EsoEinstein: A Multimodal AI System Integrating EGD and FLIP for Esophageal Motility Assessment

Achalasia diagnosis relies on integrating endoscopic, functional, and physiologic testing, yet high-resolution manometry (HRM) is invasive, uncomfortable, and resource limited. Artificial intelligence (AI) offers a potential pathway to streamline this workflow. This study aimed to develop and validate AI models for diagnosing achalasia using esophagogastroduodenoscopy (EGD), Functional Lumen Imaging Probe (FLIP) panometry, and their combined multimodal integration, using Chicago Classification v4.0 (CCv4.0) as the reference standard.

In this retrospective study, 1,802 adult patients who underwent HRM and had available EGD and/or FLIP data at a tertiary center were included. Separate deep learning models were trained for EGD videos, FLIP heatmap images, and a multimodal transformer that integrated both modalities. Primary outcomes were sensitivity, specificity, and AUC for detecting achalasia.

The 1,802-patient cohort was partitioned into training (n=1,522), validation (n=212), and test sets (n=68) for model development and evaluation. For the test set, the EGD model achieved sensitivity 0.724, specificity 0.974, and AUC 0.908. FLIP achieved sensitivity 0.862, specificity 0.821, and AUC 0.931. The multimodal model performed the best, with sensitivity 0.931, specificity 0.846, and AUC 0.969. Cochran's Q test found no statistically significant differences across models.

Multimodal AI that integrates endoscopic and FLIP data achieves expert-level accuracy and has potential as a less invasive approach for diagnosing achalasia. To realize this potential, future work must validate the system in real-world settings to evaluate its impact on diagnostic timelines and longitudinal patient outcomes.