Page 14 - Read Online

Page 139 Bektaş et al. Art Int Surg 2022;2:132-43 https://dx.doi.org/10.20517/ais.2022.20

Machine Learning models have demonstrated accuracies ranging from 63% to 89% for predicting
postoperative complications after HPB surgery. Although clinical risk prediction models have been
developed to detect postoperative complications, these models have not shown significant improvements
[79]
compared to the surgeon’s assessment . In addition, conventional regression models have predicted
postoperative complications with AUCs up to 0.74 [80,81] . Due to its promising predictive performances, ML
has illustrated the potential to surpass the surgeon’s assessment and conventional statistics. Ideally, ML
models could facilitate the implementation of prophylactic measures and improved patient monitoring
based on the predicted complications. This could prevent delayed hospital discharge for patients with severe
postoperative complications.

Machine Learning has shown accuracies between 79% and 98% for the diagnosis of HPB pathologies such as
gallstones and IPMNs. Meanwhile, logistic regressions have demonstrated accuracies between 73% and 77%
in predicting these outcomes [82,83] . As ML seems to be superior in predictive capacities, these models could
be used to preoperatively recognize patients with these pathologies, enabling the possibility to track the
most important risk factors early and improve patient monitoring. In addition, intraoperative complexities
and conversions during laparoscopic cholecystectomies have been predicted by ML models with accuracies
up to 89%, whereas logistic regression models have shown accuracies up to 83% . Recently, computer
[84]
vision models have been developed to locate anatomic landmarks and assess the grading of operative
complexities during laparoscopic cholecystectomies [85,86] . Predicting conversions and detecting complexities
during operations by using AI models could support intraoperative decision-making and secure optimal
patient safety.

For many years, conventional statistical models have been trained to predict surgical outcomes after HPB
surgery. Most ML models in this review have shown median AUCs above 0.8, possibly indicating better
discriminative abilities than conventional statistics for predicting surgical outcomes after HPB surgery.
Furthermore, ML models are able to perform better predictions if the number of input variables is large,
whereas conventional statistics function optimally with a few variables . Since clinical databases are
[87]
complex and usually contain many variables, ML would be preferred for the analysis of clinical data.
However, clinicians have been experiencing difficulties in understanding and interpreting ML methods,
which is also called the “black-box problem” . This problem could eventually hinder the development and
[88]
implementation of ML models.

This review has some limitations. Due to inconsistencies in applied ML frequencies, mean accuracies might
be underrepresented for a few ML models. Additionally, some of the studies have not reported accuracies or
AUCs for the ML algorithm; therefore, a meta-analysis could not be performed.

[89]
Since predictive accuracies above 70% indicate good discriminative abilities , ML algorithms within this
review seem to have promising predictive capabilities for outcomes after HPB surgery. However, as most
studies (85%) are missing external validation for the ML algorithms, the generalizability of these models is
not supported. The clinical integration of ML could be dependent on this external validation. Therefore,
future studies should focus on gaining external validation, which could be facilitated by retrieving large
datasets from available patient databases. In addition, interdisciplinary collaborations could be essential in
solving this “black-box problem” and support the development of efficient ML models. Data scientists and
clinicians should share clinical data to ensure proper data arrangements, data processing, and transparency
in methodologies. Sharing data between medical fields might improve the accuracy of ML models and
facilitate the procurement of external validation .
[90]