Diagnosing disorders of the esophagus has always been as much art as science. When patients complain of dysphagia—the difficulty swallowing food, liquids, or even saliva—physicians turn to high-resolution impedance manometry, or HRIM, a procedure in which a thin catheter threaded through the nose records pressure and electrical impedance along the length of the esophagus while the patient swallows. The resulting data are extraordinarily rich but notoriously difficult to interpret: complex spatiotemporal patterns that demand years of training to read, with studies showing that even experienced clinicians frequently disagree when classifying borderline cases. Now, a team of researchers at the Technical University of Munich has developed an artificial intelligence system that could bring standardization and, potentially, a new level of diagnostic accuracy to this difficult field. Their work, published in the International Journal of Computer Assisted Radiology and Surgery, combines two ideas that have never been brought together in this domain before: representing the raw manometry signals as anatomically faithful graphs, and fusing them with detailed information about the patient themselves.
The study, led by Alexander Geiger and Lars Wagner of the Research Group MITI at TUM University Hospital, with contributions from Daniel Rueckert, Alois Knoll, Dirk Wilhelm, and Alissa Jell, analyzed HRIM examinations from 104 patients with esophageal motility disorders collected between 2020 and 2025. Each examination followed the Chicago Classification protocol version 3.0, the international standard for interpreting esophageal manometry, and yielded a total of roughly 1,800 annotated swallow events. For every swallow, medical personnel labeled three distinct categories: the peristaltic pattern, the contraction type, and the pressure characteristics. Because each category can take multiple values, the problem is a multi-category, multi-class classification task—a substantially more realistic and demanding scenario than the binary problems many earlier machine learning studies addressed. The data also exhibit strong class imbalance, reflecting the natural prevalence of certain motility patterns over rare pathological ones.
What distinguishes this work from prior attempts is how the manometry data themselves are modeled. The HRIM catheter carries 36 pressure sensors and 15 impedance sensors, and for each swallow the system captures 750 time steps, spanning 15 seconds of activity. Rather than flattening this stream into an image or a vector of handcrafted features, the Munich team modeled it directly as what they call a spatiotemporal graph: a sequence of 750 graphs, one per time step, in which each node represents a pressure sensor and its recorded pressure value, while edges connect physically adjacent sensors along the catheter. Crucially, the impedance measurements— which track bolus transit, the movement of the swallowed material itself—are encoded as features on the edges connecting every second sensor pair. This design embeds the anatomy and physiology of the esophagus directly into the mathematical structure of the data, allowing the learning algorithm to exploit the organ’s intrinsic topology rather than rediscovering it from scratch.
To process these graphs, the researchers applied graph neural networks, specifically GATv2 (Graph Attention Networks v2) and GENConv layers, which propagate information between neighboring nodes during message passing, updating each node’s representation by aggregating the pressure values of its neighbors together with the impedance features on connecting edges. The resulting per-time-step graph embeddings are then fed into temporal models—a Temporal Convolutional Network and a Transformer—followed by a hierarchical attention pooling scheme. This pooling operates at two levels: node-level attention first summarizes activity across all sensors at each moment in time, and temporal attention then compresses the entire 15-second swallow into a compact, category-specific representation. Separate pooling heads for the three classification categories allow the model to learn what matters most for distinguishing, say, a failed peristalsis from a fragmented contraction, as opposed to a hypercontractile pressurization pattern.
The second pillar of the approach is the patient. Before their examinations, every patient completed a standardized questionnaire covering demographics, comorbidities, and symptoms, along with a free-text field for the examiner’s clinical notes. The researchers extracted four numeric features (such as age and body mass index) and 28 categorical or ordinal features from the structured sections using keyword detection. For the unstructured notes, they deployed a large language model—the open-source Qwen3-14B—in two stages: first asking it to compile a comprehensive list of clinically relevant attributes from all notes, then processing each individual note to determine which attributes were present. This produced 19 additional binary features, for a total of 51 patient-specific characteristics. A simple feed-forward encoder converts this vector into a dense embedding, which is concatenated with the graph-derived swallow representations before passing through category-specific classification heads.
Training the multimodal network required careful handling of the class imbalance and the small dataset. The team used a class-weighted cross-entropy loss with label smoothing for each category, in which rarer classes receive proportionally larger weights, and added a supervised contrastive loss that pulls swallows sharing the same label into tight clusters in the embedding space while pushing different classes apart. A t-SNE visualization of the learned embeddings showed clear inter-class separation, an encouraging sign that the representations capture physiologically meaningful structure. Validation employed a fivefold cross-validation scheme defined at the patient level, using an iterative multilabel stratification strategy to ensure that the joint label distribution across all three categories was preserved in every fold—an essential safeguard against information leakage between patients.
The results favor the graph-based approach. Compared against three vision-based baselines—ResNet50, Vision Transformer, and Swin Transformer V2, each fed with the HRIM data reshaped into 224-by-224 image-like tensors—the graph models, particularly the GENConv-TCN combination, consistently achieved the highest mean weighted F1 scores across all three classification categories. The authors are careful to note that statistical tests did not reach significance, likely because only five cross-validation folds were available and the dataset is small and imbalanced; Friedman tests comparing the models yielded p-values between 0.151 and 0.592. Nevertheless, the pattern of mean differences uniformly favored the graph-based architecture, a consistent trend the researchers argue warrants investigation with larger cohorts. The comparison is meaningful because vision-based CNNs and Transformers are the dominant paradigm in existing manometry AI literature, with prior studies reporting accuracies of up to 98 percent on internal datasets.
Perhaps the most clinically interesting finding comes from the ablation experiments, in which the team systematically removed each input modality from the best-performing model. The full multimodal configuration outperformed every reduced version across all three categories, and Friedman tests confirmed a significant global effect of modality composition on performance for peristaltic pattern (p = 0.0088), contraction type (p = 0.0106), and pressure characteristics (p = 0.0021). Removing manometry caused the largest degradation, confirming it as the dominant signal. Removing patient information produced a smaller but consistent drop, demonstrating for the first time that demographic, clinical, and LLM-extracted symptom data genuinely add diagnostic value on top of the pressure measurements. Intriguingly, manometry alone slightly outperformed manometry combined with only impedance or only patient features, suggesting that these secondary signals are conditionally informative but may require more sophisticated fusion strategies to be fully exploited—a puzzle the authors flag for future work.
The clinical significance of this research lies in its response to a well-documented problem. Dysphagia becomes increasingly common with advancing age and imposes substantial burdens on patients and health systems alike, from malnutrition and aspiration pneumonia to reduced quality of life. Yet the gold-standard diagnostic tool suffers from limited inter-rater reliability, with interpretation varying according to clinician expertise. Previous computational efforts, from early artificial neural networks through CNN–BiLSTM hybrids and attention-based ensembles, have attacked the classification problem but have relied almost exclusively on the manometry signal itself, ignoring the demographic and clinical context that physicians weigh in practice. Studies linking swallowing disorders to age, gender, heartburn, obesity, medication use, and comorbidities suggest exactly the kind of complementary information that the Munich pipeline now formally incorporates into a computational model.
The study’s authors are appropriately measured about the limitations. With 104 patients and uneven class distributions, the dataset cannot yet support definitive statistical conclusions, and the team acknowledges that graph construction strategies, hyperparameter optimization, and the integration of additional physiological features all leave room for refinement. The fivefold cross-validation design and patient-level splitting, however, represent rigorous methodology for a dataset of this size, and the open-access publication permits independent scrutiny. The work is framed explicitly as an initial exploration—a feasibility study demonstrating that the direction is promising rather than a finished clinical tool.
Even so, the implications extend beyond esophageal diagnostics. The core insight—that the physical structure of a measuring instrument’s sensor array can be preserved, rather than destroyed, when feeding data to a neural network, and that the patient’s story can be mined from free-text clinical notes by a language model and fused with physiological signals—offers a template applicable to many domains of medicine where multimodal, structurally organized data confront clinicians with interpretation challenges. If larger, multi-center studies confirm the trends observed in Munich, HRIM interpretation could eventually be supported by systems that never tire, never vary by institution, and consider the whole patient rather than a pressure plot alone. For the growing population of older adults facing difficulty swallowing, that could translate into faster, more consistent diagnoses and earlier, better-targeted treatment.
Subject of Research: Automated multimodal machine learning classification of esophageal motility disorders using high-resolution impedance manometry and patient-specific clinical data
Subject of Research: Medicine
Article Title: Multimodal graph-based classification of esophageal motility disorders
Article References: Geiger, A., Wagner, L., Rueckert, D., Knoll, A., Wilhelm, D., & Jell, A. (2026). Multimodal graph-based classification of esophageal motility disorders. International Journal of Computer Assisted Radiology and Surgery. https://doi.org/10.1007/s11548-026-03760-7
Image Credits: AI Generated
DOI: 10.1007/s11548-026-03760-7
Keywords: esophageal motility disorders, high-resolution impedance manometry, graph neural network, multimodal machine learning, dysphagia, large language model, spatiotemporal graphs, patient-specific features, Chicago Classification, deep learning
Cite Scienmag News
APA MLA Chicago
Ophelia Keating. (September 8, 2026). New multimodal graph approach classifies esophageal motility disorders. Scienmag. https://scienmag.com/new-multimodal-graph-approach-classifies-esophageal-motility-disorders/
Ophelia Keating. “New multimodal graph approach classifies esophageal motility disorders.” Scienmag, 8 September 2026, https://scienmag.com/new-multimodal-graph-approach-classifies-esophageal-motility-disorders/. Accessed 8 September 2026.
Ophelia Keating. “New multimodal graph approach classifies esophageal motility disorders.” Scienmag. September 8, 2026. https://scienmag.com/new-multimodal-graph-approach-classifies-esophageal-motility-disorders/
Copy citation Download RIS
Tags: advanced imaging techniques for esophageal functionadvanced medical imaging integrationAI-driven diagnostic standardization in esophageal diseasesartificial intelligence for esophageal disorder classificationartificial intelligence in esophageal function testingautomated interpretation of HRIM dataclinical implications of AI in motility disorder classificationcombined signal and patient data modelingdeep learning for dysphagia assessmentEsophageal motility disorder diagnosisgraph neural networks for medical signal analysishigh-resolution impedance manometry analysishigh-resolution impedance manometry interpretationimproving diagnostic accuracy in esophageal motinnovative approaches to borderline esophageal casesinnovative esophageal disorder classificationintegration of patient data with manometry signalsmachine learning in dysphagia assessmentmultimodal graph-based AI in gastroenterologyspatiotemporal pattern analysis in HRIM dataspatiotemporal pressure pattern interpretationstandardization of esophageal motility diagnostics
