Page 19 - Read Online

Collins et al. Art Int Surg. 2025;5:254-69 https://dx.doi.org/10.20517/ais.2024.92 Page 264

AI models must be seamlessly integrated into the clinical workflow to provide real-time assistance without
disrupting surgical autonomy or established protocols. A critical step in this transition is prospective clinical
validation, often through randomized controlled trials (RCTs) or cohort studies. These studies would
compare AI-assisted hernia repairs to traditional methods, assessing key outcome metrics such as
complication rates (e.g., infection or recurrence), surgical efficiency (e.g., operation time), and long-term
patient recovery, including pain management and return to daily activities. However, conducting such
studies is time-intensive and resource-demanding. Moreover, integrating AI into AWS requires
compatibility with hospital IT infrastructure, including EHRs and surgical planning tools.

It is also important to recognize the importance of developing AI systems that are robust enough to adapt to
anatomical variations. One strategy to improve AI’s adaptability in these cases is the inclusion of diverse
datasets during the training phase, incorporating a wide range of anatomical variations and patient-specific
characteristics. While this mitigates the issue, it may not completely eliminate it. Continuous monitoring
and recalibration of AI models, based on real-time feedback and postoperative data, will help improve the
model’s ability to provide accurate guidance across diverse patient populations. A promising future research
direction is to develop AI systems to flag cases where anatomical features deviate significantly from the
norm, prompting the surgeon to make additional assessments or adjustments.

To ensure accountability, prospective validation studies must meticulously document surgical decisions
both with and without AI assistance. This allows researchers to directly attribute any observed outcome
changes to AI intervention. Furthermore, long-term follow-up is crucial, as AI’s true impact may take time
to manifest. In hernia repairs, for instance, complications like chronic pain, mesh-related issues, or
recurrence may only emerge months or years post-surgery.

AI technologies in AWS, particularly hernia repair, hold great promise, but significant cost-effectiveness
challenges remain. While university research often produces the initial R&D results and methods, these
innovations are typically commercialized by medtech companies, which fund further development, clinical
trials, and validation. Hospitals, while not responsible for the R&D costs, must still invest in integrating
these technologies into their existing infrastructure. This includes upgrading hardware, software, and
training staff - expenses that can be substantial, especially if immediate benefits do not translate into clear
financial returns.

A critical barrier to AI adoption in hernia repair is insurance reimbursement. Insurers may hesitate to cover
these higher-cost procedures without clear evidence of cost savings or improved outcomes. The long-term
benefits of AI - such as reduced follow-up surgeries or complications - may not be immediately reflected in
short-term cost metrics, making it harder for hospitals to justify the additional investment. Reimbursement
models need to adapt to consider both short-term expenses and long-term benefits, aligning the financial
interests of hospitals, insurers, and medtech companies. Hospitals require clear financial incentives,
including reimbursement models that reflect both immediate costs and long-term benefits. Demonstrating
AI’s ability to improve patient outcomes and reduce long-term healthcare costs will be essential in
overcoming cost-effectiveness and reimbursement challenges, ensuring the broader adoption of AI in AWS.

Leveraging AI models with managed uncertainty: imperfection does not invalidate utility
It is important to state that in general, an AI model does not need to be perfect to be useful. Any ChatGPT
user will attest to that. Indeed, if this were not true, research in AI and surgery should have stopped long
ago, because the complexity of surgery data and surgical tasks practically ensures that AI model outputs will
have at least some uncertainty. The AI systems described in this article will be deployed in clinical practice