Page 59 - Read Online

Goyal et al. Art Int Surg. 2025;5:450-64 https://dx.doi.org/10.20517/ais.2024.108 Page 462

2
and fairness in model comparison. In addition, integrating quantitative measures of heterogeneity, such as I
2
statistics, τ values, and the coefficient of variation of reported accuracies into future reviews could help
identify underlying sources of inconsistency and guide model selection.

Advanced training strategies such as federated learning and self-supervised approaches may reduce the
dependence on manually labeled data while ensuring data privacy across institutions. Further research
should focus on the development of real-time AI systems with low-latency performance using edge
computing hardware such as Nvidia Jetson or A100 GPUs. These systems should be validated prospectively
in real-world, multicenter clinical settings and evaluated not only for model accuracy but for their impact
on surgical workflow, efficiency, safety, and outcomes.

Equally important is the pursuit of explainable AI to foster trust and transparency. Future models should
incorporate interpretable outputs, including heatmaps, attention weights, and confidence scores, to assist
surgical decision making and training. Regulatory pathways, including alignment with FDA and
international guidelines on software as a medical device, must be actively considered during development
and deployment. Interdisciplinary collaboration among surgeons, engineers, regulatory authorities, and
ethicists will be crucial in ensuring safe, effective, and equitable integration of AI into operative practice.

From our perspective, this systematic review represents a foundational step toward shaping the future of AI
in surgical education and operative care. We envision that real-time AI-assisted surgical phase recognition
will become an essential tool in residency training, competency evaluation, and intraoperative decision
support. For early-career robotic surgeons and general surgery trainees, such systems could offer objective,
adaptive feedback that enhances skill acquisition. The future of surgical AI lies not only in advancing
algorithmic performance but also in establishing the human-AI partnership at the core of modern surgical
practice. With early, coordinated investment in algorithm development, dataset curation, and prospective
validation, we believe this vision can be realized.

In conclusion, This systematic review underscores the transformative potential of AI and ML in advancing
real-time surgical phase recognition during minimally invasive IHR. While significant progress has been
achieved across laparoscopic and robotic procedures, the application of AI in IHR remains in its early
stages. The included studies demonstrate encouraging performance metrics, with models like TeCNO,
ResNet-50, and Video Transformer Networks achieving promising accuracies. However, challenges such as
heterogeneity in model architectures, reliance on labor-intensive annotated datasets, and variability in
surgical techniques must be addressed to enhance the clinical utility of AI. Future research should focus on
developing standardized benchmarks, conducting prospective multicenter validations, and leveraging
hardware advancements to enable latency-free, real-time applications. Collaborative efforts among
surgeons, AI developers, and policymakers will be pivotal to overcoming existing limitations, ensuring
ethical implementation, and integrating AI-driven systems into surgical workflows. By doing so, AI has the
potential to revolutionize surgical training, improve technical performance, and ultimately enhance patient
outcomes in IHR and beyond.

DECLARATIONS
Authors’ contributions
Conceptualization and study design: Goyal A, Oviedo RJ
Data acquisition and curation: Goyal A, Mendoza M, Munoz AE, Macias CA
Statistical analysis and data interpretation: Goyal A, Mendoza M, Munoz AE, Macias CA