Page 15 - Read Online

Collins et al. Art Int Surg. 2025;5:254-69 https://dx.doi.org/10.20517/ais.2024.92 Page 260

The use of preoperative imaging and digital twins extends toward other applications in AWS, including
[39]
surgical approach complexity prediction and recommendation, as well as prediction of postoperative
complications such as surgical site infection (SSI) and pulmonary failure [39,40] . The “surgical complexity”
[39]
prediction performance in was encouraging and reportedly better than expert surgeon judgments, with an
accuracy of 81.3% compared with 65.0% (P < 0.001). However, the validation of this study has several
limitations that impact the generalizability and reliability of its findings. One key concern is the lack of
external validation, as the models were only tested within the same institution where they were developed.
Without applying the deep learning models (DLMs) to independent datasets from different hospitals or
more diverse patient populations, it is unclear whether they would perform as well in broader clinical
settings. Additionally, the patient cohort was mostly white (87.5%), which limits the applicability of the
models to more diverse populations. The small validation set for the surgical complexity model (35 patients)
further raises concerns about the robustness of the reported accuracy and predictive power. While internal
validation is useful, a larger and more heterogeneous dataset would provide stronger evidence of the
model’s reliability. Another notable limitation is the lack of expert comparison for the SSI and pulmonary
complication models. Unlike the surgical complexity model, which was evaluated against expert surgeon
predictions, these other models were not benchmarked against clinical judgment, making it difficult to
assess their real-world utility. Furthermore, the pulmonary complication model performed poorly, with the
area under the receiver operator curve (ROC-AUC) of only 0.545, suggesting that the features extracted by
the model may not be clinically meaningful. This raises questions about the adequacy of the training process
and feature selection.

Outcome prediction in healthcare, particularly in complex cases like mesh infection, presents a significant
challenge due to the multifaceted nature of the factors involved. One major difficulty is the need for large-
scale, longitudinal datasets, as outcome data such as mesh infection may not manifest until months or even
[41]
years after the initial procedure . These delayed outcomes make it crucial to gather and analyze data over
extended periods, often involving patients’ health trajectories, lifestyle changes, and treatment responses
over time.

Moreover, effectively predicting these outcomes requires the integration of diverse data sources. For
instance, preoperative imaging can provide insight into the patient’s anatomy and potential risks associated
with surgery, while EHRs offer a comprehensive history of the patient’s medical conditions, previous
treatments, and comorbidities. Blood biomarkers, genetic information, and physiological data from
wearable devices can reveal critical details about the patient’s internal processes and response to treatment,
contributing to a more personalized and accurate prediction model.

In addition to these external data sources, intraoperative factors - such as the surgeon’s skill, experience, and
technique - play a vital role in the success or failure of a procedure, further complicating the prediction
process. The combination of all these data sources can provide a holistic view of the patient’s condition, but
it also introduces challenges in data integration, quality control, and analysis. More research into advanced
AI and ML models, which can handle such complex, multidimensional datasets, is essential for improving
the accuracy of outcome predictions, ultimately helping healthcare professionals make better-informed
decisions.

Intraoperative assistance
AI may contribute to intraoperative assistance in AWS in several ways: