Digital Twin Model Predicts Sepsis Mortality

A “digital twin” model successfully predicted adverse outcomes in intensive care unit (ICU) patients treated for sepsis. The research used an adaptive approach, updating time-sensitive values such as blood pressure and vitals every 15 minutes. The approach also took into account treatment decisions and has potential as a decision-making and educational tool.

The digital twin could reduce the risk for some interventions, according to Amos Lal, MD, who presented the study at the CHEST Annual Meeting. That’s because the model can predict the outcome. “You don’t actually have to make an intervention to the patient, which might be risky. By doing that, you can actually prevent a lot of harm,” said Dr. Lal, assistant professor of medicine at Mayo Clinic in Rochester, Minnesota.

The researchers used a one-dimensional convolutional neural network (CNN), similar to two-dimensional CNNs that are used to classify images, substituting the color channels used in imaging with 38 time-dependent variables. They applied it to predicting outcomes in the ICU, focusing on data generated within the first 24 hours of admission. The team made the model dynamic by adding time-sensitive data like vitals, laboratory values, and interventions every 15 minutes. That contrasts with existing models that are usually static, relying on values at admission or at 24 hours, for example. It also takes into account time-insensitive data like age, gender, and comorbidities. “Combining these two and coming up with the prediction model in real time can give you a more informed decision about how these patients are going to perform over a period of 2 weeks or 4 weeks of their stay within the ICU. And of course, as we get more and more data within the first 24 hours, the performance of the model improves as well,” said Dr. Lal.

The researchers tested the model by creating a virtual model of the patient and then performing an intervention on the patient and a simulated intervention on the virtual patient. “Then we advance the clock and the patient either improved or deteriorated, and we compared how the digital twin performed, whether the changes were concordant or discordant [between the virtual and real-world patients],” said Dr. Lal.

The model was designed to predict which patients with sepsis would be at greater risk for death or ICU stays longer than 14 days. It was created using data from 28,617 patients with critical care sepsis at a single hospital who were treated between 2011 and 2018, with 70% used as a training set, 20% as a test set, and 10% as a validation set. The researchers conducted an external validation using MIMIC-IV data on 30,903 patients from the Beth Israel Deaconess Medical Center in Boston. The model included 31 time-independent variables and 38 time-dependent variables that were collected every 15 minutes at the Mayo Clinic and every 60 minutes at Beth Israel Deaconess. Surgical patients represented 24% of the Mayo dataset and 58% of the MIMIC-IV dataset, but otherwise the two groups were demographically similar.

At 24 hours, the area under the receiver operating characteristic curve for predicting 14-day mortality was −0.82 in the Mayo validation cohort and −0.78 in the MIMIC validation cohort. The model improved in accuracy over time as more data were accumulated.

The session’s co-moderators, Sandeep Jain, MD, and Casey Cable, MD, praised the work. Dr. Cable, associate professor of pulmonary care medicine at VCU Health, Richmond, Virginia, noted that the model used both surgical patients and medical patients with sepsis, and the two groups can present quite differently. Another variable was the COVID pandemic, where some patients presented at the hospital when they were quite sick. “I’m curious how different starting points would play into it,” she said.

She called for institutions to develop such models on their own rather than relying on companies that might develop software solutions. “I think that this needs to be clinician-led, from the ground up,” said Dr. Cable.

Dr. Jain, an associate professor of pulmonary care medicine at Broward Health, suggested that such models might need to be individualized for each institution, but “my fear is it could become too expensive, so I think a group like CHEST could come together and [create] an open source system to have their researchers jumpstart the research on this,” he said.

Dr. Lal, Dr. Jain, and Dr. Cable reported no relevant financial relationships.

A version of this article appeared on Medscape.com.

A “digital twin” model successfully predicted adverse outcomes in intensive care unit (ICU) patients treated for sepsis. The research used an adaptive approach, updating time-sensitive values such as blood pressure and vitals every 15 minutes. The approach also took into account treatment decisions and has potential as a decision-making and educational tool.

The digital twin could reduce the risk for some interventions, according to Amos Lal, MD, who presented the study at the CHEST Annual Meeting. That’s because the model can predict the outcome. “You don’t actually have to make an intervention to the patient, which might be risky. By doing that, you can actually prevent a lot of harm,” said Dr. Lal, assistant professor of medicine at Mayo Clinic in Rochester, Minnesota.

The researchers used a one-dimensional convolutional neural network (CNN), similar to two-dimensional CNNs that are used to classify images, substituting the color channels used in imaging with 38 time-dependent variables. They applied it to predicting outcomes in the ICU, focusing on data generated within the first 24 hours of admission. The team made the model dynamic by adding time-sensitive data like vitals, laboratory values, and interventions every 15 minutes. That contrasts with existing models that are usually static, relying on values at admission or at 24 hours, for example. It also takes into account time-insensitive data like age, gender, and comorbidities. “Combining these two and coming up with the prediction model in real time can give you a more informed decision about how these patients are going to perform over a period of 2 weeks or 4 weeks of their stay within the ICU. And of course, as we get more and more data within the first 24 hours, the performance of the model improves as well,” said Dr. Lal.

The researchers tested the model by creating a virtual model of the patient and then performing an intervention on the patient and a simulated intervention on the virtual patient. “Then we advance the clock and the patient either improved or deteriorated, and we compared how the digital twin performed, whether the changes were concordant or discordant [between the virtual and real-world patients],” said Dr. Lal.

The model was designed to predict which patients with sepsis would be at greater risk for death or ICU stays longer than 14 days. It was created using data from 28,617 patients with critical care sepsis at a single hospital who were treated between 2011 and 2018, with 70% used as a training set, 20% as a test set, and 10% as a validation set. The researchers conducted an external validation using MIMIC-IV data on 30,903 patients from the Beth Israel Deaconess Medical Center in Boston. The model included 31 time-independent variables and 38 time-dependent variables that were collected every 15 minutes at the Mayo Clinic and every 60 minutes at Beth Israel Deaconess. Surgical patients represented 24% of the Mayo dataset and 58% of the MIMIC-IV dataset, but otherwise the two groups were demographically similar.

At 24 hours, the area under the receiver operating characteristic curve for predicting 14-day mortality was −0.82 in the Mayo validation cohort and −0.78 in the MIMIC validation cohort. The model improved in accuracy over time as more data were accumulated.

The session’s co-moderators, Sandeep Jain, MD, and Casey Cable, MD, praised the work. Dr. Cable, associate professor of pulmonary care medicine at VCU Health, Richmond, Virginia, noted that the model used both surgical patients and medical patients with sepsis, and the two groups can present quite differently. Another variable was the COVID pandemic, where some patients presented at the hospital when they were quite sick. “I’m curious how different starting points would play into it,” she said.

She called for institutions to develop such models on their own rather than relying on companies that might develop software solutions. “I think that this needs to be clinician-led, from the ground up,” said Dr. Cable.

Dr. Jain, an associate professor of pulmonary care medicine at Broward Health, suggested that such models might need to be individualized for each institution, but “my fear is it could become too expensive, so I think a group like CHEST could come together and [create] an open source system to have their researchers jumpstart the research on this,” he said.

Dr. Lal, Dr. Jain, and Dr. Cable reported no relevant financial relationships.

A version of this article appeared on Medscape.com.

A “digital twin” model successfully predicted adverse outcomes in intensive care unit (ICU) patients treated for sepsis. The research used an adaptive approach, updating time-sensitive values such as blood pressure and vitals every 15 minutes. The approach also took into account treatment decisions and has potential as a decision-making and educational tool.

The digital twin could reduce the risk for some interventions, according to Amos Lal, MD, who presented the study at the CHEST Annual Meeting. That’s because the model can predict the outcome. “You don’t actually have to make an intervention to the patient, which might be risky. By doing that, you can actually prevent a lot of harm,” said Dr. Lal, assistant professor of medicine at Mayo Clinic in Rochester, Minnesota.

The researchers used a one-dimensional convolutional neural network (CNN), similar to two-dimensional CNNs that are used to classify images, substituting the color channels used in imaging with 38 time-dependent variables. They applied it to predicting outcomes in the ICU, focusing on data generated within the first 24 hours of admission. The team made the model dynamic by adding time-sensitive data like vitals, laboratory values, and interventions every 15 minutes. That contrasts with existing models that are usually static, relying on values at admission or at 24 hours, for example. It also takes into account time-insensitive data like age, gender, and comorbidities. “Combining these two and coming up with the prediction model in real time can give you a more informed decision about how these patients are going to perform over a period of 2 weeks or 4 weeks of their stay within the ICU. And of course, as we get more and more data within the first 24 hours, the performance of the model improves as well,” said Dr. Lal.

The researchers tested the model by creating a virtual model of the patient and then performing an intervention on the patient and a simulated intervention on the virtual patient. “Then we advance the clock and the patient either improved or deteriorated, and we compared how the digital twin performed, whether the changes were concordant or discordant [between the virtual and real-world patients],” said Dr. Lal.

The model was designed to predict which patients with sepsis would be at greater risk for death or ICU stays longer than 14 days. It was created using data from 28,617 patients with critical care sepsis at a single hospital who were treated between 2011 and 2018, with 70% used as a training set, 20% as a test set, and 10% as a validation set. The researchers conducted an external validation using MIMIC-IV data on 30,903 patients from the Beth Israel Deaconess Medical Center in Boston. The model included 31 time-independent variables and 38 time-dependent variables that were collected every 15 minutes at the Mayo Clinic and every 60 minutes at Beth Israel Deaconess. Surgical patients represented 24% of the Mayo dataset and 58% of the MIMIC-IV dataset, but otherwise the two groups were demographically similar.

At 24 hours, the area under the receiver operating characteristic curve for predicting 14-day mortality was −0.82 in the Mayo validation cohort and −0.78 in the MIMIC validation cohort. The model improved in accuracy over time as more data were accumulated.

The session’s co-moderators, Sandeep Jain, MD, and Casey Cable, MD, praised the work. Dr. Cable, associate professor of pulmonary care medicine at VCU Health, Richmond, Virginia, noted that the model used both surgical patients and medical patients with sepsis, and the two groups can present quite differently. Another variable was the COVID pandemic, where some patients presented at the hospital when they were quite sick. “I’m curious how different starting points would play into it,” she said.

She called for institutions to develop such models on their own rather than relying on companies that might develop software solutions. “I think that this needs to be clinician-led, from the ground up,” said Dr. Cable.

Dr. Jain, an associate professor of pulmonary care medicine at Broward Health, suggested that such models might need to be individualized for each institution, but “my fear is it could become too expensive, so I think a group like CHEST could come together and [create] an open source system to have their researchers jumpstart the research on this,” he said.

Dr. Lal, Dr. Jain, and Dr. Cable reported no relevant financial relationships.

A version of this article appeared on Medscape.com.