Page 73 - Read Online

McGivern et al. Art Int Surg 2023;3:27-47 https://dx.doi.org/10.20517/ais.2022.39 Page 31

Studies identified were subdivided into groups focusing on diagnostics, prognostics, and interventions. We
assessed 23 papers [13-35] reporting diagnostic uses of AI in HPB surgery. Of these, five focused on the
gallbladder, 11 on the liver, and seven on the pancreas. Twenty-nine studies reported prognostication [36-64]
using AI, of which three focused on the gallbladder, 16 on the liver, one on the liver and pancreas, and nine
on the pancreas alone. Almost half of the studies identified reported on the interventional use of AI [65-110] in
HPB surgery (n = 46), with 24 studies focusing on the liver, 19 on the gallbladder alone or in conjunction
with another organ (n = 4), and three studies looking at the pancreas. A summary of the papers subdivided
into the diagnostic, prognostic and intervention cohorts can be found in Tables 1, 2, and 3, respectively.

Regarding sample size, most studies (n = 13) reporting diagnostic applications of AI in HPB surgery utilized
data from fewer than 1,000 patients. The smallest number of patients in a focused study of the ultrasound-
[14]
based classification of liver lesions was 22 . Three studies included over 5,000 patients and were included
in our big data cohort [16,30,35] . The largest number of included patients was 199,783 . Most studies (n = 16)
[30]
looking at prognostic uses of AI in HPB surgery had fewer than 500 patients. Two studies had fewer than
5,000 patients, but were included in our big data cohort due to the high number of images and image
reports included [58,63] . Eleven studies looking at interventional uses of AI in HPB surgery did not use actual
patient data, but used simulations-based approaches [69,70,75,76,81,82,84,85,91,92,104] . There was little mention or use of
“explainable AI” concepts in any of the included studies.

Conceptual mapping of AI research in HPB surgery
Following data extraction and study classification, we undertook a conceptual mapping exercise to identify
key areas and relationships in AI use [Figure 4]. Many of the identified concepts involved outcome
prediction (such as the risk of complication, or personalized survival predictions). Others utilized AI to
support clinicians in the identification of a condition before, during, or after surgery (such as identifying
malignancy, identifying complications early, or even the prevention of these by using AI to alert clinicians
to unseen structures intraoperatively). Preoperative planning and surgical simulation were particularly key
areas within the intervention grouping. Finally, within the conceptual mapping exercise, we identified
several areas where AI may be useful as either a risk stratification tool or as an intervention in future
research (purple text, Figure 4).

Diagnostic applications of artificial intelligence
Diagnostic applications of AI primarily involved interpreting images using computer vision models [Table 1
and Figure 4]. AI was used across a range of imaging modalities, including transabdominal ultrasound,
endoscopic ultrasound, MRI and CT, to identify lesions or classify lesions into different radiomic subgroups
of disease. Although the majority of preoperative, diagnostic AI work focused on imaging, there were
studies investigating perioperative risk prediction. However, there were no studies that proposed to use AI
as an intervention in preoperative care pathways. Therefore, it should be considered that preoperative AI
may also be undertaken with a broader surgical focus, rather than specifically targeted at HPB populations
and hence are not discussed in this review.

Prognostic applications of artificial intelligence
The majority of prognostic applications for AI were in the prediction of cancer recurrence and survival
[Table 2 and Figure 4]. This was achieved using a variety of input data, including imaging, genetics, and
clinical characteristics. Prediction models were developed for a variety of time points, including the first 30
days following surgery and for longer-term survival.