User login
GI Genius recently became the first Food and Drug Administration–approved device to use artificial intelligence (AI) for endoscopy. Soon, similar technology may give gastroenterologists an edge before they even walk into the procedure room.
AI can provide highly accurate risk scores for patients with suspected upper GI bleeding, and make a recommendation for discharge or hospitalization, according to Dennis Shung, MD, MHS, a clinical instructor at Yale University, New Haven, Conn. And this could provide extensive benefit.
“Acute gastrointestinal bleeding is the most common gastrointestinal diagnosis requiring hospitalization. It costs around $19.2 billion per year,” Dr. Shung said, citing a study from Gastroenterology. He made these remarks during a virtual presentation at the 2021 AGA Tech Summit sponsored by the AGA Center for GI Innovation and Technology.
Emergency department visits for upper GI bleeding increased 17% from 2006 to 2014, Dr. Shung added, suggesting a rising trend.
The trouble with using risk scores
A variety of conventional risk scores are presently available to help manage these patients. Generally, they use a composite outcome of hemostatic intervention, transfusion, or death to determine which patients should be hospitalized (high risk) and which patients can go home (low risk). Although these models can offer high sensitivity, they remain underutilized.
“[Clinical risk scores] are cumbersome, it’s difficult to calculate them, [and] you may not remember to do that in your busy workflow,” Dr. Shung said.
He pointed out that low implementation may also stem from poorly defined clinical responsibilities.
“[Observing] providers caring for patients with GI bleeding showed that there was a culture of not taking ownership,” he said. “Emergency department physicians thought that it was the gastroenterologists who needed to [perform risk scoring]. Gastroenterologists thought it was the ED [physicians’ responsibility].”
To overcome these pitfalls, Dr. Shung and colleagues are developing AI that automates risk analysis for upper GI bleeding by integrating the process into the clinical workflow. Like GI Genius, their strategy relies upon machine learning, which is a type of AI that can improve automatically without being explicitly programmed.
Their most recent study (Sci Rep. 2021 Apr 23;11[1]:8827) involved a machine learning model that could predict transfusion in patients admitted for acute GI bleeding. The model was developed and internally validated in a cohort of 2,524 patients, then shown to outperform conventional regression-based models when externally validated in 1,526 patients similarly admitted at large urban hospitals.
Google Maps for GI bleeding
“The future, as I envision it, is a Google Maps for GI bleeding,” Dr. Shung said, referring to how the popular web-mapping product analyzes real-time data, such as weather and traffic patterns, to provide the best route and an estimated time of arrival. “With the electronic health record, we have the ability to personalize care by basically using data obtained during the clinical encounter to generate risk assessment in real time.”
In other words, machine learning software reads a patient’s electronic health record, runs relevant data through an algorithm, and produces both a risk score and a clinical recommendation. In the case of suspected upper GI bleeding, the clinician is advised to either discharge for outpatient endoscopy or hospitalize for inpatient evaluation.
Because the quality and consistency of data in EHRs can vary, the most advanced form of machine learning – deep learning – is needed to make this a clinical reality. Deep learning converts simpler concepts into complex ones. In this scenario, that would mean deciding which clinical data are relevant and which are just noise. Taking this a step further, deep learning can actually “draw conclusions” from what’s missing.
“There are huge challenges in [irregular data] that need to be overcome,” Dr. Shung said in an interview. “But I see it as an opportunity. When you see things that are irregularly sampled, when you see things are missing – they mean something. They mean that a human has decided that that is not the way we should do things because this patient doesn’t need it. And I think there is a lot of value in learning how to model those things.”
The road to clinical implementation
With further research and validation, deep learning models for gastroenterology are likely to play a role in clinical decision-making, according to Dr. Shung. But to reach the clinic floor, developers will need to outsmart some more fundamental obstacles. “The main thing that’s really barring [AI risk modeling] from being used is the reimbursement issue,” he said, referring to uncertainty in how payers will cover associated costs.
In an interview, Sushovan Guha, MD, PhD, moderator of the virtual session and codirector of the center for interventional gastroenterology at UTHealth (iGUT) in Houston, pointed out another financial unknown: liability.
“What happens if there is an error?” he asked. “It’s done by the computers, but who is at fault?”
In addition to these challenges, some clinicians may need to be persuaded before they are willing to trust an algorithm with a patient’s life.
“We have to have community physicians convinced about the importance of using these tools to further improve their clinical practice,” Dr. Guha said. To this end, he added, “It’s time for us to accept and adapt, and make our decision-making process much more efficient.”
The investigators disclosed no relevant conflicts of interest.
GI Genius recently became the first Food and Drug Administration–approved device to use artificial intelligence (AI) for endoscopy. Soon, similar technology may give gastroenterologists an edge before they even walk into the procedure room.
AI can provide highly accurate risk scores for patients with suspected upper GI bleeding, and make a recommendation for discharge or hospitalization, according to Dennis Shung, MD, MHS, a clinical instructor at Yale University, New Haven, Conn. And this could provide extensive benefit.
“Acute gastrointestinal bleeding is the most common gastrointestinal diagnosis requiring hospitalization. It costs around $19.2 billion per year,” Dr. Shung said, citing a study from Gastroenterology. He made these remarks during a virtual presentation at the 2021 AGA Tech Summit sponsored by the AGA Center for GI Innovation and Technology.
Emergency department visits for upper GI bleeding increased 17% from 2006 to 2014, Dr. Shung added, suggesting a rising trend.
The trouble with using risk scores
A variety of conventional risk scores are presently available to help manage these patients. Generally, they use a composite outcome of hemostatic intervention, transfusion, or death to determine which patients should be hospitalized (high risk) and which patients can go home (low risk). Although these models can offer high sensitivity, they remain underutilized.
“[Clinical risk scores] are cumbersome, it’s difficult to calculate them, [and] you may not remember to do that in your busy workflow,” Dr. Shung said.
He pointed out that low implementation may also stem from poorly defined clinical responsibilities.
“[Observing] providers caring for patients with GI bleeding showed that there was a culture of not taking ownership,” he said. “Emergency department physicians thought that it was the gastroenterologists who needed to [perform risk scoring]. Gastroenterologists thought it was the ED [physicians’ responsibility].”
To overcome these pitfalls, Dr. Shung and colleagues are developing AI that automates risk analysis for upper GI bleeding by integrating the process into the clinical workflow. Like GI Genius, their strategy relies upon machine learning, which is a type of AI that can improve automatically without being explicitly programmed.
Their most recent study (Sci Rep. 2021 Apr 23;11[1]:8827) involved a machine learning model that could predict transfusion in patients admitted for acute GI bleeding. The model was developed and internally validated in a cohort of 2,524 patients, then shown to outperform conventional regression-based models when externally validated in 1,526 patients similarly admitted at large urban hospitals.
Google Maps for GI bleeding
“The future, as I envision it, is a Google Maps for GI bleeding,” Dr. Shung said, referring to how the popular web-mapping product analyzes real-time data, such as weather and traffic patterns, to provide the best route and an estimated time of arrival. “With the electronic health record, we have the ability to personalize care by basically using data obtained during the clinical encounter to generate risk assessment in real time.”
In other words, machine learning software reads a patient’s electronic health record, runs relevant data through an algorithm, and produces both a risk score and a clinical recommendation. In the case of suspected upper GI bleeding, the clinician is advised to either discharge for outpatient endoscopy or hospitalize for inpatient evaluation.
Because the quality and consistency of data in EHRs can vary, the most advanced form of machine learning – deep learning – is needed to make this a clinical reality. Deep learning converts simpler concepts into complex ones. In this scenario, that would mean deciding which clinical data are relevant and which are just noise. Taking this a step further, deep learning can actually “draw conclusions” from what’s missing.
“There are huge challenges in [irregular data] that need to be overcome,” Dr. Shung said in an interview. “But I see it as an opportunity. When you see things that are irregularly sampled, when you see things are missing – they mean something. They mean that a human has decided that that is not the way we should do things because this patient doesn’t need it. And I think there is a lot of value in learning how to model those things.”
The road to clinical implementation
With further research and validation, deep learning models for gastroenterology are likely to play a role in clinical decision-making, according to Dr. Shung. But to reach the clinic floor, developers will need to outsmart some more fundamental obstacles. “The main thing that’s really barring [AI risk modeling] from being used is the reimbursement issue,” he said, referring to uncertainty in how payers will cover associated costs.
In an interview, Sushovan Guha, MD, PhD, moderator of the virtual session and codirector of the center for interventional gastroenterology at UTHealth (iGUT) in Houston, pointed out another financial unknown: liability.
“What happens if there is an error?” he asked. “It’s done by the computers, but who is at fault?”
In addition to these challenges, some clinicians may need to be persuaded before they are willing to trust an algorithm with a patient’s life.
“We have to have community physicians convinced about the importance of using these tools to further improve their clinical practice,” Dr. Guha said. To this end, he added, “It’s time for us to accept and adapt, and make our decision-making process much more efficient.”
The investigators disclosed no relevant conflicts of interest.
GI Genius recently became the first Food and Drug Administration–approved device to use artificial intelligence (AI) for endoscopy. Soon, similar technology may give gastroenterologists an edge before they even walk into the procedure room.
AI can provide highly accurate risk scores for patients with suspected upper GI bleeding, and make a recommendation for discharge or hospitalization, according to Dennis Shung, MD, MHS, a clinical instructor at Yale University, New Haven, Conn. And this could provide extensive benefit.
“Acute gastrointestinal bleeding is the most common gastrointestinal diagnosis requiring hospitalization. It costs around $19.2 billion per year,” Dr. Shung said, citing a study from Gastroenterology. He made these remarks during a virtual presentation at the 2021 AGA Tech Summit sponsored by the AGA Center for GI Innovation and Technology.
Emergency department visits for upper GI bleeding increased 17% from 2006 to 2014, Dr. Shung added, suggesting a rising trend.
The trouble with using risk scores
A variety of conventional risk scores are presently available to help manage these patients. Generally, they use a composite outcome of hemostatic intervention, transfusion, or death to determine which patients should be hospitalized (high risk) and which patients can go home (low risk). Although these models can offer high sensitivity, they remain underutilized.
“[Clinical risk scores] are cumbersome, it’s difficult to calculate them, [and] you may not remember to do that in your busy workflow,” Dr. Shung said.
He pointed out that low implementation may also stem from poorly defined clinical responsibilities.
“[Observing] providers caring for patients with GI bleeding showed that there was a culture of not taking ownership,” he said. “Emergency department physicians thought that it was the gastroenterologists who needed to [perform risk scoring]. Gastroenterologists thought it was the ED [physicians’ responsibility].”
To overcome these pitfalls, Dr. Shung and colleagues are developing AI that automates risk analysis for upper GI bleeding by integrating the process into the clinical workflow. Like GI Genius, their strategy relies upon machine learning, which is a type of AI that can improve automatically without being explicitly programmed.
Their most recent study (Sci Rep. 2021 Apr 23;11[1]:8827) involved a machine learning model that could predict transfusion in patients admitted for acute GI bleeding. The model was developed and internally validated in a cohort of 2,524 patients, then shown to outperform conventional regression-based models when externally validated in 1,526 patients similarly admitted at large urban hospitals.
Google Maps for GI bleeding
“The future, as I envision it, is a Google Maps for GI bleeding,” Dr. Shung said, referring to how the popular web-mapping product analyzes real-time data, such as weather and traffic patterns, to provide the best route and an estimated time of arrival. “With the electronic health record, we have the ability to personalize care by basically using data obtained during the clinical encounter to generate risk assessment in real time.”
In other words, machine learning software reads a patient’s electronic health record, runs relevant data through an algorithm, and produces both a risk score and a clinical recommendation. In the case of suspected upper GI bleeding, the clinician is advised to either discharge for outpatient endoscopy or hospitalize for inpatient evaluation.
Because the quality and consistency of data in EHRs can vary, the most advanced form of machine learning – deep learning – is needed to make this a clinical reality. Deep learning converts simpler concepts into complex ones. In this scenario, that would mean deciding which clinical data are relevant and which are just noise. Taking this a step further, deep learning can actually “draw conclusions” from what’s missing.
“There are huge challenges in [irregular data] that need to be overcome,” Dr. Shung said in an interview. “But I see it as an opportunity. When you see things that are irregularly sampled, when you see things are missing – they mean something. They mean that a human has decided that that is not the way we should do things because this patient doesn’t need it. And I think there is a lot of value in learning how to model those things.”
The road to clinical implementation
With further research and validation, deep learning models for gastroenterology are likely to play a role in clinical decision-making, according to Dr. Shung. But to reach the clinic floor, developers will need to outsmart some more fundamental obstacles. “The main thing that’s really barring [AI risk modeling] from being used is the reimbursement issue,” he said, referring to uncertainty in how payers will cover associated costs.
In an interview, Sushovan Guha, MD, PhD, moderator of the virtual session and codirector of the center for interventional gastroenterology at UTHealth (iGUT) in Houston, pointed out another financial unknown: liability.
“What happens if there is an error?” he asked. “It’s done by the computers, but who is at fault?”
In addition to these challenges, some clinicians may need to be persuaded before they are willing to trust an algorithm with a patient’s life.
“We have to have community physicians convinced about the importance of using these tools to further improve their clinical practice,” Dr. Guha said. To this end, he added, “It’s time for us to accept and adapt, and make our decision-making process much more efficient.”
The investigators disclosed no relevant conflicts of interest.
FROM 2021 AGA TECH SUMMIT