential organ donors. We present the development and validation of that model.
Methods:
The model was derived using the electronic health record data from intensive care unit (ICU) stays from Jan 1st 2012, until Dec 31th 2019, in the CHUM (Montreal University Hospital Center, Montreal). We defined our organ donor population as (1) organ donors that were identified in the study institution (2) potential organ donors referred to the ODO but deemed ineligible for transplantation and (3) non-referred potential organ donors identified from manual chart review death audit. We included all laboratory analyses as well as the presence/absence of head scan. We developed a two-step model. First, we trained a deep convolutional neural network7 autoencoder (CNN-AE)8. This model was a non-supervised model that served as an embedding and a dimensionality reduction tool, to reduce noise9 and for transfer learning and pretraining10. The second step was to add a deep classifier to categorize each patient as donor/non-donor. We also implemented a one-layer logistical model (LM), based on the last available value, as a comparator to our neural network (NN) model. The CNN-AE was trained on 85% randomly selected non-donors. The NN model and LM were trained on the rest of the patients, randomly separated in a train (60%), validation (20%) and test (20%) sets.
We compared our NN model to our LM using a ROC curve, scaled Brier score and calibration curves11 with bootstrap for confidence intervals. We manually reviewed the files of patients that were wrongly predicted as donor but who predicted a degree of confidence over 95% in the validation set.
Results:
Our complete dataset used 19717 patients with 392 donors. After excluding rare laboratory analysis, the NN model and LM were trained on 105 distinct laboratory analyses. On the test set, NN model and LM performed similarly with a ROC-AUC of respectively of 0.950 (95%CI 0.923-0.974) and 0.947 (95%CI 0.9169-0.9730) and scaled Brier score of 0.313 (0.134-0.472) and 0.432 (0.270-0.580). At best accuracy, sensitivity and specificity were 82% (95%CI 73-90) and 92% (95%CI 89-94) for the NN model and 82% (95%CI 73-90) and 94% (95%CI 91-95) for the LM. The NN model accuracy was more consistent across subgroups compared to the LM (fig 1).
Conclusion:
We present preliminary evidence that routinely collected medical data can be used to screen for potential organ donors. When comparing our more complex temporal model (NN) with its non-temporal simpler version (LM) both performed as well globally. However, the NN keeps a good accuracy in the more complex clinical patterns. Prospective and external validation of these models are required for further calibration before any potential clinical application.
Image 1
Image 2
1. Canadian Institute for Health Information. Annual Statistics on Organ Replacement in Canada: Dialysis, Transplantation and Donation, 2009 to 2018. (2019).
2. Krmpotic, K., Payne, C., Isenor, C. & Dhanani, S. Delayed Referral Results in Missed Opportunities for Organ Donation After Circulatory Death. Crit. Care Med. 45, 989–992 (2017).
3. Kutsogiannis, D. J., Asthana, S., Townsend, D. R., Singh, G. & Karvellas, C. J. The incidence of potential missed organ donors in intensive care units and emergency rooms: A retrospective cohort. Intensive Care Med. 39, 1452–1459 (2013).
4. Sairanen, T. et al. Lost potential of kidney and liver donors amongst deceased intracerebral hemorrhage patients. Eur. J. Neurol. (2014).
5. Opdam, H. I. & Silvester, W. Potential for organ donation in Victoria: An audit of hospital deaths. Med. J. Aust. 185, 250–254 (2006).
6. Beam, A. L. & Kohane, I. S. Big data and machine learning in health care. JAMA - Journal of the American Medical Association (2018).
7. He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (2016).
8. Hinton, G. E. & Salakhutdinov, R. R. Reducing the dimensionality of data with neural networks. Science (80-. ). (2006). doi:10.1126/science.1127647
9. Charte, D., Charte, F., del Jesus, M. J. & Herrera, F. A Showcase of the Use of Autoencoders in Feature Learning Applications. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 11487 LNCS, (2019).
10. Ferreira, M. F., Camacho, R. & Teixeira, L. F. Autoencoders as Weight Initialization of Deep Classification Networks for Cancer versus Cancer Studies. (2020).
11. Leisman, D. E. et al. Development and Reporting of Prediction Models. Crit. Care Med. (2020).
Topic: End-of-Life and Supportive Care
Sauthier, Nicolas, MD BEng MSc c.1;Bouchakri, Rima, PhD1; Carrier, François-Martin, MD MSc1; Chassé, Michaël, MD PhD1
1 Centre de Recherche du Centre Hospitalier de l'Université de Montréal (CRCHUM), Montréal, Canada
CRCHUM
Introduction:
Despite a slow improvement in total organ donors over the last 20 years1, organ donation is still not meeting the demand. Organ transplantation depends on potential organ donor identification which is still a major challenge. Retrospective studies show that we may miss between 30% and 50 % of potential organ donors2–5. By increasing the total number of potential donors referred to an Organ Donation Organization (ODO), the number of organ donors could substantially increase.
Objectives:
We aimed to develop a deep learning6 model using only routinely collected data that could help with the screening of potential organ donors. We present the development and validation of that model.
Methods:
The model was derived using the electronic health record data from intensive care unit (ICU) stays from Jan 1st 2012, until Dec 31th 2019, in the CHUM (Montreal University Hospital Center, Montreal). We defined our organ donor population as (1) organ donors that were identified in the study institution (2) potential organ donors referred to the ODO but deemed ineligible for transplantation and (3) non-referred potential organ donors identified from manual chart review death audit. We included all laboratory analyses as well as the presence/absence of head scan. We developed a two-step model. First, we trained a deep convolutional neural network7 autoencoder (CNN-AE)8. This model was a non-supervised model that served as an embedding and a dimensionality reduction tool, to reduce noise9 and for transfer learning and pretraining10. The second step was to add a deep classifier to categorize each patient as donor/non-donor. We also implemented a one-layer logistical model (LM), based on the last available value, as a comparator to our neural network (NN) model. The CNN-AE was trained on 85% randomly selected non-donors. The NN model and LM were trained on the rest of the patients, randomly separated in a train (60%), validation (20%) and test (20%) sets.
We compared our NN model to our LM using a ROC curve, scaled Brier score and calibration curves11 with bootstrap for confidence intervals. We manually reviewed the files of patients that were wrongly predicted as donor but who predicted a degree of confidence over 95% in the validation set.
Results:
Our complete dataset used 19717 patients with 392 donors. After excluding rare laboratory analysis, the NN model and LM were trained on 105 distinct laboratory analyses. On the test set, NN model and LM performed similarly with a ROC-AUC of respectively of 0.950 (95%CI 0.923-0.974) and 0.947 (95%CI 0.9169-0.9730) and scaled Brier score of 0.313 (0.134-0.472) and 0.432 (0.270-0.580). At best accuracy, sensitivity and specificity were 82% (95%CI 73-90) and 92% (95%CI 89-94) for the NN model and 82% (95%CI 73-90) and 94% (95%CI 91-95) for the LM. The NN model accuracy was more consistent across subgroups compared to the LM (fig 1).
Conclusion:
We present preliminary evidence that routinely collected medical data can be used to screen for potential organ donors. When comparing our more complex temporal model (NN) with its non-temporal simpler version (LM) both performed as well globally. However, the NN keeps a good accuracy in the more complex clinical patterns. Prospective and external validation of these models are required for further calibration before any potential clinical application.
Image 1
Image 2
1. Canadian Institute for Health Information. Annual Statistics on Organ Replacement in Canada: Dialysis, Transplantation and Donation, 2009 to 2018. (2019).
2. Krmpotic, K., Payne, C., Isenor, C. & Dhanani, S. Delayed Referral Results in Missed Opportunities for Organ Donation After Circulatory Death. Crit. Care Med. 45, 989–992 (2017).
3. Kutsogiannis, D. J., Asthana, S., Townsend, D. R., Singh, G. & Karvellas, C. J. The incidence of potential missed organ donors in intensive care units and emergency rooms: A retrospective cohort. Intensive Care Med. 39, 1452–1459 (2013).
4. Sairanen, T. et al. Lost potential of kidney and liver donors amongst deceased intracerebral hemorrhage patients. Eur. J. Neurol. (2014).
5. Opdam, H. I. & Silvester, W. Potential for organ donation in Victoria: An audit of hospital deaths. Med. J. Aust. 185, 250–254 (2006).
6. Beam, A. L. & Kohane, I. S. Big data and machine learning in health care. JAMA - Journal of the American Medical Association (2018).
7. He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (2016).
8. Hinton, G. E. & Salakhutdinov, R. R. Reducing the dimensionality of data with neural networks. Science (80-. ). (2006). doi:10.1126/science.1127647
9. Charte, D., Charte, F., del Jesus, M. J. & Herrera, F. A Showcase of the Use of Autoencoders in Feature Learning Applications. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 11487 LNCS, (2019).
10. Ferreira, M. F., Camacho, R. & Teixeira, L. F. Autoencoders as Weight Initialization of Deep Classification Networks for Cancer versus Cancer Studies. (2020).
11. Leisman, D. E. et al. Development and Reporting of Prediction Models. Crit. Care Med. (2020).
ePoster
Topic: End-of-Life and Supportive Care
Sauthier, Nicolas, MD BEng MSc c.1;Bouchakri, Rima, PhD1; Carrier, François-Martin, MD MSc1; Chassé, Michaël, MD PhD1
1 Centre de Recherche du Centre Hospitalier de l'Université de Montréal (CRCHUM), Montréal, Canada
CRCHUM
Introduction:
Despite a slow improvement in total organ donors over the last 20 years1, organ donation is still not meeting the demand. Organ transplantation depends on potential organ donor identification which is still a major challenge. Retrospective studies show that we may miss between 30% and 50 % of potential organ donors2–5. By increasing the total number of potential donors referred to an Organ Donation Organization (ODO), the number of organ donors could substantially increase.
Objectives:
We aimed to develop a deep learning6 model using only routinely collected data that could help with the screening of potential organ donors. We present the development and validation of that model.
Methods:
The model was derived using the electronic health record data from intensive care unit (ICU) stays from Jan 1st 2012, until Dec 31th 2019, in the CHUM (Montreal University Hospital Center, Montreal). We defined our organ donor population as (1) organ donors that were identified in the study institution (2) potential organ donors referred to the ODO but deemed ineligible for transplantation and (3) non-referred potential organ donors identified from manual chart review death audit. We included all laboratory analyses as well as the presence/absence of head scan. We developed a two-step model. First, we trained a deep convolutional neural network7 autoencoder (CNN-AE)8. This model was a non-supervised model that served as an embedding and a dimensionality reduction tool, to reduce noise9 and for transfer learning and pretraining10. The second step was to add a deep classifier to categorize each patient as donor/non-donor. We also implemented a one-layer logistical model (LM), based on the last available value, as a comparator to our neural network (NN) model. The CNN-AE was trained on 85% randomly selected non-donors. The NN model and LM were trained on the rest of the patients, randomly separated in a train (60%), validation (20%) and test (20%) sets.
We compared our NN model to our LM using a ROC curve, scaled Brier score and calibration curves11 with bootstrap for confidence intervals. We manually reviewed the files of patients that were wrongly predicted as donor but who predicted a degree of confidence over 95% in the validation set.
Results:
Our complete dataset used 19717 patients with 392 donors. After excluding rare laboratory analysis, the NN model and LM were trained on 105 distinct laboratory analyses. On the test set, NN model and LM performed similarly with a ROC-AUC of respectively of 0.950 (95%CI 0.923-0.974) and 0.947 (95%CI 0.9169-0.9730) and scaled Brier score of 0.313 (0.134-0.472) and 0.432 (0.270-0.580). At best accuracy, sensitivity and specificity were 82% (95%CI 73-90) and 92% (95%CI 89-94) for the NN model and 82% (95%CI 73-90) and 94% (95%CI 91-95) for the LM. The NN model accuracy was more consistent across subgroups compared to the LM (fig 1).
Conclusion:
We present preliminary evidence that routinely collected medical data can be used to screen for potential organ donors. When comparing our more complex temporal model (NN) with its non-temporal simpler version (LM) both performed as well globally. However, the NN keeps a good accuracy in the more complex clinical patterns. Prospective and external validation of these models are required for further calibration before any potential clinical application.
Image 1
Image 2
1. Canadian Institute for Health Information. Annual Statistics on Organ Replacement in Canada: Dialysis, Transplantation and Donation, 2009 to 2018. (2019).
2. Krmpotic, K., Payne, C., Isenor, C. & Dhanani, S. Delayed Referral Results in Missed Opportunities for Organ Donation After Circulatory Death. Crit. Care Med. 45, 989–992 (2017).
3. Kutsogiannis, D. J., Asthana, S., Townsend, D. R., Singh, G. & Karvellas, C. J. The incidence of potential missed organ donors in intensive care units and emergency rooms: A retrospective cohort. Intensive Care Med. 39, 1452–1459 (2013).
4. Sairanen, T. et al. Lost potential of kidney and liver donors amongst deceased intracerebral hemorrhage patients. Eur. J. Neurol. (2014).
5. Opdam, H. I. & Silvester, W. Potential for organ donation in Victoria: An audit of hospital deaths. Med. J. Aust. 185, 250–254 (2006).
6. Beam, A. L. & Kohane, I. S. Big data and machine learning in health care. JAMA - Journal of the American Medical Association (2018).
7. He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (2016).
8. Hinton, G. E. & Salakhutdinov, R. R. Reducing the dimensionality of data with neural networks. Science (80-. ). (2006). doi:10.1126/science.1127647
9. Charte, D., Charte, F., del Jesus, M. J. & Herrera, F. A Showcase of the Use of Autoencoders in Feature Learning Applications. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 11487 LNCS, (2019).
10. Ferreira, M. F., Camacho, R. & Teixeira, L. F. Autoencoders as Weight Initialization of Deep Classification Networks for Cancer versus Cancer Studies. (2020).
11. Leisman, D. E. et al. Development and Reporting of Prediction Models. Crit. Care Med. (2020).
Topic: End-of-Life and Supportive Care
Sauthier, Nicolas, MD BEng MSc c.1;Bouchakri, Rima, PhD1; Carrier, François-Martin, MD MSc1; Chassé, Michaël, MD PhD1
1 Centre de Recherche du Centre Hospitalier de l'Université de Montréal (CRCHUM), Montréal, Canada
CRCHUM
Introduction:
Despite a slow improvement in total organ donors over the last 20 years1, organ donation is still not meeting the demand. Organ transplantation depends on potential organ donor identification which is still a major challenge. Retrospective studies show that we may miss between 30% and 50 % of potential organ donors2–5. By increasing the total number of potential donors referred to an Organ Donation Organization (ODO), the number of organ donors could substantially increase.
Objectives:
We aimed to develop a deep learning6 model using only routinely collected data that could help with the screening of potential organ donors. We present the development and validation of that model.
Methods:
The model was derived using the electronic health record data from intensive care unit (ICU) stays from Jan 1st 2012, until Dec 31th 2019, in the CHUM (Montreal University Hospital Center, Montreal). We defined our organ donor population as (1) organ donors that were identified in the study institution (2) potential organ donors referred to the ODO but deemed ineligible for transplantation and (3) non-referred potential organ donors identified from manual chart review death audit. We included all laboratory analyses as well as the presence/absence of head scan. We developed a two-step model. First, we trained a deep convolutional neural network7 autoencoder (CNN-AE)8. This model was a non-supervised model that served as an embedding and a dimensionality reduction tool, to reduce noise9 and for transfer learning and pretraining10. The second step was to add a deep classifier to categorize each patient as donor/non-donor. We also implemented a one-layer logistical model (LM), based on the last available value, as a comparator to our neural network (NN) model. The CNN-AE was trained on 85% randomly selected non-donors. The NN model and LM were trained on the rest of the patients, randomly separated in a train (60%), validation (20%) and test (20%) sets.
We compared our NN model to our LM using a ROC curve, scaled Brier score and calibration curves11 with bootstrap for confidence intervals. We manually reviewed the files of patients that were wrongly predicted as donor but who predicted a degree of confidence over 95% in the validation set.
Results:
Our complete dataset used 19717 patients with 392 donors. After excluding rare laboratory analysis, the NN model and LM were trained on 105 distinct laboratory analyses. On the test set, NN model and LM performed similarly with a ROC-AUC of respectively of 0.950 (95%CI 0.923-0.974) and 0.947 (95%CI 0.9169-0.9730) and scaled Brier score of 0.313 (0.134-0.472) and 0.432 (0.270-0.580). At best accuracy, sensitivity and specificity were 82% (95%CI 73-90) and 92% (95%CI 89-94) for the NN model and 82% (95%CI 73-90) and 94% (95%CI 91-95) for the LM. The NN model accuracy was more consistent across subgroups compared to the LM (fig 1).
Conclusion:
We present preliminary evidence that routinely collected medical data can be used to screen for potential organ donors. When comparing our more complex temporal model (NN) with its non-temporal simpler version (LM) both performed as well globally. However, the NN keeps a good accuracy in the more complex clinical patterns. Prospective and external validation of these models are required for further calibration before any potential clinical application.
Image 1
Image 2
1. Canadian Institute for Health Information. Annual Statistics on Organ Replacement in Canada: Dialysis, Transplantation and Donation, 2009 to 2018. (2019).
2. Krmpotic, K., Payne, C., Isenor, C. & Dhanani, S. Delayed Referral Results in Missed Opportunities for Organ Donation After Circulatory Death. Crit. Care Med. 45, 989–992 (2017).
3. Kutsogiannis, D. J., Asthana, S., Townsend, D. R., Singh, G. & Karvellas, C. J. The incidence of potential missed organ donors in intensive care units and emergency rooms: A retrospective cohort. Intensive Care Med. 39, 1452–1459 (2013).
4. Sairanen, T. et al. Lost potential of kidney and liver donors amongst deceased intracerebral hemorrhage patients. Eur. J. Neurol. (2014).
5. Opdam, H. I. & Silvester, W. Potential for organ donation in Victoria: An audit of hospital deaths. Med. J. Aust. 185, 250–254 (2006).
6. Beam, A. L. & Kohane, I. S. Big data and machine learning in health care. JAMA - Journal of the American Medical Association (2018).
7. He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (2016).
8. Hinton, G. E. & Salakhutdinov, R. R. Reducing the dimensionality of data with neural networks. Science (80-. ). (2006). doi:10.1126/science.1127647
9. Charte, D., Charte, F., del Jesus, M. J. & Herrera, F. A Showcase of the Use of Autoencoders in Feature Learning Applications. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 11487 LNCS, (2019).
10. Ferreira, M. F., Camacho, R. & Teixeira, L. F. Autoencoders as Weight Initialization of Deep Classification Networks for Cancer versus Cancer Studies. (2020).
11. Leisman, D. E. et al. Development and Reporting of Prediction Models. Crit. Care Med. (2020).
Detection of Potential Organ Donors; An Automatic Approach on Temporal Data
Nicolas Sauthier
Abstract
Discussion Forum (0)
ePoster
Topic: End-of-Life and Supportive Care
Sauthier, Nicolas, MD BEng MSc c.1;Bouchakri, Rima, PhD1; Carrier, François-Martin, MD MSc1; Chassé, Michaël, MD PhD1
1 Centre de Recherche du Centre Hospitalier de l'Université de Montréal (CRCHUM), Montréal, Canada
CRCHUM
Introduction:
Despite a slow improvement in total organ donors over the last 20 years1, organ donation is still not meeting the demand. Organ transplantation depends on potential organ donor identification which is still a major challenge. Retrospective studies show that we may miss between 30% and 50 % of potential organ donors2–5. By increasing the total number of potential donors referred to an Organ Donation Organization (ODO), the number of organ donors could substantially increase.
Objectives:
We aimed to develop a deep learning6 model using only routinely collected data that could help with the screening of potential organ donors. We present the development and validation of that model.
Methods:
The model was derived using the electronic health record data from intensive care unit (ICU) stays from Jan 1st 2012, until Dec 31th 2019, in the CHUM (Montreal University Hospital Center, Montreal). We defined our organ donor population as (1) organ donors that were identified in the study institution (2) potential organ donors referred to the ODO but deemed ineligible for transplantation and (3) non-referred potential organ donors identified from manual chart review death audit. We included all laboratory analyses as well as the presence/absence of head scan. We developed a two-step model. First, we trained a deep convolutional neural network7 autoencoder (CNN-AE)8. This model was a non-supervised model that served as an embedding and a dimensionality reduction tool, to reduce noise9 and for transfer learning and pretraining10. The second step was to add a deep classifier to categorize each patient as donor/non-donor. We also implemented a one-layer logistical model (LM), based on the last available value, as a comparator to our neural network (NN) model. The CNN-AE was trained on 85% randomly selected non-donors. The NN model and LM were trained on the rest of the patients, randomly separated in a train (60%), validation (20%) and test (20%) sets.
We compared our NN model to our LM using a ROC curve, scaled Brier score and calibration curves11 with bootstrap for confidence intervals. We manually reviewed the files of patients that were wrongly predicted as donor but who predicted a degree of confidence over 95% in the validation set.
Results:
Our complete dataset used 19717 patients with 392 donors. After excluding rare laboratory analysis, the NN model and LM were trained on 105 distinct laboratory analyses. On the test set, NN model and LM performed similarly with a ROC-AUC of respectively of 0.950 (95%CI 0.923-0.974) and 0.947 (95%CI 0.9169-0.9730) and scaled Brier score of 0.313 (0.134-0.472) and 0.432 (0.270-0.580). At best accuracy, sensitivity and specificity were 82% (95%CI 73-90) and 92% (95%CI 89-94) for the NN model and 82% (95%CI 73-90) and 94% (95%CI 91-95) for the LM. The NN model accuracy was more consistent across subgroups compared to the LM (fig 1).
Conclusion:
We present preliminary evidence that routinely collected medical data can be used to screen for potential organ donors. When comparing our more complex temporal model (NN) with its non-temporal simpler version (LM) both performed as well globally. However, the NN keeps a good accuracy in the more complex clinical patterns. Prospective and external validation of these models are required for further calibration before any potential clinical application.
Image 1
Image 2
1. Canadian Institute for Health Information. Annual Statistics on Organ Replacement in Canada: Dialysis, Transplantation and Donation, 2009 to 2018. (2019).
2. Krmpotic, K., Payne, C., Isenor, C. & Dhanani, S. Delayed Referral Results in Missed Opportunities for Organ Donation After Circulatory Death. Crit. Care Med. 45, 989–992 (2017).
3. Kutsogiannis, D. J., Asthana, S., Townsend, D. R., Singh, G. & Karvellas, C. J. The incidence of potential missed organ donors in intensive care units and emergency rooms: A retrospective cohort. Intensive Care Med. 39, 1452–1459 (2013).
4. Sairanen, T. et al. Lost potential of kidney and liver donors amongst deceased intracerebral hemorrhage patients. Eur. J. Neurol. (2014).
5. Opdam, H. I. & Silvester, W. Potential for organ donation in Victoria: An audit of hospital deaths. Med. J. Aust. 185, 250–254 (2006).
6. Beam, A. L. & Kohane, I. S. Big data and machine learning in health care. JAMA - Journal of the American Medical Association (2018).
7. He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (2016).
8. Hinton, G. E. & Salakhutdinov, R. R. Reducing the dimensionality of data with neural networks. Science (80-. ). (2006). doi:10.1126/science.1127647
9. Charte, D., Charte, F., del Jesus, M. J. & Herrera, F. A Showcase of the Use of Autoencoders in Feature Learning Applications. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 11487 LNCS, (2019).
10. Ferreira, M. F., Camacho, R. & Teixeira, L. F. Autoencoders as Weight Initialization of Deep Classification Networks for Cancer versus Cancer Studies. (2020).
11. Leisman, D. E. et al. Development and Reporting of Prediction Models. Crit. Care Med. (2020).
Topic: End-of-Life and Supportive Care
Sauthier, Nicolas, MD BEng MSc c.1;Bouchakri, Rima, PhD1; Carrier, François-Martin, MD MSc1; Chassé, Michaël, MD PhD1
1 Centre de Recherche du Centre Hospitalier de l'Université de Montréal (CRCHUM), Montréal, Canada
CRCHUM
Introduction:
Despite a slow improvement in total organ donors over the last 20 years1, organ donation is still not meeting the demand. Organ transplantation depends on potential organ donor identification which is still a major challenge. Retrospective studies show that we may miss between 30% and 50 % of potential organ donors2–5. By increasing the total number of potential donors referred to an Organ Donation Organization (ODO), the number of organ donors could substantially increase.
Objectives:
We aimed to develop a deep learning6 model using only routinely collected data that could help with the screening of potential organ donors. We present the development and validation of that model.
Methods:
The model was derived using the electronic health record data from intensive care unit (ICU) stays from Jan 1st 2012, until Dec 31th 2019, in the CHUM (Montreal University Hospital Center, Montreal). We defined our organ donor population as (1) organ donors that were identified in the study institution (2) potential organ donors referred to the ODO but deemed ineligible for transplantation and (3) non-referred potential organ donors identified from manual chart review death audit. We included all laboratory analyses as well as the presence/absence of head scan. We developed a two-step model. First, we trained a deep convolutional neural network7 autoencoder (CNN-AE)8. This model was a non-supervised model that served as an embedding and a dimensionality reduction tool, to reduce noise9 and for transfer learning and pretraining10. The second step was to add a deep classifier to categorize each patient as donor/non-donor. We also implemented a one-layer logistical model (LM), based on the last available value, as a comparator to our neural network (NN) model. The CNN-AE was trained on 85% randomly selected non-donors. The NN model and LM were trained on the rest of the patients, randomly separated in a train (60%), validation (20%) and test (20%) sets.
We compared our NN model to our LM using a ROC curve, scaled Brier score and calibration curves11 with bootstrap for confidence intervals. We manually reviewed the files of patients that were wrongly predicted as donor but who predicted a degree of confidence over 95% in the validation set.
Results:
Our complete dataset used 19717 patients with 392 donors. After excluding rare laboratory analysis, the NN model and LM were trained on 105 distinct laboratory analyses. On the test set, NN model and LM performed similarly with a ROC-AUC of respectively of 0.950 (95%CI 0.923-0.974) and 0.947 (95%CI 0.9169-0.9730) and scaled Brier score of 0.313 (0.134-0.472) and 0.432 (0.270-0.580). At best accuracy, sensitivity and specificity were 82% (95%CI 73-90) and 92% (95%CI 89-94) for the NN model and 82% (95%CI 73-90) and 94% (95%CI 91-95) for the LM. The NN model accuracy was more consistent across subgroups compared to the LM (fig 1).
Conclusion:
We present preliminary evidence that routinely collected medical data can be used to screen for potential organ donors. When comparing our more complex temporal model (NN) with its non-temporal simpler version (LM) both performed as well globally. However, the NN keeps a good accuracy in the more complex clinical patterns. Prospective and external validation of these models are required for further calibration before any potential clinical application.
Image 1
Image 2
1. Canadian Institute for Health Information. Annual Statistics on Organ Replacement in Canada: Dialysis, Transplantation and Donation, 2009 to 2018. (2019).
2. Krmpotic, K., Payne, C., Isenor, C. & Dhanani, S. Delayed Referral Results in Missed Opportunities for Organ Donation After Circulatory Death. Crit. Care Med. 45, 989–992 (2017).
3. Kutsogiannis, D. J., Asthana, S., Townsend, D. R., Singh, G. & Karvellas, C. J. The incidence of potential missed organ donors in intensive care units and emergency rooms: A retrospective cohort. Intensive Care Med. 39, 1452–1459 (2013).
4. Sairanen, T. et al. Lost potential of kidney and liver donors amongst deceased intracerebral hemorrhage patients. Eur. J. Neurol. (2014).
5. Opdam, H. I. & Silvester, W. Potential for organ donation in Victoria: An audit of hospital deaths. Med. J. Aust. 185, 250–254 (2006).
6. Beam, A. L. & Kohane, I. S. Big data and machine learning in health care. JAMA - Journal of the American Medical Association (2018).
7. He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (2016).
8. Hinton, G. E. & Salakhutdinov, R. R. Reducing the dimensionality of data with neural networks. Science (80-. ). (2006). doi:10.1126/science.1127647
9. Charte, D., Charte, F., del Jesus, M. J. & Herrera, F. A Showcase of the Use of Autoencoders in Feature Learning Applications. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 11487 LNCS, (2019).
10. Ferreira, M. F., Camacho, R. & Teixeira, L. F. Autoencoders as Weight Initialization of Deep Classification Networks for Cancer versus Cancer Studies. (2020).
11. Leisman, D. E. et al. Development and Reporting of Prediction Models. Crit. Care Med. (2020).
ePoster
Topic: End-of-Life and Supportive Care
Sauthier, Nicolas, MD BEng MSc c.1;Bouchakri, Rima, PhD1; Carrier, François-Martin, MD MSc1; Chassé, Michaël, MD PhD1
1 Centre de Recherche du Centre Hospitalier de l'Université de Montréal (CRCHUM), Montréal, Canada
CRCHUM
Introduction:
Despite a slow improvement in total organ donors over the last 20 years1, organ donation is still not meeting the demand. Organ transplantation depends on potential organ donor identification which is still a major challenge. Retrospective studies show that we may miss between 30% and 50 % of potential organ donors2–5. By increasing the total number of potential donors referred to an Organ Donation Organization (ODO), the number of organ donors could substantially increase.
Objectives:
We aimed to develop a deep learning6 model using only routinely collected data that could help with the screening of potential organ donors. We present the development and validation of that model.
Methods:
The model was derived using the electronic health record data from intensive care unit (ICU) stays from Jan 1st 2012, until Dec 31th 2019, in the CHUM (Montreal University Hospital Center, Montreal). We defined our organ donor population as (1) organ donors that were identified in the study institution (2) potential organ donors referred to the ODO but deemed ineligible for transplantation and (3) non-referred potential organ donors identified from manual chart review death audit. We included all laboratory analyses as well as the presence/absence of head scan. We developed a two-step model. First, we trained a deep convolutional neural network7 autoencoder (CNN-AE)8. This model was a non-supervised model that served as an embedding and a dimensionality reduction tool, to reduce noise9 and for transfer learning and pretraining10. The second step was to add a deep classifier to categorize each patient as donor/non-donor. We also implemented a one-layer logistical model (LM), based on the last available value, as a comparator to our neural network (NN) model. The CNN-AE was trained on 85% randomly selected non-donors. The NN model and LM were trained on the rest of the patients, randomly separated in a train (60%), validation (20%) and test (20%) sets.
We compared our NN model to our LM using a ROC curve, scaled Brier score and calibration curves11 with bootstrap for confidence intervals. We manually reviewed the files of patients that were wrongly predicted as donor but who predicted a degree of confidence over 95% in the validation set.
Results:
Our complete dataset used 19717 patients with 392 donors. After excluding rare laboratory analysis, the NN model and LM were trained on 105 distinct laboratory analyses. On the test set, NN model and LM performed similarly with a ROC-AUC of respectively of 0.950 (95%CI 0.923-0.974) and 0.947 (95%CI 0.9169-0.9730) and scaled Brier score of 0.313 (0.134-0.472) and 0.432 (0.270-0.580). At best accuracy, sensitivity and specificity were 82% (95%CI 73-90) and 92% (95%CI 89-94) for the NN model and 82% (95%CI 73-90) and 94% (95%CI 91-95) for the LM. The NN model accuracy was more consistent across subgroups compared to the LM (fig 1).
Conclusion:
We present preliminary evidence that routinely collected medical data can be used to screen for potential organ donors. When comparing our more complex temporal model (NN) with its non-temporal simpler version (LM) both performed as well globally. However, the NN keeps a good accuracy in the more complex clinical patterns. Prospective and external validation of these models are required for further calibration before any potential clinical application.
Image 1
Image 2
1. Canadian Institute for Health Information. Annual Statistics on Organ Replacement in Canada: Dialysis, Transplantation and Donation, 2009 to 2018. (2019).
2. Krmpotic, K., Payne, C., Isenor, C. & Dhanani, S. Delayed Referral Results in Missed Opportunities for Organ Donation After Circulatory Death. Crit. Care Med. 45, 989–992 (2017).
3. Kutsogiannis, D. J., Asthana, S., Townsend, D. R., Singh, G. & Karvellas, C. J. The incidence of potential missed organ donors in intensive care units and emergency rooms: A retrospective cohort. Intensive Care Med. 39, 1452–1459 (2013).
4. Sairanen, T. et al. Lost potential of kidney and liver donors amongst deceased intracerebral hemorrhage patients. Eur. J. Neurol. (2014).
5. Opdam, H. I. & Silvester, W. Potential for organ donation in Victoria: An audit of hospital deaths. Med. J. Aust. 185, 250–254 (2006).
6. Beam, A. L. & Kohane, I. S. Big data and machine learning in health care. JAMA - Journal of the American Medical Association (2018).
7. He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (2016).
8. Hinton, G. E. & Salakhutdinov, R. R. Reducing the dimensionality of data with neural networks. Science (80-. ). (2006). doi:10.1126/science.1127647
9. Charte, D., Charte, F., del Jesus, M. J. & Herrera, F. A Showcase of the Use of Autoencoders in Feature Learning Applications. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 11487 LNCS, (2019).
10. Ferreira, M. F., Camacho, R. & Teixeira, L. F. Autoencoders as Weight Initialization of Deep Classification Networks for Cancer versus Cancer Studies. (2020).
11. Leisman, D. E. et al. Development and Reporting of Prediction Models. Crit. Care Med. (2020).
Topic: End-of-Life and Supportive Care
Sauthier, Nicolas, MD BEng MSc c.1;Bouchakri, Rima, PhD1; Carrier, François-Martin, MD MSc1; Chassé, Michaël, MD PhD1
1 Centre de Recherche du Centre Hospitalier de l'Université de Montréal (CRCHUM), Montréal, Canada
CRCHUM
Introduction:
Despite a slow improvement in total organ donors over the last 20 years1, organ donation is still not meeting the demand. Organ transplantation depends on potential organ donor identification which is still a major challenge. Retrospective studies show that we may miss between 30% and 50 % of potential organ donors2–5. By increasing the total number of potential donors referred to an Organ Donation Organization (ODO), the number of organ donors could substantially increase.
Objectives:
We aimed to develop a deep learning6 model using only routinely collected data that could help with the screening of potential organ donors. We present the development and validation of that model.
Methods:
The model was derived using the electronic health record data from intensive care unit (ICU) stays from Jan 1st 2012, until Dec 31th 2019, in the CHUM (Montreal University Hospital Center, Montreal). We defined our organ donor population as (1) organ donors that were identified in the study institution (2) potential organ donors referred to the ODO but deemed ineligible for transplantation and (3) non-referred potential organ donors identified from manual chart review death audit. We included all laboratory analyses as well as the presence/absence of head scan. We developed a two-step model. First, we trained a deep convolutional neural network7 autoencoder (CNN-AE)8. This model was a non-supervised model that served as an embedding and a dimensionality reduction tool, to reduce noise9 and for transfer learning and pretraining10. The second step was to add a deep classifier to categorize each patient as donor/non-donor. We also implemented a one-layer logistical model (LM), based on the last available value, as a comparator to our neural network (NN) model. The CNN-AE was trained on 85% randomly selected non-donors. The NN model and LM were trained on the rest of the patients, randomly separated in a train (60%), validation (20%) and test (20%) sets.
We compared our NN model to our LM using a ROC curve, scaled Brier score and calibration curves11 with bootstrap for confidence intervals. We manually reviewed the files of patients that were wrongly predicted as donor but who predicted a degree of confidence over 95% in the validation set.
Results:
Our complete dataset used 19717 patients with 392 donors. After excluding rare laboratory analysis, the NN model and LM were trained on 105 distinct laboratory analyses. On the test set, NN model and LM performed similarly with a ROC-AUC of respectively of 0.950 (95%CI 0.923-0.974) and 0.947 (95%CI 0.9169-0.9730) and scaled Brier score of 0.313 (0.134-0.472) and 0.432 (0.270-0.580). At best accuracy, sensitivity and specificity were 82% (95%CI 73-90) and 92% (95%CI 89-94) for the NN model and 82% (95%CI 73-90) and 94% (95%CI 91-95) for the LM. The NN model accuracy was more consistent across subgroups compared to the LM (fig 1).
Conclusion:
We present preliminary evidence that routinely collected medical data can be used to screen for potential organ donors. When comparing our more complex temporal model (NN) with its non-temporal simpler version (LM) both performed as well globally. However, the NN keeps a good accuracy in the more complex clinical patterns. Prospective and external validation of these models are required for further calibration before any potential clinical application.
Image 1
Image 2
1. Canadian Institute for Health Information. Annual Statistics on Organ Replacement in Canada: Dialysis, Transplantation and Donation, 2009 to 2018. (2019).
2. Krmpotic, K., Payne, C., Isenor, C. & Dhanani, S. Delayed Referral Results in Missed Opportunities for Organ Donation After Circulatory Death. Crit. Care Med. 45, 989–992 (2017).
3. Kutsogiannis, D. J., Asthana, S., Townsend, D. R., Singh, G. & Karvellas, C. J. The incidence of potential missed organ donors in intensive care units and emergency rooms: A retrospective cohort. Intensive Care Med. 39, 1452–1459 (2013).
4. Sairanen, T. et al. Lost potential of kidney and liver donors amongst deceased intracerebral hemorrhage patients. Eur. J. Neurol. (2014).
5. Opdam, H. I. & Silvester, W. Potential for organ donation in Victoria: An audit of hospital deaths. Med. J. Aust. 185, 250–254 (2006).
6. Beam, A. L. & Kohane, I. S. Big data and machine learning in health care. JAMA - Journal of the American Medical Association (2018).
7. He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (2016).
8. Hinton, G. E. & Salakhutdinov, R. R. Reducing the dimensionality of data with neural networks. Science (80-. ). (2006). doi:10.1126/science.1127647
9. Charte, D., Charte, F., del Jesus, M. J. & Herrera, F. A Showcase of the Use of Autoencoders in Feature Learning Applications. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 11487 LNCS, (2019).
10. Ferreira, M. F., Camacho, R. & Teixeira, L. F. Autoencoders as Weight Initialization of Deep Classification Networks for Cancer versus Cancer Studies. (2020).
11. Leisman, D. E. et al. Development and Reporting of Prediction Models. Crit. Care Med. (2020).
{{ help_message }}
{{filter}}