Page 9 - Read Online

Bektaş et al. Art Int Surg 2022;2:132-43 https://dx.doi.org/10.20517/ais.2022.20 Page 134

Table 1. Definitions AI methods
General term
Machine Learning Machine Learning is an umbrella term and is commonly described as computational techniques that are able to perform
(ML) complex tasks by analyzing large-scale data [6]
Analytical approaches
Decision Tree Within Decision Tree models, data is divided into smaller nodes and branches. Each node represents a variable, and each
branch contains a feature of the variable. Features contain two outcomes such as yes or no. Following one outcome at
each feature will eventually form the prediction tree for the desired task. In the end, the smallest tree that optimally fits
[7]
the data will be produced
Gradient Boosting Gradient Boosting models begin with forming a model that fits the data. Afterwards, a consecutive model is constructed
(GBM) that concentrates only on inaccurately predicted aspects of data. Models are then combined to form an improved model.
[8]
This process is repeated until a final model is established with a minimal error in prediction
Random Forest In a Random Forest model, Decision Trees are present for the desired outcome. Each Decision Tree contains a different
prediction path based on the values for the selected variables. By combining all Decision Trees, the final most accurate
model will be built [9]
Support Vector Support Vector Machines are capable of making predictions by finding the optimal border to classify variables or
Machine (SVM) outcomes in two groups [10]
Artificial Neural Artificial Neural Networks are models in which datasets are analyzed by multiple processing layers. In each layer, features
Networks (ANNs) of each data point are extracted to recognize patterns; these features contain weighting factors within each layer. After
[11]
repeating this training process on multiple datasets, a final model is produced for the complex task
Convolutional Neural Convolutional Neural Networks are similar to ANNs, except these models do not use weights for extracted features.
Networks (CNNs) Instead, specific filters are applied to detect patterns in datasets. Additionally, connections are present to provide
[12]
feedback in each training process
Deep Learning Deep Learning algorithms function similarly to Neural Networks; however, Deep Learning models have more layers or
depth than Neural Networks [13]
Area of AI that could benefit from ML
Radiomics In Radiomics models, images are analyzed to detect various quantitative features. Afterwards, these features are used for
predictions or associations of several medical outcomes [14] (editors, comment #4)

METHODS
Search strategy
Literature was retrieved and systematically reviewed in conformity with the PRISMA guidelines and
Cochrane Handbook for Systematic Reviews of Interventions version 6.0. Databases PubMed, Embase.com,
Clarivate Analytics/Web of Science Core Collection, and the Wiley/Cochrane Library were used to perform
a systematic search. The timeframe within the databases was from inception to the 7th of July 2021. The
systematic search was performed by Bektaş M and Burchell GL. The search included keywords and free text
terms for (synonyms of) “Machine Learning” combined with (synonyms of) “digestive system surgical
procedures”. A comprehensive overview of the search terms per database is available in the supplementary
materials [Supplementary Tables 1-4]. The search and protocol of this review were not registered in
PROSPERO.

Study selection
During the initial step, articles were included when they described ML within general surgery to secure
studies with overlapping content. Subsequently, studies were only qualified if they met the following
criteria: (1) describing ML methods within HPB surgery; (2) clinical study; and (3) conducted on adults.
Studies were excluded when they: (1) reported on reviews, children, and study abstracts; (2) described
regression models; and (3) were not written in English. No specific study design was preferred in the
inclusion criteria. Two reviewers (Bektaş M, Costa Pereira J) independently performed the title and abstract
screening in conformity with the inclusion and exclusion criteria. Studies were qualified for full-text
screening when both reviewers agreed on inclusion. Disagreements were resolved by means of discussion
between reviewers, resulting in an agreement.