Machine learning may enable noninvasive prediction of Barrett’s esophagus

A new risk prediction panel based on machine learning may enable routine, noninvasive identification of patients with a high risk of Barrett’s esophagus, according to investigators.

The methods used in the present study could potentially be applied to other conditions to aid in diagnosis and limit low-yield invasive testing, reported lead author Avi Rosenfeld, PhD, of Jerusalem College of Technology in Israel, and colleagues.

“The currently used system for identifying patients with Barrett’s esophagus, or those at risk of esophageal adenocarcinoma, is flawed because it is based on symptoms that trigger expensive and unpleasant invasive tests,” the investigators wrote. Their report is in The Lancet Digital Health.

In an effort to improve early clinical clarity, the investigators turned to machine learning. Algorithm development and testing relied upon data from more than 1,600 patients involved in two case-control trials, BEST2 and BOOST. From BEST2, 1,299 patients were randomized in a 6:4 ratio to generate a training dataset (n = 776) and a testing dataset (n = 523). Data from all 398 patients in the BOOST trial were used for external validation. Barrett’s esophagus was common in both trials, with a prevalence of 68% and 50% in BEST2 and BOOST, respectively.

During the training phase, machine learning identified independent patient characteristics associated with a diagnosis of Barrett’s esophagus. Specific artificial intelligence techniques included information gain and correlation-based feature selection.

“Both information gain and correlation-based feature selection are filter feature selection methods and thus have the advantage of being fast, scalable, and independent of the classifier,” the investigators noted. They explained that independence from the classifier is “crucial,” as this characteristic has been associated with improved interpretability and more stable algorithms than conventional statistical approaches.

The training process revealed eight distinct diagnostic features: sex, age, waist circumference, cigarette smoking, duration of acidic taste, duration of heartburn, frequency of stomach pain, and use of antireflux medication. All of these were directly correlated with Barrett’s esophagus, except for frequency of stomach pain, which had an inverse relationship. The investigators noted that this inverse relationship may initially seem counterintuitive, but a closer look suggests that the relationship is appropriate.

“Most patients with esophageal adenocarcinoma are not identified before cancer develops despite many of them having Barrett’s esophagus,” the investigators wrote. “Indeed, 40% of patients with esophageal adenocarcinoma have not previously had symptomatic reflux and many probably had Barrett’s esophagus. Therefore, Barrett’s esophagus has been hypothesized to not be associated with severity of reflux symptoms; which fits with the model determined from our data.”

To provide a test of the model’s predictive ability, the investigators arbitrarily set sensitivity at 90%. Within this context, logistic regression was used to provide an upper estimate of the model’s predictive ability; this resulted in an area under the receiver-operator curve (AUC) of 0.87 and a specificity of 68%. In the validation phase, these figures decreased slightly. In the testing dataset, AUC was 0.86, while specificity was 65%. External validation was associated with an AUC of 0.81 and a specificity of 58%.

“We have shown that a panel with eight features, including detailed stomach and chest symptoms, can identify the presence of Barrett’s esophagus with high sensitivity and specificity in a case-control population,” the investigators concluded.

“Simple triaging of individuals might be possible on the basis of predictive panels that include variables that are widely available or easy to obtain,” they added. “Patient age and sex, together with medication and smoking history, are routinely captured in primary care systems. Additionally, asking about duration of heartburn and acidic taste, frequency of stomach pain, and measuring waist circumference should be simple for physicians. Alternatively, a patient could do a self-assessment using a web-based app and generate a personalized risk profile for having Barrett’s esophagus.”

The study was funded by the Charles Wolfson Charitable Trust and Guts UK, the National Institute for Health Research Biomedical Research Centre, Cancer Research UK, and the Wellcome/EPSRC Centre for Interventional and Surgical Sciences at University College London. Dr. Fitzgerald reported a relationship with Medtronic via licensing of the cytosponge device.

SOURCE: Rosenfeld A et al. Lancet Digital Health. 2019 Dec 5. doi: 10.1016/S2589-7500(19)30216-X.

A new risk prediction panel based on machine learning may enable routine, noninvasive identification of patients with a high risk of Barrett’s esophagus, according to investigators.

The methods used in the present study could potentially be applied to other conditions to aid in diagnosis and limit low-yield invasive testing, reported lead author Avi Rosenfeld, PhD, of Jerusalem College of Technology in Israel, and colleagues.

“The currently used system for identifying patients with Barrett’s esophagus, or those at risk of esophageal adenocarcinoma, is flawed because it is based on symptoms that trigger expensive and unpleasant invasive tests,” the investigators wrote. Their report is in The Lancet Digital Health.

In an effort to improve early clinical clarity, the investigators turned to machine learning. Algorithm development and testing relied upon data from more than 1,600 patients involved in two case-control trials, BEST2 and BOOST. From BEST2, 1,299 patients were randomized in a 6:4 ratio to generate a training dataset (n = 776) and a testing dataset (n = 523). Data from all 398 patients in the BOOST trial were used for external validation. Barrett’s esophagus was common in both trials, with a prevalence of 68% and 50% in BEST2 and BOOST, respectively.

During the training phase, machine learning identified independent patient characteristics associated with a diagnosis of Barrett’s esophagus. Specific artificial intelligence techniques included information gain and correlation-based feature selection.

“Both information gain and correlation-based feature selection are filter feature selection methods and thus have the advantage of being fast, scalable, and independent of the classifier,” the investigators noted. They explained that independence from the classifier is “crucial,” as this characteristic has been associated with improved interpretability and more stable algorithms than conventional statistical approaches.

The training process revealed eight distinct diagnostic features: sex, age, waist circumference, cigarette smoking, duration of acidic taste, duration of heartburn, frequency of stomach pain, and use of antireflux medication. All of these were directly correlated with Barrett’s esophagus, except for frequency of stomach pain, which had an inverse relationship. The investigators noted that this inverse relationship may initially seem counterintuitive, but a closer look suggests that the relationship is appropriate.

“Most patients with esophageal adenocarcinoma are not identified before cancer develops despite many of them having Barrett’s esophagus,” the investigators wrote. “Indeed, 40% of patients with esophageal adenocarcinoma have not previously had symptomatic reflux and many probably had Barrett’s esophagus. Therefore, Barrett’s esophagus has been hypothesized to not be associated with severity of reflux symptoms; which fits with the model determined from our data.”

To provide a test of the model’s predictive ability, the investigators arbitrarily set sensitivity at 90%. Within this context, logistic regression was used to provide an upper estimate of the model’s predictive ability; this resulted in an area under the receiver-operator curve (AUC) of 0.87 and a specificity of 68%. In the validation phase, these figures decreased slightly. In the testing dataset, AUC was 0.86, while specificity was 65%. External validation was associated with an AUC of 0.81 and a specificity of 58%.

“We have shown that a panel with eight features, including detailed stomach and chest symptoms, can identify the presence of Barrett’s esophagus with high sensitivity and specificity in a case-control population,” the investigators concluded.

“Simple triaging of individuals might be possible on the basis of predictive panels that include variables that are widely available or easy to obtain,” they added. “Patient age and sex, together with medication and smoking history, are routinely captured in primary care systems. Additionally, asking about duration of heartburn and acidic taste, frequency of stomach pain, and measuring waist circumference should be simple for physicians. Alternatively, a patient could do a self-assessment using a web-based app and generate a personalized risk profile for having Barrett’s esophagus.”

The study was funded by the Charles Wolfson Charitable Trust and Guts UK, the National Institute for Health Research Biomedical Research Centre, Cancer Research UK, and the Wellcome/EPSRC Centre for Interventional and Surgical Sciences at University College London. Dr. Fitzgerald reported a relationship with Medtronic via licensing of the cytosponge device.

SOURCE: Rosenfeld A et al. Lancet Digital Health. 2019 Dec 5. doi: 10.1016/S2589-7500(19)30216-X.

A new risk prediction panel based on machine learning may enable routine, noninvasive identification of patients with a high risk of Barrett’s esophagus, according to investigators.

The methods used in the present study could potentially be applied to other conditions to aid in diagnosis and limit low-yield invasive testing, reported lead author Avi Rosenfeld, PhD, of Jerusalem College of Technology in Israel, and colleagues.

“The currently used system for identifying patients with Barrett’s esophagus, or those at risk of esophageal adenocarcinoma, is flawed because it is based on symptoms that trigger expensive and unpleasant invasive tests,” the investigators wrote. Their report is in The Lancet Digital Health.

In an effort to improve early clinical clarity, the investigators turned to machine learning. Algorithm development and testing relied upon data from more than 1,600 patients involved in two case-control trials, BEST2 and BOOST. From BEST2, 1,299 patients were randomized in a 6:4 ratio to generate a training dataset (n = 776) and a testing dataset (n = 523). Data from all 398 patients in the BOOST trial were used for external validation. Barrett’s esophagus was common in both trials, with a prevalence of 68% and 50% in BEST2 and BOOST, respectively.

During the training phase, machine learning identified independent patient characteristics associated with a diagnosis of Barrett’s esophagus. Specific artificial intelligence techniques included information gain and correlation-based feature selection.

“Both information gain and correlation-based feature selection are filter feature selection methods and thus have the advantage of being fast, scalable, and independent of the classifier,” the investigators noted. They explained that independence from the classifier is “crucial,” as this characteristic has been associated with improved interpretability and more stable algorithms than conventional statistical approaches.

The training process revealed eight distinct diagnostic features: sex, age, waist circumference, cigarette smoking, duration of acidic taste, duration of heartburn, frequency of stomach pain, and use of antireflux medication. All of these were directly correlated with Barrett’s esophagus, except for frequency of stomach pain, which had an inverse relationship. The investigators noted that this inverse relationship may initially seem counterintuitive, but a closer look suggests that the relationship is appropriate.

“Most patients with esophageal adenocarcinoma are not identified before cancer develops despite many of them having Barrett’s esophagus,” the investigators wrote. “Indeed, 40% of patients with esophageal adenocarcinoma have not previously had symptomatic reflux and many probably had Barrett’s esophagus. Therefore, Barrett’s esophagus has been hypothesized to not be associated with severity of reflux symptoms; which fits with the model determined from our data.”

To provide a test of the model’s predictive ability, the investigators arbitrarily set sensitivity at 90%. Within this context, logistic regression was used to provide an upper estimate of the model’s predictive ability; this resulted in an area under the receiver-operator curve (AUC) of 0.87 and a specificity of 68%. In the validation phase, these figures decreased slightly. In the testing dataset, AUC was 0.86, while specificity was 65%. External validation was associated with an AUC of 0.81 and a specificity of 58%.

“We have shown that a panel with eight features, including detailed stomach and chest symptoms, can identify the presence of Barrett’s esophagus with high sensitivity and specificity in a case-control population,” the investigators concluded.

“Simple triaging of individuals might be possible on the basis of predictive panels that include variables that are widely available or easy to obtain,” they added. “Patient age and sex, together with medication and smoking history, are routinely captured in primary care systems. Additionally, asking about duration of heartburn and acidic taste, frequency of stomach pain, and measuring waist circumference should be simple for physicians. Alternatively, a patient could do a self-assessment using a web-based app and generate a personalized risk profile for having Barrett’s esophagus.”

The study was funded by the Charles Wolfson Charitable Trust and Guts UK, the National Institute for Health Research Biomedical Research Centre, Cancer Research UK, and the Wellcome/EPSRC Centre for Interventional and Surgical Sciences at University College London. Dr. Fitzgerald reported a relationship with Medtronic via licensing of the cytosponge device.

SOURCE: Rosenfeld A et al. Lancet Digital Health. 2019 Dec 5. doi: 10.1016/S2589-7500(19)30216-X.