Page 13 - Read Online

Collins et al. Art Int Surg. 2025;5:254-69 https://dx.doi.org/10.20517/ais.2024.92 Page 258

learning from their mistakes. During supervised learning, the initial “untrained” AI model makes
predictions on the training data, and a performance metric is used to evaluate how well it performs, such as
sensitivity, specificity, precision, recall, accuracy, or F1 (the harmonic mean of precision and recall). Based
on this evaluation, an automatic training process is launched, which makes small modifications to the
model’s parameters to improve its performance metric. The process repeats iteratively, and the model
gradually improves its ability to make accurate predictions on the training data.

Alternative ML methods exist, and they can be used alongside supervised learning to enhance model
[29]
training. One such approach is self-supervised learning , where the model generates its own labels from
the input data, allowing the model to identify inherent structures or relationships in the data. Self-
supervised learning is often used as a prelude to supervised learning, where its main benefit is to reduce the
amount of required labeled data. Another notable approach is reinforcement learning (RL), which
complements supervised learning by enabling the model to learn through trial and error. In RL, the AI
model learns strategies, or decision-making policies, designed to maximize a virtual reward. A virtual
reward refers to a feedback signal that the model uses to gauge the effectiveness of its actions. Unlike in
supervised learning, where the model is provided with labeled training data, in RL, the model explores
actions and receives rewards (or penalties) based on the success or failure of those actions. This virtual
reward can be a numerical value that increases when the model takes actions that bring it closer to achieving
its goal, or decreases when it moves further away from the desired outcome. Over time, the model learns to
maximize these rewards by refining its action/decision-making policy.

To draw an analogy, RL is similar to training in a surgical simulation. Initially, the surgeon might perform
steps in a procedure without knowing whether the action is optimal or whether the outcome will be
successful. However, the simulator provides feedback that may be positive (if the surgeon performs a task
correctly, reinforcing the trainee’s good decisions/actions) or negative (e.g., if an action leads to a
complication), prompting the surgeon to adjust their approach on their next trial. As the surgeon continues
to practice, they refine their skills based on this continuous cycle of action, feedback, and adjustment,
ultimately leading to mastery. Similarly, in RL, the model continuously learns from its experiences by
receiving virtual rewards, allowing it to optimize its performance over time. The main advantage of RL over
supervised learning is that RL allows the model to learn optimal decision-making strategies through trial
and error, without requiring explicit labels to tell it which actions or decisions should be performed at each
step.

[28]
The second phase of AI model development is internal validation , where the performance of a trained AI
model is tested using data similar to the training data, typically collected concurrently with the training
dataset. This helps catch common problems early, such as overfitting or model bias. Overfitting happens
when the model learns random details or noise in the training data instead of general patterns, causing it to
perform well on the training data but poorly on new, unseen data. Model bias, on the other hand, occurs
when the model makes incorrect assumptions about the data, leading to inaccurate predictions.

The third phase is external validation, where the model is tested on unseen, independent data. This step is
crucial for assessing the model’s ability to generalize to new, real-world scenarios and ensuring that it can
perform reliably outside the training environment. The fourth phase involves the regulatory approval
process, where models must meet strict standards and regulations regarding performance, intended
population, and safety before they can be deployed for real-world use . Finally, the fifth phase is
[30]
performance monitoring, which focuses on tracking the model’s performance over time. This phase is
required to detect “drift”, or any decline in performance caused by changes in the data or the environment