Page 10 - Read Online

Page 135 Bektaş et al. Art Int Surg 2022;2:132-43 https://dx.doi.org/10.20517/ais.2022.20

Risk of bias assessment
The methodological quality assessment of included studies was independently performed by two reviewers
[20]
(Bektaş M, Costa Pereira J) using the ROBINS-I assessment tool . This tool measures the risk of bias in the
domains: confounding, participant selection, intervention classification, deviations from intended
interventions, missing outcome data, measurement of outcomes, and selection of reported results. Based on
these domains, overall risk of bias is determined for each study. Moreover, the PROBAST risk of bias tool
was used to evaluate the quality of ML models within studies . Risk of bias domains included participant
[21]
selection, predictors, outcomes, and analysis.

Data synthesis and outcome assessment
The following data aspects were independently retrieved from each study by two reviewers (Bektaş M, Costa
Pereira J): first author, year, country of research, number of patients, study design, surgical procedure, type
of ML, purpose of ML, outcome measurements, and predictive performance. The categorization of studies
was based on surgical domains, such as liver, biliary, and pancreatic surgery. Subsequently, accuracies of ML
studies were reported within each surgical domain. For each study, the mean accuracy (ACC) and area
under the curve (AUC) were calculated to represent the predictive performances of ML models. In addition,
descriptive statistics were used to calculate the median and range of accuracies for every ML model.

RESULTS
The search strategy identified a total of 1821 studies after the disposal of duplicates [Figure 1]. The 1821
studies were screened for eligibility based on the title and abstract. Subsequently, 104 studies were eligible
for full-text assessment, resulting in the inclusion of 52 studies.

Several subclasses of ML have been used within HPB surgery. A vast majority of studies have applied Neural
Networks models (n = 16), Radiomics (n = 13), or multiple ML methods (n = 13). Remaining studies
involved Decision Trees (n = 7), GBM (n = 1), Random Forest (n = 1), and SVM (n = 1).

Within studies addressing liver surgery, studies predominantly involved hepatocellular carcinomas (HCC)
(n = 21), intrahepatic cholangiocarcinomas (ICC) (n = 5), and colorectal liver metastasis (CRLM) (n = 3).
Remaining studies perihilar cholangiocarcinomas (PHCC) (n = 1), and extrahepatic cholangiocarcinomas
(ECC) (n = 1). Pathologies in biliary surgery included acute cholecystitis (n = 2), gallstones (n = 2), and
cholesterolosis and polyps (n = 1). For pancreatic surgery, studies included pancreatic cancer (n = 5),
pancreatic fistulas (n = 4), intraductal papillary mucinous neoplasms (IPMN) (n = 2), and acute pancreatitis
(n = 1). Additionally, four studies have not specified the pathology that was treated.

The purposes of ML algorithms mostly included predicting the course of disease (n = 26), postoperative
complications (n = 13), diagnosis (n = 4), and intraoperative complexities (n = 3). Additionally, ML was
used to determine essential predictors (n = 5) and to predict the postoperative quality of life (n = 1).

An overview of study characteristics for liver, biliary, and pancreatic surgery is separately presented in
Supplementary Tables 5-7, respectively. In addition, the median and range of accuracies for included ML
models are presented in Figure 2.

Risk of bias assessment
Within the 52 included articles, 47 (90%) retrospective cohort studies have been detected. Additionally, five
prospective cohort studies (10%) were present. Therefore, only the ROBINS-I assessment tool was used for
the methodological quality assessment. It was discovered that most studies (92%) received a low overall bias,