AI IN HEALTHCARE

Nevertheless, there are several blind spots that deserve further scrutiny. First, while the review acknowledges that most gastric cancer datasets are Asia-centric, it fails to quantify algorithmic bias or propose equity metrics for cross-ethnic deployment. Therefore, cross-ethnic drops in positive-predictive value could be expected when models trained on East Asian cohorts are deployed in Western populations. To future-proof such systems, equity metrics, such as disaggregated calibration curves or subgroup-specific false-negative rates, should be reported alongside global AUC. Second, the economic costs of training ever larger multimodal models are mentioned only tangentially, even though these factors will strongly influence adoption in resource-limited settings. Third, environmental sustainability warrants equal attention. Truhn and colleagues shed light on the ecological footprint of medical AI, particularly in radiology, where the training and deployment of large neural networks demand substantial energy and emit significant CO2 [10]. Based on their estimates, an average EU resident produces about 6.8 t CO2 annually, and running a single MRI scanner for one year emits roughly 58 t CO2, while training a model the size of GPT-3 yields a comparable footprint to operating an MRI machine for 9.5 years or flying 262 round trips between Munich and New York [10]. Although a single inference on images of a patient may release only ~0.5 g CO2, which is tiny compared to the 14,600 g generated by acquiring the MRI scan itself, the authors warn that large-scale clinical deployment could magnify these emissions to consequential levels [10]. To mitigate this, they urge adoption of efficiency-oriented model compression, high-efficiency or renewable-powered data centers, specialized low-power inference hardware, and the strategic use of healthcare purchasing and investment decisions, such as divesting from fossil fuels, to accelerate sector-wide decarbonization. Potentially, by publishing ‘Model Cards' that include demographic performance slices, separate models for different subgroups, and carbon emission data, gastroenterology AI could align with the EU AI Act's risk-based transparency requirements [11] and the WHO guidance on Ethics and Governance of AI for Health [12]. These cards should be in compliance with Article 32 of the GDPR, which requires controllers and processors to apply risk-proportionate technical and organizational safeguards. This includes the pseudonymization and encryption of personal data, ensuring ongoing confidentiality, integrity, availability and resilience of processing systems, the ability to restore data access in a timely manner after incidents, and a process for regularly testing and evaluating security measures [13].

The first frameworks for these ‘Model Cards' have the potential to increase transparency and accountability within the ML community by including metadata on model details, underlying assumptions, intended use, potential risks or harms, demographics, performance, evaluation data, training data, and ethical considerations. These were proposed some years ago [14,15]. However, these templates now require optimization. In particular, they should incorporate standardized energy-use reporting, life-cycle versioning, automated alerts for performance drift, harmonized fairness metrics that align with the EU AI Act's "high-risk” criteria, and machine-readable sections that can be directly plugged into regulatory submissions and electronic health record audit logs. Continuous community feedback, iterative updating, and formal alignment with emerging international standards (e.g., ISO/IEC 42001) [16] and the NIST AI RMF [17] will be essential to ensure that Model Cards evolve from static disclosure documents into living instruments of accountability and sustainability. Future work should integrate cost-effectiveness analysis and carbon accounting into model-reporting checklists. A more complex issue, briefly mentioned in the review, is determining who will be responsible for payment when an “invisible" algorithm replaces a billable human action. In radiology we have seen both ends of the spectrum: computer-aided detection for screening mammography received a separate current procedural terminology (CPT) add-on in 2004, but the extra payment was later revoked when its added value was minimal. In contrast, the stroke triage platform Viz for large-vessel occlusion (LVO) [18] received a New-Technology Add-on Payment of up to $1040 per use from the U.S. Centers for Medicare & Medicaid Services, and the autonomous FDA-approved fundus photography algorithm IDx-DR used for effective diabetic retinopathy screening was assigned its own CPT code (92229) in 2021 [19]. Will a vision-transformer that identifies sub-centimeter gastric neoplasia be reimbursed like a laboratory test, included in the endoscopy DRG, or licensed to hospitals on a software-as-a-service basis? Clear guidance will be crucial, as should payers opt for "bundling" AI outputs without additional compensation, adoption may slow down despite proven clinical benefits. Conversely, if every inference comes with a fee, there is a risk of perverse incentives that could increase healthcare costs without improving outcomes.

Many AI systems to improve clinical workflows are currently awaiting approval and the incorporation of defined reimbursement models for AI costs into Diagnosis-Related Groups (DRGs) is pending. Meanwhile, value-based care contracts link reimbursement to measurable outcome improvements. In the future, reimbursement models that can be integrated into current payment schemes might include per-patient rates, direct payment, value-based payments (e.g., increased rate of adenoma detection, reduced readmissions), targeted competition rewards, transitional add-on payments, and performance standards [20]. Early evidence from a U.S. integrated delivery network demonstrates that integrating polyp-detection AI into a capitated colonoscopy bundle would be able to reduce the average colonoscopy cost and the gross annual reimbursement for colonoscopies by 18.9% and US$149.2 million in Japan, 6.9% and US$12.3 million in England, 7.6% and US$1.1 million in Norway, and 10.9% and US$85.2 million in the United States, respectively, compared with the resect-all-polyp strategy [21]. This illustrates that properly aligned incentives can reward both innovation and stewardship. In European countries, reimbursement processes can be complex. In Germany, the digital healthcare act allows for AI diagnostic listing in the digital health application (DiGA) and reimbursement by statutory health insurance. In France, AI-based medical devices are assessed by the National Authority for Health and the price is negotiated by the Minister of Health. Reimbursement is based on a classification system that determines the level of reimbursement for a device [22]. In the UK, reimbursement for AI-based medical devices is covered by the National Health Service (NHS) if approved by the National Institute for Health and Care Excellence [22]. These models spread costs across large patient populations and tie bonus payments to real-world quality metrics, such as reduced readmission rates or shorter lengths of stay. To navigate these new pathways, future AI submissions should routinely include cost-utility and budget impact analyses. This will provide regulators and insurers with transparent evidence that algorithmic assistance offers a net clinical benefit without inflating overall healthcare expenditure. Equally unsettling is the data safety gap that arises when the servers, Graphics Processing Units (GPUs), and code bases supporting an AI diagnostic are managed by third-party vendors who are not part of the formal circle of medical confidentiality. These engineers often require privileged "root" access for updates, potentially exposing protected health information to networks and devices outside hospital firewalls, increasing GDPR-related privacy risks. Robust governance frameworks, including zero-trust architectures, real-time audit logging, and legally binding business associate agreements, will be crucial for patient safety, just like the accuracy of the algorithms themselves. For clinicians, the message is that AI tools are poised to migrate from experimental add-ons to workflow defaults; one example is a polyp-detection module that adds merely 30 s to a colonoscopy yet reduces miss rates substantially. For investigators, the review serves as a grant-writing blueprint: proposals will increasingly be judged on how convincingly they integrate heterogeneous data and how transparently they link algorithmic features to biological mechanisms. Looking forward, closed-loop digital twins that fuse real-time ctDNA methylation, spatial transcriptomics, and wearable data could make dynamic treatment adaptation a reality for both inflammatory bowel disease flares and HCC surveillance. Achieving that vision will require standardized acquisition protocols, interpretable architectures, and regulatory sandboxes that can accommodate continuous model updating without compromising patient safety. The application of multimodal data and deep learning techniques, which have the capacity to simultaneously evaluate vast quantities of complex data, including medical images and electronic health records has been shown to be helpful in the diagnosis, prediction, and treatment response assessment of hepatocellular carcinoma (HCC) [23]. Similarly, AI-powered bowel sound analysis has been shown to be a promising, non-invasive tool for diagnosing, monitoring, and managing inflammatory bowel disease (IBD), including Crohn's disease and ulcerative colitis [24]. To integrate multimodal AI into day-to-day care, clinician-retraining modules should be offered. CME-accredited workshops that walk endoscopists through hands-on cases and outline fallback procedures when the algorithm fails would also be beneficial. Interoperability can be ensured by exposing model outputs via an electronically exchange healthcare data system, allowing for the removal of critical barriers and trusted data flows between a wide range of different healthcare information systems such as the interoperability standard Fast Healthcare Interoperability Resources (FHIRs) developed by the Health Level 7 standards organization (HL7) [25-27]. This will allow for the automatic flow of probability scores, heatmaps and audit trails into the electronic endoscopy report. Finally, a ‘continuous-learning' governance pathway is needed in which hospital change-control boards treat each new model weight as regulated ‘digital firmware', approving version upgrades only after silent shadow testing confirms non-inferiority on local data and updating rollback plans are in place. Wu and colleagues have provided a compass that not only celebrates current accomplishments but also pinpoints the infrastructural and ethical ground we still must cover if multimodal Al is to fulfill its transformative promise in digestive disease care. To turn that direction into day-to-day reality, the gastroenterology community now needs a practical itinerary, one that ensures multimodal AI becomes trustworthy, affordable and globally accessible. Five priorities stand out: Standardized data capture and formats: Ensure that endoscopy videos, imaging files, pathology slides and multi-omics results are generated with shared protocols and vendor-neutral standards so that data remain comparable across centers. Explainable modeling: Combine high-performance algorithms with transparent, clinician-friendly rationales, including heat-maps, feature rankings or rule overlays, so users can verify and trust each prediction. Privacy-preserving collaboration: Engage in federated-learning networks that keep data behind local firewalls while sharing model updates, enhancing performance without compromising patient privacy. Outcome-aligned reimbursement: Collaborate with payers to incorporate AI costs into diagnosis-related groups or value-based contracts, rewarding measurable benefits (e.g., higher adenoma-detection rates) rather than imposing per-click charges. Fairness and sustainability tracking: Release or publish model cards that disclose subgroup accuracy and CO2 footprints, and implement mitigation carbon-offset strategies to ensure fair and environmentally responsible innovation.

Multimodal AI is rapidly transitioning from research laboratories to everyday practice in gastroenterology and hepatology. It integrates endoscopic video, radiomics, whole-slide pathology, and multi-omics data to create cohesive decision-making platforms. This commentary acknowledges these advancements while also addressing key issues: dataset variability, “black box" opacity, ethical concerns, uncertain reimbursement processes, and potential risks to third-party data security. To promote the application of AI in Gastroenterology and Hepatology, federated learning consortia, explainable neural-symbolic models, standardized data collection protocols, and fair payment structures will be essential to ensure safe, transparent and universally accessible AI-driven care for digestive diseases.