**3. Perioperative risk assessment and surgical planning**

Due to the ability to quickly and efficiently incorporate and compile large amounts of data, AI/ML paradigms are likely to be heavily involved in preoperative risk assessment in all fields of surgery. Through the collection of patient data and characteristics, such as weight, heart rate, blood pressure, comorbidities, and other factors, highly sophisticated models can be used in algorithms that predict the risk of the patient before undergoing a surgical procedure. With the ability to calculate risk, AI/ ML may also bring the potential for appropriate mitigating strategies that could decrease patient morbidity and mortality [4, 15]. By utilizing large data sets organized by specific surgical procedures and procedure types, AI/ML-powered algorithms could be used to modify models that carry out statistical weight optimization for different variables associated with morbidity and/or mortality for each type of surgery, within a specific set of clinical circumstances (e.g., emergency versus nonemergency) or within a certain population (e.g., demographic). Assuming a representative sample, an effective AI/ML algorithm would allow surgeons and other perioperative medicine experts to input values for individual patients and return an objective preoperative risk assessment, leading to potential applications in precision medicine. For instance, there are multiple different bariatric surgeries available to patients, including sleeve gastrectomy, Roux-en-Y gastric bypass, adjustable gastric band, and biliopancreatic diversion [16]. Though sleeve gastrectomy is now the most common approach, each technique has trade-offs between cost, short-term morbidity, long-term morbidity, and long-term weight loss, and this can sometimes lead to complex decisions in choosing the optimal procedure [17]. Machine learning algorithms could help address this issue using preoperative data to provide individualized recommendations, potentially leading to more optimal bariatric surgery prescriptions [16]. Recent studies have investigated the use of similarly structured and implemented algorithms across many different types of surgeries and surgical challenges, from predicting preoperative risk of cardiac complications, identification of a difficult airway prior to intubation, and the general risk–benefit estimations of different procedural or surgical interventions [18–22]. When properly designed and implemented, such algorithms would allow for risk stratification and, thus, better preparation for adverse outcomes following surgery. Future improvements would increase the specificity and sensitivity of these algorithms, facilitating a more accurate prediction of perioperative risk. Additionally, AI algorithms may be able to provide quantitative predictions about outcomes with and without surgery, providing both surgeons and patients with the information for objective decision-making [23].

Additional preoperative risk assessment could take the form of dedicated ML analysis of the radiologic imaging [24]. Preoperative imaging is utilized before surgery to give surgeons more information about the patient's pathology and anatomy and is essential for preoperative planning. ML algorithms can be used in the preoperative setting to predict prognoses and augment surgical decision-making across various surgical specialties [25–27]. An example of the implementation of preoperative ML models is the utilization of computed tomography (CT) scans to diagnose lung cancer. Using ML to evaluate CT scans has shown comparable to even better sensitivities and specificities compared to radiologists [28]. Such models can be further augmented to provide data about each identified tumor and suggestions for surgical planning [29]. More widespread adoption of ML algorithms that read imaging could lead to advancements in surgical planning in interventions such as lumbar decompression in spinal stenosis to assessing characteristics of corneal endothelium in specular microscopy for treatment of corneal edema [20, 30]. The utilization of ML algorithms could transform how surgeons interpret CT scans preoperatively and could, in return, improve patient care and surgical outcomes.

Advances in the algorithmic interpretation of medical imaging have led to the emergence of radiomics, a field involving the analysis of medical imaging to provide information about the physiology or pathology of the disease [31]. Radiomics contributes an additional layer to how ML algorithms can interpret medical imaging and has shown unique promise in surgical oncology, where minute changes in image features can be associated with various prognoses. Typical features used in radiomic workflow may include the intensity of signals and the distribution of these signals [32]. Because benign and malignant tumors have different microenvironments and expression of specific markers, magnetic resonance imaging (MRI) radiomics shows promise in being able to differentiate malignant or benign tumors from normal tissue [32]. Radiomics could therefore improve patient outcomes through early identification of disease.

In terms of specific examples, radiomics can be used to determine axillary lymph node (ALN) metastases in patients with breast cancer. The most common site of breast cancer metastasis is to the axillary lymph nodes (ALN). Early detection of ALN metastases can inform the surgical management of breast cancer [33]. Based on the Z0011 clinical trial results, the current diagnostic procedure for ALN metastases for most patients is sentinel lymph node biopsy (SLNB) [34]. Although this procedure is less invasive than ALN dissection, SLNB still carries the risk of lymphedema, axillary paresthesia, and reduced range of motion in the involved upper extremity [35]. Furthermore, in some cases, SLNB has been shown to have false negative rates in the range of 5–10% [36]. Thus, finding more effective alternative ways to identify ALN metastases is increasingly important. Radiomics has shown the ability to identify malignant tissues and determine ALN metastases at a higher rate than radiologists [37]. In the future, radiologists equipped with radiomics capabilities may be able to more efficiently and more accurately identify ALN metastases, leading to more prompt medical and surgical therapeutic interventions. Evidence suggests that radiomics may be able to differentiate between different subtypes of cancer based on the unique molecular profile and the resulting appearance on imaging of each subtype [38]. The ability to specifically diagnose different subtypes of cancer from their respective radiologic imaging characteristics may allow surgeons to stratify patient prognoses and better determine medical and surgical management (e.g., precision medicine/surgery).

Preoperative uses of ML and AI could also improve patient outcomes for those who are awaiting organ transplants. More specifically, ML algorithms trained to analyze
