User login
BALTIMORE – Neural networks are the building blocks of machine learning and artificial intelligence, and researchers from the University of Minnesota have identified a panel of “simple, readily known” preoperative patient factors that they fed into an artificial neural network model that can be predictive of 30-day outcomes after laparoscopic sleeve gastrectomy, one of the researchers reported at the annual meeting of the Society of American Gastrointestinal and Endoscopic Surgeons.
“The biggest limitation to using neural networks clinically is the fact that they’re algorithmic complex,” said Eric S. Wise, MD, of the University of Minnesota, Minneapolis, in presenting the research. “There is an underlying algorithm that’s developed, but it’s very difficult to understand.” He called it “a black box problem.”
Nonetheless, the researchers drew upon 101,721 laparoscopic sleeve gastrectomy cases from the 2016 Metabolic and Bariatric Surgery Accreditation and Quality Improvement Program national database to extract factors that were associated with postoperative complications. “More pertinently, we wanted to optimize predictability of a panel of readily obtainable, easily qualifiable preoperative factors and maximize the variants that are contained within those variables to predict the outcome of 30-day morbidity and mortality,” Dr. Wise said.
Essentially, neural networks recognize patterns through a machine-learning process in a manner modeled on the human brain. As Dr. Wise explained, they first emerged in the 1960s to simulate the human brain’s psychological-neurologic systems.
Through bivariate and multivariate analyses, the research identified eight preoperative variables strongly associated with the 30-day endpoints. After univariate analysis, seven of those variables were statistically significant: older age (P = .03), nonwhite race, higher initial body mass index, severe hypertension, history of diabetes, nonindependent functional status, and previous foregut/bariatric surgery (all P less than .001). “Gender was the only factor that was not predictive,” Dr. Wise said.
The factors held up under logistic regression modeling. “We were able to use a traditional logistic regression model that came up with a reasonable area under the curve of 0.572,” he said. Using artificial neural network analysis, the training set, which comprised 80% of patients, was more accurate than logistic regression, with an area under the curve of 0.582.
One limitation was that this was a “small study,” Dr. Wise said, influenced by selection bias inherent in any retrospective data selection. Other major factors that may exist were not considered.
However, he noted, “in the past we’ve had some success translating neural networks into something that’s clinically useful.” His group at Vanderbilt University published a report of artificial neural network modeling to identify five factors predictive of weight loss after Roux-en-Y gastric bypass 2 years ago (Surg Endosc. 2016;30:480-8). “There are ways to translate neural networks clinically,” Dr. Wise said.
Dr. Wise had no financial relationships to disclose.
SOURCE: Wise ES et al. SAGES 2019, Abstract S053.
BALTIMORE – Neural networks are the building blocks of machine learning and artificial intelligence, and researchers from the University of Minnesota have identified a panel of “simple, readily known” preoperative patient factors that they fed into an artificial neural network model that can be predictive of 30-day outcomes after laparoscopic sleeve gastrectomy, one of the researchers reported at the annual meeting of the Society of American Gastrointestinal and Endoscopic Surgeons.
“The biggest limitation to using neural networks clinically is the fact that they’re algorithmic complex,” said Eric S. Wise, MD, of the University of Minnesota, Minneapolis, in presenting the research. “There is an underlying algorithm that’s developed, but it’s very difficult to understand.” He called it “a black box problem.”
Nonetheless, the researchers drew upon 101,721 laparoscopic sleeve gastrectomy cases from the 2016 Metabolic and Bariatric Surgery Accreditation and Quality Improvement Program national database to extract factors that were associated with postoperative complications. “More pertinently, we wanted to optimize predictability of a panel of readily obtainable, easily qualifiable preoperative factors and maximize the variants that are contained within those variables to predict the outcome of 30-day morbidity and mortality,” Dr. Wise said.
Essentially, neural networks recognize patterns through a machine-learning process in a manner modeled on the human brain. As Dr. Wise explained, they first emerged in the 1960s to simulate the human brain’s psychological-neurologic systems.
Through bivariate and multivariate analyses, the research identified eight preoperative variables strongly associated with the 30-day endpoints. After univariate analysis, seven of those variables were statistically significant: older age (P = .03), nonwhite race, higher initial body mass index, severe hypertension, history of diabetes, nonindependent functional status, and previous foregut/bariatric surgery (all P less than .001). “Gender was the only factor that was not predictive,” Dr. Wise said.
The factors held up under logistic regression modeling. “We were able to use a traditional logistic regression model that came up with a reasonable area under the curve of 0.572,” he said. Using artificial neural network analysis, the training set, which comprised 80% of patients, was more accurate than logistic regression, with an area under the curve of 0.582.
One limitation was that this was a “small study,” Dr. Wise said, influenced by selection bias inherent in any retrospective data selection. Other major factors that may exist were not considered.
However, he noted, “in the past we’ve had some success translating neural networks into something that’s clinically useful.” His group at Vanderbilt University published a report of artificial neural network modeling to identify five factors predictive of weight loss after Roux-en-Y gastric bypass 2 years ago (Surg Endosc. 2016;30:480-8). “There are ways to translate neural networks clinically,” Dr. Wise said.
Dr. Wise had no financial relationships to disclose.
SOURCE: Wise ES et al. SAGES 2019, Abstract S053.
BALTIMORE – Neural networks are the building blocks of machine learning and artificial intelligence, and researchers from the University of Minnesota have identified a panel of “simple, readily known” preoperative patient factors that they fed into an artificial neural network model that can be predictive of 30-day outcomes after laparoscopic sleeve gastrectomy, one of the researchers reported at the annual meeting of the Society of American Gastrointestinal and Endoscopic Surgeons.
“The biggest limitation to using neural networks clinically is the fact that they’re algorithmic complex,” said Eric S. Wise, MD, of the University of Minnesota, Minneapolis, in presenting the research. “There is an underlying algorithm that’s developed, but it’s very difficult to understand.” He called it “a black box problem.”
Nonetheless, the researchers drew upon 101,721 laparoscopic sleeve gastrectomy cases from the 2016 Metabolic and Bariatric Surgery Accreditation and Quality Improvement Program national database to extract factors that were associated with postoperative complications. “More pertinently, we wanted to optimize predictability of a panel of readily obtainable, easily qualifiable preoperative factors and maximize the variants that are contained within those variables to predict the outcome of 30-day morbidity and mortality,” Dr. Wise said.
Essentially, neural networks recognize patterns through a machine-learning process in a manner modeled on the human brain. As Dr. Wise explained, they first emerged in the 1960s to simulate the human brain’s psychological-neurologic systems.
Through bivariate and multivariate analyses, the research identified eight preoperative variables strongly associated with the 30-day endpoints. After univariate analysis, seven of those variables were statistically significant: older age (P = .03), nonwhite race, higher initial body mass index, severe hypertension, history of diabetes, nonindependent functional status, and previous foregut/bariatric surgery (all P less than .001). “Gender was the only factor that was not predictive,” Dr. Wise said.
The factors held up under logistic regression modeling. “We were able to use a traditional logistic regression model that came up with a reasonable area under the curve of 0.572,” he said. Using artificial neural network analysis, the training set, which comprised 80% of patients, was more accurate than logistic regression, with an area under the curve of 0.582.
One limitation was that this was a “small study,” Dr. Wise said, influenced by selection bias inherent in any retrospective data selection. Other major factors that may exist were not considered.
However, he noted, “in the past we’ve had some success translating neural networks into something that’s clinically useful.” His group at Vanderbilt University published a report of artificial neural network modeling to identify five factors predictive of weight loss after Roux-en-Y gastric bypass 2 years ago (Surg Endosc. 2016;30:480-8). “There are ways to translate neural networks clinically,” Dr. Wise said.
Dr. Wise had no financial relationships to disclose.
SOURCE: Wise ES et al. SAGES 2019, Abstract S053.
REPORTING FROM SAGES 2019