Benjamin Rogers1, Sabyasachi Samanta2, Kevan Ghobadi2, Amit Patel3, Edoardo Savarino4, Sabine Roman5,6,7, Daniel Sifrim8, C Prakash Gyawali9. 1. Division of Gastroenterology, Washington University School of Medicine, 660 South Euclid Ave., Campus Box 8124, Saint Louis, MO, 63110, USA. 2. Crosswave Solutions, LLC., Lexington, KY, USA. 3. Division of Gastroenterology, Duke University School of Medicine, The Durham Veterans Affairs Medical Center, Durham, NC, USA. 4. Division of Gastroenterology, Department of Surgery, Oncology and Gastroenterology, University of Padua, Padua, Italy. 5. Digestive Physiology, Hospices Civils de Lyon, Hopital E Herriot, Université de Lyon, 69437, Lyon, France. 6. Digestive Physiology, Université de Lyon, Lyon I University, 69008, Lyon, France. 7. Université de Lyon, Inserm U1032, LabTAU, 69008, Lyon, France. 8. Barts and The London School of Medicine and Dentistry Queen Mary, University of London, London, UK. 9. Division of Gastroenterology, Washington University School of Medicine, 660 South Euclid Ave., Campus Box 8124, Saint Louis, MO, 63110, USA. cprakash@wustl.edu.
Abstract
BACKGROUND: Artificial intelligence (AI) has potential to streamline interpretation of pH-impedance studies. In this exploratory observational cohort study, we determined feasibility of automated AI extraction of baseline impedance (AIBI) and evaluated clinical value of novel AI metrics. METHODS: pH-impedance data from a convenience sample of symptomatic patients studied off (n = 117, 53.1 ± 1.2 years, 66% F) and on (n = 93, 53.8 ± 1.3 years, 74% F) anti-secretory therapy and from asymptomatic volunteers (n = 115, 29.3 ± 0.8 years, 47% F) were uploaded into dedicated prototypical AI software designed to automatically extract AIBI. Acid exposure time (AET) and manually extracted mean nocturnal baseline impedance (MNBI) were compared to corresponding total, upright, and recumbent AIBI and upright:recumbent AIBI ratio. AI metrics were compared to AET and MNBI in predicting ≥ 50% symptom improvement in GERD patients. RESULTS: Recumbent, but not upright AIBI, correlated with MNBI. Upright:recumbent AIBI ratio was higher when AET > 6% (median 1.18, IQR 1.0-1.5), compared to < 4% (0.95, IQR 0.84-1.1), 4-6% (0.89, IQR 0.72-0.98), and controls (0.93, IQR 0.80-1.09, p ≤ 0.04). While MNBI, total AIBI, and the AIBI ratio off PPI were significantly different between those with and without symptom improvement (p < 0.05 for each comparison), only AIBI ratio segregated management responders from other cohorts. On ROC analysis, off therapy AIBI ratio outperformed AET in predicting GERD symptom improvement when AET was > 6% (AUC 0.766 vs. 0.606) and 4-6% (AUC 0.563 vs. 0.516) and outperformed MNBI overall (AUC 0.661 vs. 0.313). CONCLUSIONS: BI calculation can be automated using AI. Novel AI metrics show potential in predicting GERD treatment outcome.
BACKGROUND: Artificial intelligence (AI) has potential to streamline interpretation of pH-impedance studies. In this exploratory observational cohort study, we determined feasibility of automated AI extraction of baseline impedance (AIBI) and evaluated clinical value of novel AI metrics. METHODS: pH-impedance data from a convenience sample of symptomatic patients studied off (n = 117, 53.1 ± 1.2 years, 66% F) and on (n = 93, 53.8 ± 1.3 years, 74% F) anti-secretory therapy and from asymptomatic volunteers (n = 115, 29.3 ± 0.8 years, 47% F) were uploaded into dedicated prototypical AI software designed to automatically extract AIBI. Acid exposure time (AET) and manually extracted mean nocturnal baseline impedance (MNBI) were compared to corresponding total, upright, and recumbent AIBI and upright:recumbent AIBI ratio. AI metrics were compared to AET and MNBI in predicting ≥ 50% symptom improvement in GERD patients. RESULTS: Recumbent, but not upright AIBI, correlated with MNBI. Upright:recumbent AIBI ratio was higher when AET > 6% (median 1.18, IQR 1.0-1.5), compared to < 4% (0.95, IQR 0.84-1.1), 4-6% (0.89, IQR 0.72-0.98), and controls (0.93, IQR 0.80-1.09, p ≤ 0.04). While MNBI, total AIBI, and the AIBI ratio off PPI were significantly different between those with and without symptom improvement (p < 0.05 for each comparison), only AIBI ratio segregated management responders from other cohorts. On ROC analysis, off therapy AIBI ratio outperformed AET in predicting GERD symptom improvement when AET was > 6% (AUC 0.766 vs. 0.606) and 4-6% (AUC 0.563 vs. 0.516) and outperformed MNBI overall (AUC 0.661 vs. 0.313). CONCLUSIONS: BI calculation can be automated using AI. Novel AI metrics show potential in predicting GERD treatment outcome.
Authors: Rena Yadlapati; C Prakash Gyawali; Melina Masihi; Dustin A Carlson; Peter J Kahrilas; Billy Darren Nix; Anand Jain; Joseph R Triggs; Michael F Vaezi; Leila Kia; Alexander Kaizer; John E Pandolfino Journal: Am J Gastroenterol Date: 2022-06-10 Impact factor: 12.045