Page 82 - Read Online

Page 40 McGivern et al. Art Int Surg 2023;3:27-47 https://dx.doi.org/10.20517/ais.2022.39

The intersection of AI and big data in HPB surgery
We identified eleven studies utilizing large datasets in HPB surgery applications [Table 4]. Nine have been
published since 2020. Eight of the 11 identified papers utilized NLP to extract data from large numbers of
reports, mainly with the aim of identifying patients with a specific condition, either for phenotyping or to
identify patient cohorts. The majority originated from the USA (n = 9; 82%), with one study from China and
one from South Korea [Figure 5].

An example of the use of an NLP algorithm to identify patient cohorts and devise a means of following-up
[35]
incidental scan results was by Kooragayala et al. . This study used a keyword search associated with
suspicious pancreatic lesions in over 18,000 patients who underwent a CT scan following trauma over a 10-
year period. The approach identified pancreatic lesions in the reports of 232 patients, of which 48 were
intraductal papillary mucinous neoplasms (IPMNs). In addition, this paper proposed a management
flowchart for incidentally found pancreatic lesions. A further example of the use of NLP in high-volume
[62]
data was demonstrated by Morris-Stiff et al. , who used NLP to identify asymptomatic gallstones from a
cohort of 49,414 patients. They were then able to identify risk factors for progression to symptomatic
gallstone disease in this asymptomatic cohort and showed an approximately 2% risk of symptomatic
progression per year.

DISCUSSION
This review has identified a rapid increase in the quantity of AI research conducted within HPB surgery.
Much of this is focused on intraoperative applications of AI, such as the use of image analysis and computer
vision to address diagnostic and prognostic uncertainties. In addition, the use of 3D reconstruction and
augmented reality models coupled with data-driven prediction algorithms has emerged as an important
area, particularly in preoperative planning and intraoperative decision-making in liver surgery. Artificial
intelligence methods have the most to offer in the distillation of multi-dimensional information to tractable
knowledge that can be applied to individual treatment decisions. HPB surgery represents a good target for
these technologies, given the frequently complex disease patterns and diverse treatment pathways employed.

Most artificial intelligence approaches rely on large volumes of data for training purposes. Of the commonly
described features of big data, the included studies reflect “volume” and “variety” with fewer utilizing real-
time rapidly changing data (“velocity”). Data sources included large pre-existing databases, collated images
and imaging reports. Two notable databases used were the Cancer Genome Atlas and the American College
of Surgeons National Surgical Quality Improvement Program (NSQIP) database, which were widely used
across a range of studies. Natural language processing was frequently employed to extract information from
imaging reports and other healthcare text sources. In one study, NLP was used to identify concerning
pancreatic lesions in historical imaging reports . This demonstrates the depth and flexibility in AI
[35]
techniques to adapt to changes in patient management over time - the malignant potential of particular
pancreatic cysts has only been appreciated in recent years. Moreover, these approaches may be adapted to
help non-specialists managing HPB conditions, particularly in low-resource settings with limited access to
tertiary HPB services. As computer vision approaches improve, the supplementation of local imaging and
pathology reporting with AI-derived diagnostic support may leapfrog the requirement for massive and
often unaffordable training of humans to perform these tasks.

There are, however, genuine risks of bias arising with the development of these techniques. We found
significant geographical variation in current research, with no studies incorporating data from low- and
middle-income countries (LMICs). If the benefits of AI are to be shared equitably across contexts, then
investigators must consider how solutions can broadly generalize between populations and avoid