Page 64 - Read Online

Page 250 Elhage et al. Art Int Surg. 2025;5:247-53 https://dx.doi.org/10.20517/ais.2024.87

complexity and surgical site infection models performed very well, with AUCs of 0.744 and 0.898,
respectively. Additionally, the surgical complexity model outperformed a panel of expert surgeons on a
separate validation set, with an accuracy of 81.3% compared to the expert surgeon accuracy of 65.0%.
However, the model for predicting pulmonary failure was unsuccessful, with an AUC of 0.545, and this was
thought to be due to the low sample size of patients who developed postoperative pulmonary failure .
[25]
A follow-up study by Ayuso et al. in 2023 aimed to utilize specialized techniques to improve the accuracy of
DL models for predicting rare but often devastating outcomes in AWR. Querying a database of 510 patients,
they utilized framework-augmented DL models that incorporated image augmentation and anomaly
detection, as a way to give an algorithm more data for learning when original unique data may be limited.
Despite the variables of interest occurring at low rates, pulmonary failure at 5.6% and mesh infection at
3.7%, they were able to successfully predict these outcomes with the augmented DL models. In comparison,
[26]
these models significantly outperformed conventional DL models assessing the same outcomes .
In 2022, McAuliffe et al. utilized ML to successfully predict the development of incisional hernia after
colorectal operations. Rather than AI analyzing images on its own, the authors identified features on
preoperative CT images, 21 of which were capable of discriminating between incisional hernia and non-
incisional hernia patients. With these features, they utilized ML to generate predictive algorithms based on
three unique pathophysiologic domains: structural widening of the rectus complex, increased visceral
volume, and atrophy of the abdominopelvic skeletal muscle. Based on the importance of these domains, the
authors hypothesized a mechanism for hernia formation based on the premise of early fascial microgap and
failed healing, in the setting of decreased fascial integrity and higher forces on the fascia via increased intra-
abdominal pressure. Work such as this can not only help with the prediction of hernia development, but
[27]
can also improve our understanding of the underlying pathophysiology . Furthermore, this study
highlights important traits of supervised ML algorithms, where the significance of specific features can be
[11]
uncovered, as opposed to DL models where feature significance is often hidden in a black box .
Talwar et al. similarly aimed to predict the risk of hernia development after abdominal operations, utilizing
optimal biomarkers identified on preoperative CT imaging to train their predictive ML model. Optimal
biomarker methodology can utilize unstructured data to “identify a small subset from a large set of features
that are both discriminative for the outcome of interest and independent from other features”. Using this
technique, their most successful ML model was able to predict incisional hernia formation well, with an
AUC of 0.85, accuracy of 0.83, sensitivity of 0.86, and specificity of 0.81 .
[28]

An interesting study by Wilson et al. published in 2024 was the first in the hernia literature to compare
three DL predictive models using images alone, clinical data alone (age, sex, body mass index, diabetes
status, and history of tobacco use), and a combined model using images and clinical data. All three models
were found to be poorly predictive of hernia recurrence; however, a noteworthy conclusion from this study
was that the DL model based only on clinical data outperformed the image-based model and the combined
model. This observation suggests that certain outcomes may benefit more from image analysis or clinical
data, and that the combination of the two does not necessarily purport an additive effect in terms of
[29]
accuracy .
GENERATIVE AI
Generative AI is a fascinating field within AI that focuses on AI’s ability to create new content. Much of the
AI technology that has infiltrated our day-to-day lives, such as Chatbots and music-, image-, and video-
generating software, are examples of generative AI. There has been much speculation as to the future