Page 49 - Read Online

Page 453 Goyal et al. Art Int Surg. 2025;5:450-64 https://dx.doi.org/10.20517/ais.2024.108

Data extraction
Three independent reviewers performed the data extraction using a standardized data collection form.
Extracted variables included:

• Study characteristics (authors, year of publication, country, study design)

• Procedure details (TAPP, TEP, laparoscopic vs. robotic, unilateral vs. bilateral)

• AI model specifics (model type, architecture, training dataset, ground-truth annotations)

• Performance metrics (accuracy, F1 score, precision, recall, latency)

• Surgical phase recognition outcomes

• Number of videos analyzed, surgeons involved, and video collection period

Any disagreements during data extraction were resolved by consensus or consultation with a third reviewer.

Bias and quality assessment
The Newcastle-Ottawa Scale (NOS) was used to assess the risk of bias in the included studies. Three
independent reviewers rated each study based on three domains: selection, comparability, and outcome
assessment. Studies scoring 7-9 were considered high quality, 4-6 as moderate quality, and < 4 as low
quality . Inter-rater reliability was assessed using the Fleiss’ kappa analysis. Discrepancies were resolved
[12]
through discussion or input from a third reviewer.

Outcomes and effect measures
The primary outcomes sought were AI model performance metrics: accuracy, F1 score, precision, recall,
and latency. The F1 score, a metric that combines precision and recall, was specifically collected for studies
employing classification models. Secondary outcomes included surgical phase recognition performance and
details of edge computing devices, where applicable.

Data synthesis and presentation
Due to the heterogeneity of AI models and outcome measures, a narrative synthesis was performed. Data
were tabulated to summarize study characteristics and outcomes. Meta-analysis and sensitivity analysis were
not conducted due to insufficient data and variability in study methodologies. Findings are presented in
descriptive tables and figures to facilitate interpretation.

RESULTS
Study selection
A comprehensive search was conducted across six databases, yielding 903 records in the initial literature
search. After removing 169 duplicates, 734 records remained for screening. Based on pre-determined
inclusion and exclusion criteria, we conducted a title and abstract review, selecting six studies for full-text
analysis. Of these, one study was excluded because it utilized images for surgical phase recognition . After
[13]
a thorough full-text review, data were extracted from five articles, all published in English. Additionally, a
citation search was performed to ensure comprehensive coverage of relevant literature, identifying five
additional references within the scope of our study. Upon reviewing these, one was excluded as it was a
review article , and three others were excluded for employing image-based surgical phase recognition
[14]