**6. Postoperative risk assessment**

The use of ML and AI in postoperative risk assessment would work similar to periand intraoperative risk assessment using patient vital signs and characteristics. After performing a surgery, the surgeon must be able to triage patients by likelihood of postoperative complications. Improperly triaged high-risk patients may be sent to hospital floors where there is a high patient-to-clinician ratio, which can limit the frequency of patient assessments and lead to higher rates of morbidity and mortality [92]. Loftus et al. were able to develop an AI model capable of using pre- and perioperative labs and vital signs, intraoperative anesthesia variables (such as intraoperative high inspired oxygen fraction (FIo2)), and postoperative evaluations (including scheduled postop location) to identify undertriaged patients at risk of postoperative complication [92]. In the future, similar technology could be integrated into the electronic health record and send mobile alerts to physicians, allowing for quicker alterations to patient care [93]. Because postoperative risk assessments may utilize more complete information, they have been shown to provide a more accurate prediction of postsurgical prognoses and complications [94, 95].

Machine learning models for postoperative care will also be better suited for predicting pain management needs of the surgical patient. Opiates are common medications prescribed for postoperative pain. However, the opioid epidemic affects over 3 million people in the USA, and it is estimated that 500,000 people in the USA are dependent on opiates [96]. Physicians are now much more aware of the risks of opioid addiction; therefore, opioid dependence and abuse are important considerations to make when prescribing opioids for postoperative pain. A few studies have investigated the use of ML to predict long-term opioid use. One study developed a model to predict long-term opioid use, defined as opioid prescriptions that were requested in addition to the original prescription, in patients who underwent elective hip arthroplasty. Internal validation indicated that the model had good predictive value for the testing cohorts in the study [97]. Other studies have looked at the use of similar algorithms in breast cancer surgery, anterior cruciate ligament (ACL) reconstruction, and joint arthroplasty [98–100]. While these studies did not utilize external validation, these proof-of-concept studies indicate that ML in the future may have utility in predicting long-term opioid use, allowing for more informed prescription of pain medications and potentially earlier identification of patients at risk for opioid dependence.

Machine learning algorithms may also be used for gait analysis in postoperative care. For most elective joint surgeries, postoperative assessment involves patientreported outcome measures or performance-based metrics like the range of motion and mobility [101]. These assessment methods may introduce bias through subjective ratings of outcomes measures by the patient or through biased ratings of performance metrics by physicians [101]. Gait analysis using ML may be able to provide ancillary objective analysis of postsurgical outcomes. One study showed that an ML model incorporating walking speed, gait cycle, maximum force of a step, and other biomechanical variables was able to separate patients who had total knee arthroplasty with patients who underwent unicompartmental knee arthroplasty [102]. Other studies have shown similar potential in total knee arthroplasty and ACL reconstruction [103, 104]. Furthermore, computer vision can likely be leveraged to increase the power of these models. Currently, there exist programs that allow users to mark parts of the body in videos, such as the knees and elbows, and follow the motion of these structures throughout the video. However, manual input of data is time-consuming and prone to human error. To alleviate these concerns, multiple markerless models have been developed to map out patient gait, tracking the movement of anatomical structures such as the ankles, knees, hips, shoulders, head, and arms that do not require human input [105–107]. Based on gait estimation from video, future ML algorithms may be able to stratify patients based on how well they will regain function following surgery. Algorithms may also be able to identify which patients might experience recurring issues or may be at higher risk of falls based on their gait (**Table 4**) [108].


