Page 44 - Read Online

Page 2 Shapey et al. Art Int Surg 2023;3:1-13 https://dx.doi.org/10.20517/ais.2022.31

INTRODUCTION
Machine Learning (ML) relates to the use of computer-derived algorithms and systems to enhance
knowledge in order to facilitate decision making. In surgery, ML has the potential to shape clinical decision
making and the management of postoperative complications in three ways: (a) by using the predicted
probability of postoperative complications or survival to determine and guide optimal treatment; (b) by
identifying anomalous data and patterns representing high-risk physiological states during the perioperative
period and taking measures to minimise the impact of the existing risks; (c) to facilitate post-hoc
identification of physiological trends, phenotypic patient characteristics, morphological characteristics of
diseases, and human factors that may help alert surgeons to relevant risk factors in future patients. Here we
aim to review the potential clinical relevance of ML to improving the prediction of postoperative
complications in hepato-biliary and pancreatic surgery.

THE CURRENT LANDSCAPE OF PREDICTING POSTOPERATIVE COMPLICATIONS
Preoperative prediction
The occurrence of postoperative complications in pancreatic surgery is a major determinant of outcomes,
not least because of the impact of complications on the non-completion of adjunctive therapies. Predicting
postoperative complications prior to making a commitment towards surgical therapy is important because
it has the potential to change the sequence of therapies provided and the options considered. Several
decision making dilemmas exist in pancreatic surgery, which include debates surrounding upfront
chemotherapy vs. upfront surgery for pancreatic ductal adenocarcinoma (PDAC), fast-track surgery vs.
preoperative biliary drainage for head of pancreas tumours, and parenchymal preserving vs. oncological
[1-3]
resection in small neuro-endocrine tumours amongst many others . It is also challenging to determine the
resectability of malignant disease of the pancreas and ML could help play a role in reducing futile surgery -
this could have a beneficial impact on patients in terms of reducing avoidable morbidity on the one hand
and maximising healthcare resources on the other. There is also a need to improve the interpretation of
complex multivariable patterns that represent clinical response to chemo- and immuno-therapies so that
the treatment regimens and the timing of surgery could be optimised.

In liver surgery, accurate preoperative prediction of post-hepatectomy liver failure and the functioning liver
remnant (FLR) could change decision making by supporting the use of adjunctive methods to increase the
FLR or by counselling against higher-risk surgery in favour of other lower-risk therapeutic options (e.g.,
[4-8]
ablation or hepatic artery pump chemotherapy) where the difference in outcome may be equivocal . ML
could help identify patients better suited to more aggressive therapeutic options such as transplantation and
predict which grafts and recipients are at higher risk of failure, immune rejection and mortality.

In considering the potential application of ML to predict postoperative complications prior to surgery, it is
helpful to appreciate the limitations of existing models, which are primarily based on regression analyses.
Three important limitations of regression based scoring systems that are commonly encountered include:
(a) insufficient statistical power, often arising when the number of recorded events relative to outcomes is
low; (b) when the traditional rules of frequentist classical statistics are not met, e.g., the 10-to-1 rule of 10
events for each variable included in a multivariable model; (c) where reporting of the area under the curve
(AUC) is not accompanied by the standard error and p-value when making direct comparisons between
models; (d) where a new variable is added to existing prediction models, but the discretional value of the
additional variable is not evaluated through techniques such as Net Reclassification Improvement .
[9]
Regression models have struggled to translate data related to predictor variables into robust and reliable
tools to improve decision making in “real-world” situations .
[10]