From sequential prediction to clinical utility: Reframing admission-time length-of-stay modeling for ICU care

Dear Editor,

We read with great interest the study by Chang et al., which compared an ImageNet-pretrained vision transformer (ViT) with human endoscopists for predicting endoscopic and histologic activity in ulcerative colitis (UC) from white-light colonoscopy videos. ¹ The study is clinically relevant because treat-to-target strategies in UC increasingly require reproducible assessment of mucosal healing and deeper inflammatory control. ² Its major value lies not only in reporting competitive diagnostic performance, but also in exploring whether routinely available white-light endoscopy can support AI-assisted inference of histologic activity without specialized imaging systems.

A key methodological issue, however, is whether the selected model architecture truly matches the data form and clinical task. Although the study is framed as video-based AI, the implemented pipeline appears closer to frame-level image classification: original 512 × 512 images were resized to 224 × 224, and videos were downsampled from 30 to 3 frames per second before ViT inference. ¹ This design may weaken precisely the visual cues that matter in UC, such as fine vascular pattern loss, granularity, superficial erosions, friability, and subtle mucosal texture. ViTs are powerful for global dependency modeling, but their patch-based tokenization and natural-image pretraining do not automatically guarantee preservation of small, clinically decisive endoscopic features. ³ Therefore, the superiority over ResNeSt does not by itself prove that a standard ViT is the optimal inductive bias for UC videos. Future work should include ablations on input resolution, patch size, frame sampling rate, mucosa-focused region selection, and temporal aggregation. A temporal transformer, video-level multiple-instance learning framework, or quality-gated high-resolution tiling strategy may better reflect how endoscopists integrate dynamic mucosal inspection across a segment.

A second issue concerns the label mechanism. Predicting histologic healing from white-light endoscopy is not a conventional fully supervised image task. Histology is derived from biopsy samples, whereas the video segment may contain heterogeneous mucosal regions, motion, blur, fluid, specular highlights, and adjacent areas not sampled by biopsy. Thus, the training signal is inherently weakly supervised and spatially imperfect. The use of binary Nancy Histological Index thresholds simplifies the endpoint, but does not resolve label noise or spatial mismatch between biopsy tissue and video frames. The class-distance weighted loss addresses ordinal class separation, yet it does not explicitly model uncertainty in the reference standard. More robust strategies could include biopsy-site localization, frame-to-biopsy spatial registration, multi-biopsy aggregation, soft labels reflecting pathologist uncertainty, and multiple-instance learning in which a video clip is treated as a bag of heterogeneous observations. Such designs would help distinguish genuine mucosal phenotype learning from local annotation conventions or biopsy sampling bias. Previous video-based UC AI studies also highlight the importance of prospective video validation and clinically grounded endpoints. ⁴

Third, translation into clinical practice requires more than accuracy, AUROC, sensitivity, specificity, and F1 score. For treatment escalation, de-escalation, or biopsy triage, a false prediction of healing may carry very different clinical risk from a false prediction of activity. Model calibration, Brier score, expected calibration error, decision-curve analysis, threshold-specific net benefit, and uncertainty-guided abstention should therefore be reported. ⁵ Similarly, Grad-CAM examples are useful for illustration, but heatmaps should not be equated with faithful clinical reasoning unless validated against expert-defined mucosal regions, occlusion tests, counterfactual perturbations, or biopsy-relevant areas. Generalization should also be specified algorithmically: device shift, imaging-protocol shift, bowel-preparation and image-quality shift, population shift, and annotation-culture shift may each affect model behavior differently.

In summary, Chang et al. provide an encouraging step toward AI-assisted UC activity assessment. The next stage should move from performance comparison toward robust, interpretable, calibrated, uncertainty-aware, and workflow-compatible endoscopic AI systems that can support—not merely imitate—clinical decision-making.

ORCID iD

Kaijian Yang https://orcid.org/0009-0007-9502-1654

Data Availability Statement

Not applicable to this article.*