P197. Triaging Pulmonary Nodules: AI Accurately Predicts Which Patients Need Surgical Consultation Following Robotic Bronchoscopy
Presented During: Thoracic Poster Session II
Cedars-Sinai Medical Center
Los Angeles, CA
United States - Contact Me
Zach is a cardiothoracic surgery resident at Cedars-Sinai Medical Center. Prior to medical school, he was in the army and worked as a federal agent. When not training to to be a cardiothoracic surgeon, Zach hangs out with his dog, scuba dives, and goes sailing in the beautiful Los Angeles sun.
Sunday, May 4, 2025: 9:00 AM - 4:00 PM
Seattle Convention Center | Summit
Room: Poster Area, Exhibit Hall
Description
Objective
Robotic navigational bronchoscopy (RNB) and biopsy can provide diagnosis prior to surgical resection for suspicious pulmonary nodules. There are limited patient and nodule characteristics known to predict those requiring surgery after robotic bronchoscopy. We applied a machine learning model to predict those who were eligible for curative surgery after RNB.
Methods
Using a peer reviewed, automated machine learning tool that automatically prepares data for model development, we assessed our institutional robotic bronchoscopy patients. The primary endpoint was curative surgery or no surgery after RNB. Area Under the Curve (AUC) was utilized to assess algorithm and model performance and predictive ability. Also known as receiver operating characteristics, AUC graphs show classifiers' performance by plotting the true positive rate and false positive rate. Machine learning algorithms in our model included Category Gradient Boosting, Decision Tree, Extreme Gradient Boosting, Logistic Regression, and Random Forest. Mutual Information Median Scores and MultiSURF Median Scores were assessed for each model to provide a unique and specific level of importance to each datapoint and variable.
Results
A total of 1,091 RNB patients were included; 262 (24%) underwent curative intent surgery and 829 (76%) did not. Our analysis included 41 variables, consisting of 38,257 unique datapoints. AUC graphs demonstrated all algorithms performed well (Figure 1): Extreme Gradient Boosting had an AUC of 0.838, Category Gradient Boosting had an AUC of 0.823, Decision Tree had an AUC of 0.794, Logistic Regression had an AUC of 0.808, and Random Forest had an AUC of 0.824. As a comparison, industry standard AUCs as well as Society of Thoracic Surgeons Risk Calculator AUCs range between 0.7 and 0.8.
Based on Mutual Information Median Scores and MultiSURF Median Scores, the most important variables were whether a radial endobronchial ultrasound was used, whether the bronchoscopy was diagnostic or not, the location of the nodule, and number of nodules biopsied.
Conclusions
Our machine learning models, evaluating 41 different variables with an AUC ranging between 0.794 - 0.838, can effectively aid in post-bronchoscopy triage, especially in non-diagnostic cases, helping direct patients to medical or surgical clinics by determining likelihood of requiring subsequent surgery.
Authors
Zachary Brennan (1), Claire Perez (2), Lucas Weiser (3), Kellie Knabe (2), Charles Fuller (2), Sevannah Soukiasian (4), Rafaelle Rocco (2), Andrew Brownlee (5), Harmik Soukiasian (2)
Institutions
(1) Cedars-Sinai Medical Center, Gainesville, CA, (2) Cedars-Sinai Medical Center, Los Angeles, CA, (3) Cedars Sinai Medical Center, Los Angeles, CA, (4) Cedars Sinai Medical Center, Encino, CA, (5) Cedars-Sinai Medical Center, Studio City, CA
Robotic navigational bronchoscopy (RNB) and biopsy can provide diagnosis prior to surgical resection for suspicious pulmonary nodules. There are limited patient and nodule characteristics known to predict those requiring surgery after robotic bronchoscopy. We applied a machine learning model to predict those who were eligible for curative surgery after RNB.
Methods
Using a peer reviewed, automated machine learning tool that automatically prepares data for model development, we assessed our institutional robotic bronchoscopy patients. The primary endpoint was curative surgery or no surgery after RNB. Area Under the Curve (AUC) was utilized to assess algorithm and model performance and predictive ability. Also known as receiver operating characteristics, AUC graphs show classifiers' performance by plotting the true positive rate and false positive rate. Machine learning algorithms in our model included Category Gradient Boosting, Decision Tree, Extreme Gradient Boosting, Logistic Regression, and Random Forest. Mutual Information Median Scores and MultiSURF Median Scores were assessed for each model to provide a unique and specific level of importance to each datapoint and variable.
Results
A total of 1,091 RNB patients were included; 262 (24%) underwent curative intent surgery and 829 (76%) did not. Our analysis included 41 variables, consisting of 38,257 unique datapoints. AUC graphs demonstrated all algorithms performed well (Figure 1): Extreme Gradient Boosting had an AUC of 0.838, Category Gradient Boosting had an AUC of 0.823, Decision Tree had an AUC of 0.794, Logistic Regression had an AUC of 0.808, and Random Forest had an AUC of 0.824. As a comparison, industry standard AUCs as well as Society of Thoracic Surgeons Risk Calculator AUCs range between 0.7 and 0.8.
Based on Mutual Information Median Scores and MultiSURF Median Scores, the most important variables were whether a radial endobronchial ultrasound was used, whether the bronchoscopy was diagnostic or not, the location of the nodule, and number of nodules biopsied.
Conclusions
Our machine learning models, evaluating 41 different variables with an AUC ranging between 0.794 - 0.838, can effectively aid in post-bronchoscopy triage, especially in non-diagnostic cases, helping direct patients to medical or surgical clinics by determining likelihood of requiring subsequent surgery.
Authors
Zachary Brennan (1), Claire Perez (2), Lucas Weiser (3), Kellie Knabe (2), Charles Fuller (2), Sevannah Soukiasian (4), Rafaelle Rocco (2), Andrew Brownlee (5), Harmik Soukiasian (2)
Institutions
(1) Cedars-Sinai Medical Center, Gainesville, CA, (2) Cedars-Sinai Medical Center, Los Angeles, CA, (3) Cedars Sinai Medical Center, Los Angeles, CA, (4) Cedars Sinai Medical Center, Encino, CA, (5) Cedars-Sinai Medical Center, Studio City, CA