Page 34 - Read Online

Boutros et al. Art Int Surg 2022;2:213-23 https://dx.doi.org/10.20517/ais.2022.32 Page 215

specific groups, such as patients most likely to experience a complication.

Semi-supervised learning combines elements of both supervised and unsupervised learning. A small
amount of labeled data can be combined with unlabeled data to achieve weakly supervised analysis of
surgical data. For example, one application of semi-supervised learning is to use a small amount of labeled
data to train an algorithm and to augment algorithmic performance with the unlabeled data. For example,
unlabeled cholecystectomy videos have been used to improve the performance of an algorithm that has
been trained on a small number of labeled videos of cholecystectomy .
[5]

APPLICATIONS OF ARTIFICIAL INTELLIGENCE IN HEPATOPANCREATICOBILIARY
SURGERY
While AI and machine learning have great success in other fields of medicine, their use in surgery has just
started to increase. More manuscripts that incorporate AI and ML principles to advance the practice of
surgery, and specifically HPB surgery, are being published. Bektas et al. have recently reviewed applications
of AI in HPB surgery . Three important areas of application to HPB surgery that clinicians should
[6]
familiarize themselves with include machine learning for tabular data, natural language processing, and
computer vision [Figure 1].

Machine learning for tabular data
Tabular data simply refers to data that can be organized into rows and columns and is perhaps the type of
data with which clinicians are most familiar. Tabular data is utilized in data collected from claims records,
electronic medical records, or large national datasets. The performance of machine learning on such data is
variable and the literature has been quite heterogeneous regarding the state-of-the-art, with manuscripts
suggesting either very poor or impossibly good performance of predictive algorithms .
[7]

Common sources of tabular data include the electronic medical record of a given institution, claims
databases such as the Nationwide Inpatient Sample, or specific registries such as the National Cancer
Database [Table 1]. Specific registries can take into account clinical variables of interest for given
procedures. For example, Merath et al. utilized the American College of Surgeons (ACS) National Surgical
Quality Improvement Program to identify patients undergoing liver, pancreatic and colorectal surgery from
2014 to 2016 . Decision tree models were utilized to predict the occurrence of a broad range of
[8]
complications, including stroke, wound dehiscence, cardiac arrest, and progressive renal failure, with
accuracy outperforming known risk stratification tools like the ASA and ACS surgical risk calculator .
[8]
Machine learning has been used to create algorithms to predict specific issues such as postoperative diabetes
mellitus after partial pancreatectomy. All patients undergoing partial pancreatectomy in a single institution
were analyzed from 2015 to 2019. Machine learning was utilized to predict the development of
postoperative diabetes mellitus after one year. Their algorithm was able to accurately predict the
[9]
development of diabetes with 87% and 85% sensitivity and specificity, respectively .

Natural language processing
Natural language processing (NLP) is a field that focuses on using automated methods to organize and
make sense of language - both spoken and written. Clinically, NLP algorithms are especially useful for
interpreting and structuring text from electronic health records (EHR). Historically, vocabulary- and rule-
based NLP approaches have relied on lists of words or phrases with multiple variations of a phrase or word.
This, however, becomes untenable as clinical scenarios become more complex. Furthermore, such
approaches require more predictable formats of free text (e.g., structured notes) to optimize performance.
For example, Al-Haddad et al. utilized natural language processing to create a registry of patients with