Page 20 - Read Online

Page 265 Collins et al. Art Int Surg. 2025;5:254-69 https://dx.doi.org/10.20517/ais.2024.92

only if the risks associated with uncertain predictions have been established in clinical trials and the benefits
to the patient outweigh the risks. Understanding, communicating, and managing the risks associated with
prediction uncertainty will be key to success.

AI model prediction uncertainty arises from two main factors: epistemic uncertainty and aleatory
uncertainty. It is crucial for all stakeholders and users of AI, including surgeons making critical decisions
based on AI results, to understand these concepts.

Epistemic Uncertainty: This type of uncertainty comes from a lack of knowledge or information. It is
considered reducible, meaning that adding more training data can help the model better capture underlying
patterns or relationships, leading to a more accurate representation of the system.

Aleatory Uncertainty: In contrast, aleatory uncertainty is caused by the inherent randomness and variability
present in the data. This type of uncertainty is irreducible - no matter how many additional data are
collected, the inherent variability in the data remains.

Considering epistemic uncertainty, if an AI system is trained on a dataset that lacks diversity in terms of
patient demographics or surgical techniques, or patient abnormalities due to prior surgeries, surgery
technique recommendations may be untrustworthy due to high epistemic uncertainty. Ensuring that AI
models are trained on diverse and representative datasets reduces epistemic uncertainty, and the only way
to understand if the uncertainty is sufficiently reduced is through rigorous model testing with appropriate
[67]
external validation .

However, reducing epistemic uncertainty alone is not sufficient to guarantee trustworthy results. For
instance, a digital twin’s accuracy is limited by the imaging modality from which it was generated. Various
anatomical structures may not have sufficient contrast for unambiguous delineation, especially in non-
contrast-enhanced CT. Consequently, a digital twin may be imperfect, no matter how many data are used
for training. This is an example of aleatory uncertainty. By definition, this is irreducible, and whether it is a
problem depends, crucially, on the intended use of the model’s predictions. Before investing considerable
research effort in developing a model and gathering training data, this should be fully understood.

Pragmatically, one should always estimate the aleatory uncertainty before investing time in AI model
development. One of the best ways to do this is to have multiple expert surgeons perform the task
independently and then measure their inter-rater agreement. A high disagreement indicates high aleatory
uncertainty, which may prohibit the possibility of a useful AI model ever being developed. We performed
this kind of analysis when we were researching our system to automatically verify the completion of the
[16]
critical view of the myopectineal orifice (CVMPO) , which helped steer the research toward a more
focused goal of detecting critical structures in the CVMPO. Initially, we aimed to automatically verify the
[68]
completion of every step according to CVMPO guidelines . However, our preliminary inter-rater analysis
revealed that several steps, such as assuring dissection clearance of 2 cm from Cooper’s ligament, had a low
agreement among surgeons. This was due to the high difficulty in visually determining the absolute size of
structures in a scene from single-lens (monocular) laparoscopic images. Instead, we focused on detecting
critical structures, including Cooper’s ligament, which was a substantially more objective task (with less
aleatory uncertainty). Aleatory uncertainty can sometimes be reduced by altering the nature of the data - for
example, by using contrast-enhanced CT scans to improve image clarity and image segmentation
performance. However, in other situations, reducing this uncertainty may not be feasible without disrupting
the surgeon’s workflow or making the process impractical. For instance, we have shown that AI models can