Mayo: Multi-Domain Research Insights

• Mayo is a multifaceted term encompassing an ordinal endoscopic score for ulcerative colitis, a rich institutional clinical research platform, a statistical philosophy by Deborah Mayo, and contested materials science studies.
• In medical imaging and machine learning, Mayo scoring drives advancements through curriculum learning, weak supervision, and synthetic progression models that address ordinal and labeling challenges.
• Mayo Clinic serves as a hub for large-scale, multimodal clinical research and model development, while Deborah Mayo’s error-statistical approach reshapes debates on hypothesis testing and scientific inference.

Searching arXiv for the provided Mayo-related sources to ground the article. “Mayo” appears in contemporary arXiv literature in several technically distinct senses. In gastrointestinal imaging, it denotes the Mayo Endoscopic Subscore (MES), a four-point ordinal scale for ulcerative colitis severity. In biomedical AI and clinical informatics, it denotes Mayo Clinic as an institutional source of large multimodal datasets, foundation models, and operational prediction systems. In the foundations of statistics, it denotes Deborah Mayo, whose error-statistical and severe-testing program has been central to debates over likelihood, sufficiency, and conditionality. In computational materials science, it also appears as Mayo et al., referring to a Li–Sn structure-prediction study later criticized for inconsistencies (Ozdemir et al., 18 Aug 2025).

1. Mayo as the Mayo Endoscopic Subscore in ulcerative colitis

The Mayo Endoscopic Subscore is an integer score in $\{0,1,2,3\}$ used to grade mucosal inflammation in ulcerative colitis. In the MES formulation used in recent endoscopy papers, MES $=0$ denotes normal or inactive disease, MES $=1$ mild disease, MES $=2$ moderate disease, and MES $=3$ severe disease. The scale is explicitly ordinal: errors between adjacent grades are less severe than errors between distant grades. In video-based clinical use, a colonoscopy video is assigned an MES by the most severe frame, so the video label is effectively max-pooled over frame-level severity (Schwab et al., 2021).

This clinical definition creates several modeling difficulties. Standard cross-entropy ignores ordering among grades, even though the distinction between $1 \rightarrow 2$ and $0 \rightarrow 3$ is clinically non-equivalent. Label noise is also substantial because expert gastroenterologists disagree, especially on moderate cases $\text{MES}=1,2$, poor bowel preparation can obscure mucosal patterns, and severe inflammation $\text{MES}=3$ is underrepresented in some datasets (Ozdemir et al., 18 Aug 2025).

A further limitation of the conventional MES is that it summarizes a video by its most severe segment. As Schwab et al. note, severe inflammation spread throughout the colon receives the same MES as otherwise healthy colon with severe inflammation restricted to a small localized segment. This means that disease extent and overall response to treatment may not be fully captured by MES alone (Schwab et al., 2021).

2. Machine learning methods built around Mayo scoring

Recent MES modeling has focused on the combination of ordinal supervision, label uncertainty, and domain-specific pretraining. In "CLoE: Curriculum Learning on Endoscopic Images for Robust MES Classification" (Ozdemir et al., 18 Aug 2025), CLoE defines the task as predicting $y_i \in \{0,1,2,3\}$ from an endoscopic image $=0$0, but replaces uniform training with a difficulty-aware curriculum. Image quality is estimated by a lightweight MobileNetV2 classifier $=0$1 trained on Boston Bowel Preparation Scale labels from HyperKvasir, where BBPS $=0$2 is mapped to clean and BBPS $=0$3 or impacted-stool to noisy. The resulting cleanliness score $=0$4 is thresholded at $=0$5 to partition training images into $=0$6 and $=0$7. Training then follows a three-stage schedule: clean only, clean plus noisy, and noisy only, with stage transition when validation accuracy stalls for 5 epochs. ResizeMix with $=0$8 is used throughout all stages because MixUp and CutMix may confuse ordinal boundaries. On LIMUC, CLoE with ConvNeXt-Tiny reaches $=0$9 accuracy, $=1$0 QWK, and $=1$1 Macro-F1; on HyperKvasir, the best CLoE result is $=1$2 accuracy, $=1$3 QWK, and $=1$4 Macro-F1. The reported ablations show that progressive curriculum learning improves accuracy by about $=1$5 over all-data training, while ResizeMix plus curriculum yields about $=1$6 additional accuracy and about $=1$7 QWK over curriculum alone (Ozdemir et al., 18 Aug 2025).

A complementary line replaces dataset-specific supervision with large-scale endoscopic pretraining. "EndoDINO: A Foundation Model for GI Endoscopy" (Dermyer et al., 8 Jan 2025) pre-trains ViT-B/14, ViT-L/14, and ViT-g/14 models on data curated from 130,037 de-identified GI endoscopy videos, corresponding to about 3.5 billion frames. The curation pipeline downsamples to 5 fps, removes duplicates and near-duplicates to obtain 112 million unique frames, then uses hierarchical $=1$8-means clustering and balanced sampling to construct pretraining sets ranging from 100 k to 10 M images. For MES, only a linear classification head on the frozen CLS embedding is trained, with no data augmentation during head training. On HyperKvasir 3-class MES, EndoDINO-g/14 (10 M) reports $=1$9 accuracy, $=2$0, Macro-F1 $=2$1, and Micro-F1 $=2$2. On LIMUC 4-class MES, EndoDINO-g/14 (10 M HKM) reports AUROC $=2$3, Macro-F1 $=2$4, accuracy $=2$5, and $=2$6. Failure cases are concentrated at class boundaries, especially $=2$7 versus $=2$8, when mucosa is partly obscured by specular highlights (Dermyer et al., 8 Jan 2025).

Weak supervision has also been used to infer frame-level Mayo severity from only video-level labels. Schwab et al. formulate each video as a bag of frames and use max-based multi-instance learning with either an ensemble of ranked binary heads or an ordinal regression head. For the ranked formulation, each head estimates $=2$9, the bag-level score is $=3$0, and the video-level MES is reconstructed from thresholded or converted frame-level probabilities. The best video-level ordinal agreement is obtained by Regression+MIL with quadratic weighted Cohen’s $=3$1 $=3$2, while the Ensemble-Threshold model achieves $=3$3. On frame-level validation against a 4-reader consensus, Regression+MIL reaches $=3$4, and all reported methods fall within the $=3$5 CI of human raters. The same framework reports AUC $=3$6 for mucosal healing and AUC $=3$7 for remission prediction (Schwab et al., 2021).

Generative modeling has extended Mayo scoring from classification to controllable synthesis. "Disentangled Anatomy-Disease Diffusion (DADD) for Controllable Ulcerative Colitis Progression Synthesis" (Dundar et al., 3 May 2026) models UC severity as a continuous ordinal progression along the Mayo Endoscopic Score. It uses an Additive Ordinal Embedder $=3$8, a Feature Purifier that suppresses disease-correlated channels in CLIP-based image tokens, and Triple-Pathway Cross-Attention with anatomy, disease, and delta pathways. The delta pathway implements single-pass steering of $=3$9. On LIMUC, DADD-H reports FID $1 \rightarrow 2$0, CMMD $1 \rightarrow 2$1, Precision $1 \rightarrow 2$2, and Recall $1 \rightarrow 2$3. For downstream MES classification, mixed real plus synthetic training raises accuracy from $1 \rightarrow 2$4 to $1 \rightarrow 2$5 and QWK from $1 \rightarrow 2$6 to $1 \rightarrow 2$7. A regression judge trained on real data rates DADD-H outputs at QWK $1 \rightarrow 2$8 and RMSE $1 \rightarrow 2$9, near the reported real-data ceiling of $0 \rightarrow 3$0 and $0 \rightarrow 3$1. This suggests that Mayo scoring has become not only a diagnostic label but also a controllable latent axis for disease-progression synthesis (Dundar et al., 3 May 2026).

A recurring misconception in this literature is to treat MES as an ordinary multiclass target. The recent work consistently argues otherwise: MES is ordinal, noisy, and sensitive to acquisition artifacts, so methods that ignore ordering, confidence, or temporal/video structure lose clinically relevant information (Ozdemir et al., 18 Aug 2025).

3. Mayo Clinic as a multimodal clinical research infrastructure

In another major sense, “Mayo” denotes Mayo Clinic as a data-producing and method-deploying institution. "Launching Insights: A Pilot Study on Leveraging Real-World Observational Data from the Mayo Clinic Platform to Advance Clinical Research" describes the Mayo Clinic Platform (MCP) as a scalable ecosystem integrating structured EHR, unstructured text, imaging, and genomics/omics. The reported volumes are substantial: more than 2.7 billion lab-result rows, more than 400 million medication orders, 1.25 billion clinical notes, and 3.9 billion images across radiology, pathology, and ECG waveforms. Institutional sources include Mayo Clinic and Mercy Hospital via the Mayo Clinic Care Network (Yu et al., 21 Mar 2025).

MCP’s de-identification pipeline removes direct identifiers, shifts dates by a random offset per patient, replaces dates with “days from index,” applies NLP redaction to free text, and restricts analysis to a FISMA-moderate secure enclaved compute environment. Data standardization aligns terminologies to ICD-10-CM, LOINC, and RxNorm, and organizes sentence-level note facts into the FACT_SYN_SENTENCES table with section labels in SYN_SECTION_HEADER_DK. Although the paper does not explicitly state use of OMOP CDM or FHIR, it states that MCP Discover implements a bespoke CDM with the same principles (Yu et al., 21 Mar 2025).

The platform’s reported analytic workflow has three steps: cohort definition in Cohort Visualizer, data extraction via SparkSQL in a Jupyter Notebook, and analysis in MCP Workspace using R or Python. Four pilot studies illustrate the range of supported designs. Heart-failure drug emulation uses propensity-score matching followed by Cox modeling and reports HR $0 \rightarrow 3$2 $0 \rightarrow 3$3 for ACE-inhibitors versus ARBs, with concordance index $0 \rightarrow 3$4. An antihypertensive-medication study on ADRD risk reports overall HR $0 \rightarrow 3$5 $0 \rightarrow 3$6 and interaction AHM$0 \rightarrow 3$7metformin HR $0 \rightarrow 3$8 $0 \rightarrow 3$9, with model $\text{MES}=1,2$0-index $\text{MES}=1,2$1. A BiGRU model for MCI-to-AD progression reports $\text{MES}=1,2$2 and $\text{MES}=1,2$3. A deep network for MACE after liver transplantation reports AUC $\text{MES}=1,2$4 $\text{MES}=1,2$5, sensitivity $\text{MES}=1,2$6, and specificity $\text{MES}=1,2$7 (Yu et al., 21 Mar 2025).

Compared to Mayo’s internal Electronic Health Record data, MCP is reported to provide broader accessibility, enhanced data standardization, and multi-institutional integration. A plausible implication is that, in current biomedical AI usage, “Mayo” often indexes not a single dataset or clinic but a translational platform for cohort construction, model validation, and cross-institutional evidence generation (Yu et al., 21 Mar 2025).

4. Mayo Clinic in large-scale models: pathology, oncology, and forecasting

Mayo Clinic also appears in arXiv as a site of large-scale model development. "Atlas: A Novel Pathology Foundation Model by Mayo Clinic, Charité, and Aignostics" presents a self-supervised ViT-H/14 with 632 M parameters trained with the RudolfV/DINOv2 paradigm on 1.2 million de-identified histopathology whole-slide images from Mayo Clinic and Charité. The training corpus covers over 70 tissue or organ types, over 100 distinct staining protocols, slides from 7 whole-slide scanners, and four magnifications $\text{MES}=1,2$8. In total, 3.4 billion tiles were generated, 520 million training tiles were sampled, training used AdamW with cosine decay for 100 epochs, and the run took about 3 weeks on a 128-GPU Nvidia H100 cluster inside the Mayo Clinic Platform. Frozen linear probes on 21 public benchmarks report Molecular-Average $\text{MES}=1,2$9, Morphology-Average $\text{MES}=3$0, and Overall Average $\text{MES}=3$1, with state of the art on 11 of 21 tasks (Alber et al., 9 Jan 2025).

Mayo Clinic data have also been used for graph-based precision oncology. A study of 794 adult cancer patients from Mayo Clinic Rochester combines EHR-derived phenotypic features and FoundationOne genetic test reports in a joint phenotypic/genetic graph. Eight GNNs are compared with conventional baselines. Mean performance for GNNs versus baselines is reported as accuracy $\text{MES}=3$2 versus $\text{MES}=3$3, precision $\text{MES}=3$4 versus $\text{MES}=3$5, recall $\text{MES}=3$6 versus $\text{MES}=3$7, and F1-score $\text{MES}=3$8 versus $\text{MES}=3$9. Among individual models, ChebNet, GraphSAGE, and TAGCN each report accuracy $y_i \in \{0,1,2,3\}$0, while GAT underperforms at accuracy $y_i \in \{0,1,2,3\}$1 and F1 $y_i \in \{0,1,2,3\}$2 (Oniani et al., 2021).

A different Mayo Clinic modeling tradition is exemplified by the hierarchical Bayesian spatiotemporal SIR model for COVID-19 cases and hospitalizations. That system explicitly treats the transmission rate $y_i \in \{0,1,2,3\}$3 as time-varying, uses county-level SIR state dynamics with Binomial recovery and vaccination flows, models spatial structure with CAR/GMRF priors, and performs Bayesian inference with a hybrid Gibbs–Metropolis–Hastings sampler. Forecast quality is assessed by RMSE, MAE, coverage, quantile score, and CRPS. The paper states that the model accurately predicted the timing and extent of the summer and fall surges at Mayo Clinic sites and was useful to the state of Minnesota for policy decisions (Storlie et al., 2021).

Taken together, these studies show that “Mayo” in contemporary ML literature frequently denotes institutional scale: pathology foundation models, graph-based clinical stratification, and uncertainty-aware operational forecasting. This suggests a research profile organized around very large clinical corpora, heterogeneous modalities, and deployment-oriented evaluation (Alber et al., 9 Jan 2025).

5. Deborah Mayo and the error-statistical tradition

In the philosophy of statistics, “Mayo” refers to Deborah Mayo, whose work centers on error statistics and severe testing. Severe testing is defined by the idea that a hypothesis passes a test to the extent that the test would probably have found flaws if they were present. One formalization writes, for a testing problem with statistic $y_i \in \{0,1,2,3\}$4, observed data $y_i \in \{0,1,2,3\}$5, and a specific alternative $y_i \in \{0,1,2,3\}$6,

$y_i \in \{0,1,2,3\}$7

This yields a severity profile over a grid of alternatives rather than a single binary decision. The associated program emphasizes long-run error probabilities, explicit discrepancies from the null, and model checking, rather than taking a large $y_i \in \{0,1,2,3\}$8-value as confirmation without regard to power (Gelman et al., 2019).

In the error-statistical formulation reviewed by Christian Robert, the central quantities are the Type I error

$y_i \in \{0,1,2,3\}$9

the Type II error

$=0$00

and power $=0$01. Mayo’s severity function is then used to assess whether a non-rejection is genuinely probative against departures of practical importance. The review also notes Mayo’s criticism of interpreting a $=0$02-value as a posterior probability that $=0$03 is true, and her objection that Bayesian procedures do not directly guarantee the same long-run error control (Robert, 2011).

Debates over Mayo’s work are especially prominent in discussions of Birnbaum’s theorem. Birnbaum’s 1962 claim was that the Sufficiency Principle and the Conditionality Principle imply the Likelihood Principle. Deborah Mayo’s critique argues that violations of the Likelihood Principle need not violate either sufficiency or conditionality, because Birnbaum’s proof over-extends conditionality across hypothetical mixtures. Martin and Liu use this to reopen foundational questions and propose an inferential model framework based on validity and efficiency principles (Martin et al., 2013). Fraser’s discussion similarly argues that standard, within-experiment forms of conditioning, especially in continuity-based inference using natural ancillaries and second-order $=0$04-value functions, do not force acceptance of the strong Likelihood Principle (Fraser, 2014).

The controversy is therefore not about whether sufficiency, conditioning, and likelihood matter, but about how they are formulated and whether they jointly compel likelihoodism. In that literature, “Mayo” denotes a substantive philosophical position rather than an institution or scoring system (Gelman et al., 2019).

6. Mayo et al. in computational materials science

A further use of “Mayo” occurs in materials modeling. A 2018 arXiv comment on a 2017 study by Mayo et al. concerns ab-initio random structure searching of Li–Sn intermetallics at $=0$05 GPa and $=0$06 K. According to the comment, Mayo et al. recovered experimentally known phases and reported two previously unknown stable structures on the convex hull: Li$=0$07Sn$=0$08 in $=0$09 and Li$=0$10Sn$=0$11 in $=0$12. They also reported ten metastable compositions within 20 meV/atom of the hull (Sen et al., 2018).

Sen et al. re-evaluated the CIF files under tighter DFT convergence and phonon-stability checks and reported widespread contradictions in space-group assignment. The comment highlights five disputed compositions. For example, Li$=0$13Sn$=0$14 is reported by Mayo et al. as “P3$=0$15” in the Abstract and Table 1 and “P3m1” in Supporting Information, whereas Sen et al. find $=0$16. Li$=0$17Sn$=0$18 is reported as “$=0$19” in the Abstract and Table 1 and “Cmmm” elsewhere, whereas Sen et al. identify $=0$20. The comment attributes the disagreements to loose convergence criteria, inconsistent reporting across text, tables, and CIF archives, and incorrect CIF formats (Sen et al., 2018).

The significance of this dispute is practical. Hull-contacting phases in Li–Sn determine predicted voltage plateaus, volume expansion, and Li-ion diffusivity in tin-anode models. The comment therefore treats the Mayo et al. paper not as a terminological curiosity but as a contested contribution to first-principles materials prediction (Sen et al., 2018).