AI in plastic surgery: customizing care for each patient

AI increasing patient communication

In PRS, patient-surgeon communication is essential to address expectations of procedures, approach patients’ concerns and goals, and manage patients’ health. Communication between clinicians and patients has a demonstrable impact on patient satisfaction, clinical outcomes, and litigation^[16]. However, there is often a discrepancy between surgeons’ level of communication and patient’s level of comprehension. One of the direct applications of AI in transforming medical care is optimizing the creation and delivery of patient information, as well as medical documentation. AI tools, including ChatGPT and artificial intelligence virtual assistants (AIVAs), have recently been utilized to create relevant and tailored medical information to support patients' needs. These NLP systems have proven their ability to increase the readability of medical information, respond to frequently asked surgical questions, and address the benefits and risks of PRS operations^[17-24]. By producing patient-relevant information, AI tools can also reduce the need for healthcare workers to respond to frequently asked questions during medical appointments. Additionally, ChatGPT can be utilized 24/7 to address patients’ concerns and promptly provide essential medical information^[25]. As a result, chatbots may help decrease the need for in-person medical consultations and increase access to healthcare in rural regions at a reduced cost^[26,27].

Breast reconstruction

Patient satisfaction is a central goal of breast reconstruction, and delivering patient-centered treatment during the reconstructive process can help improve the perception of quality of care^[16]. ML methods have accordingly enabled more patient-specific care in PRS and other disciplines of surgery [Figure 2]. For example, ML has been used to assist clinicians in reconstructive method selection, preoperative planning, facilitation of postoperative monitoring, enhancement of patient outcomes, and to decrease hospital readmissions [Table 1]^[33]. In a pilot study by Mavioso et al., the preoperative utility of ML was evaluated for semi-automatic assessment of Angio CT imaging for forty patients scheduled for deep inferior epigastric perforators (DIEP) breast reconstruction^[34] [Table 2]. Specifically, Mavioso et al. utilized a paired sample t-test and Wilcoxon test to compare the blood vessel sizes determined using semi-automatic identification against manual identification^[34]. Additionally, a one sample t-test was performed to evaluate the estimated location of the blood vessels when utilizing semi-automatic identification^[34]. When compared to the manual procedure performed by the imaging team, ML analysis of vessel caliber, orientation, and location significantly reduced the time spent on preoperative planning for DIEP flap reconstruction. However, the software could not accurately estimate the caliber of small vessels (< 1.5 mm)^[34]. Additionally, the vertical component of vessel location differed by 2-3 mm from the manual method, although this discrepancy did not impact the dissection. Overall, this study demonstrates that ML may decrease the time spent on surgical planning and simplify the overall process.

Figure 2. AI-supported patient-specific processes in plastic surgery. AI: Artificial intelligence.

ML algorithms can also be valuable for the prompt detection of complications following breast surgery. By analyzing available patient data, these algorithms can identify patterns and determine the associations among relevant variables^[1]. Kiranantawat et al. developed the first smartphone application for microsurgery monitoring by training the algorithm with photographic data of fingers undergoing venous or arterial congestion^[35]. Across forty-two participants, the application successfully assessed the vascular status of fingers with a sensitivity and specificity of 94% and 98%, respectively^[35]. This study suggests that ML could enhance early detection of postoperative flap failure and help optimize monitoring of the flap after surgery. Another study by Myung et al. developed an ML model to determine patient-specific characteristics and surgical factors that lead to an increased risk of donor site complications after the performance of abdominal flaps for breast reconstruction^[36]. After analyzing 568 patients, Myung et al. discovered that the algorithm was able to accurately predict complications [area under the curve (AUC): 0.89] and could further be used as a reference for assessing the individual risk associated with abdominal flaps^[36].

Additionally, ML has been applied to minimize postoperative infection following implant-based reconstruction, including the development of internal algorithms to guide clinical decisions such as the need for reoperation or introduction of antibiotics^[1]. Hassan et al. developed an algorithm using ML to predict periprosthetic infection and explantation^[37]. The study demonstrated that ML models can provide a higher predictive accuracy compared to multivariate logistic regression for periprosthetic infection and explantation^[37,38]. Therefore, ML could help reduce postoperative burden and promote better outcomes in breast reconstruction.

Facial surgery

In the context of facial surgery, photographic data provides a means to assess surgical success or failure, and it can also be used as a tool for enhancing communication between patient and physician. ML has been explored in combination with photographic data to maintain proper standardization of procedures and offer more precise postoperative assessment^[39]. Using pre- and postoperative pictures, Zhang et al. showed that neural networks could identify preoperative age and facial age reduction following facelift surgery^[39]. A positive correlation between the algorithmically determined result and patient satisfaction after facelift was identified, representing a validated method of quantifying postoperative results and efficacy for plastic surgeons^[39]. In another study, Boonipat et al. used ML to assess postoperative facial expression improvement after facial reanimation surgery^[40]. Recording of facial expressions was performed for each patient in a video clip and analyzed with ML software to detect facial expressions^[40]. ML algorithms were found to be capable of reading facial emotional expressions and providing a quantification of those expressions. These tools may thus be helpful in assessing facial palsy and the success of postoperative outcomes^[40]. Moreover, corrective procedures, the use of neurotoxins, or soft tissue fillers could utilize ML as an assessment tool for photographic or recorded data^[39,40].

ML has also been applied to perform automatic detection of craniofacial conditions, facilitating early diagnosis based on photographic images and annotated datasets. In an early study by Geisler et al., neural networks have successfully achieved an overall testing accuracy of 90.6% for the detection of craniosynostosis, opening the field for earlier diagnosis and minimizing the need for CT scans^[41]. Another study by Knoops et al. described a computer-assisted model framework involving supervised learning for diagnostic, predictive outcome and treatment stimulation in craniofacial surgery^[42]. The algorithm was trained on non-ionizing 3D face scans of healthy faces and orthognathic patients, and it provided an accurate classification with a 95.5% sensitivity and 95.2% specificity^[42]. The algorithm was also able to stimulate patient-specific postoperative outcomes with a mean accuracy of 1.1 +/- 0.3 mm. compared to conventional surgical planning, suggesting that the model could predict the postoperative shape of the face in a single step and reduce the time for the planning process^[42].

In addition to data classification, the advent of generative AI tools such as DALL-E2 can enable the creation of various types of synthetic images or text on demand^[43]. Cosmetic surgery, given its inherently visual nature, can, therefore, take advantage of generative AI to simulate post-surgery results even prior to the procedure. In one study by Lim et al., DALL-E2, Midjourney and Blue Willow were evaluated in their utility to provide images clinically relevant after cosmetic surgery^[44]. In future cases, surgeons could virtually simulate different interventions and examine the AI outcomes with patients for preoperative scoring and evaluation, which may help surgeons fine-tune the planned procedure and aim for specific modeled outcomes.

Hand surgery

Similar to radiology and other imaging-dominated disciplines, AI has been applied in hand surgery for fracture detection^[45]. Despite the performance of standard clinical examination and X-ray characterization, scaphoid fractures, representing 15% of acute wrist fractures, are missed initially in nearly 16% of cases^[45]. Therefore, there is an existing need for ML algorithms to improve the detection of scaphoid fractures, wrist fractures, and other cases within emergency departments. In a recent study by Ozkaya et al., an ANN model for scaphoid fracture detection on anteroposterior wrist radiographs was compared to three physicians (two orthopedic specialists and one physician in the emergency department)^[46]. The ANN showed a 76% sensitivity and 92% specificity, which exceeded the performance of emergency department physicians, but still lagged behind that of an experienced orthopedic specialist^[46]. However, the addition of clinical examination findings in the algorithm, as well as lateral views, could enhance the sensitivity of the ML algorithm. Further, Oeding et al. discovered that recent AI models have demonstrated excellent performance in detecting scaphoid fractures and radius fractures, with AUC values of 0.77-0.96 and 0.90-0.99, respectively^[47]. The majority of AI models have currently demonstrated comparable or better performance than many clinical experts, and further improvements in speed and performance may result from larger data sets, more powerful computing resources, and increasingly open-source code toolboxes^[47].

Based on patient-specific data such as age, smoking status, dominant hand, occupation, and subtype of fracture, ML could also provide useful information in the acute setting, such as deciding whether to perform one type of hand surgery over another (e.g., replantation vs. amputation)^[48,49]. Recently, two studies by Hoogendam et al. and Loos et al. from the Hand and Wrist Study Group in the Netherlands have introduced user-friendly graphical applications, supported by ML algorithms, for predicting postoperative function^[48,49]. These ML approaches were applied to data from 2,119 patients with carpal tunnel syndrome (CTS) and 2,653 patients with thumb carpometacarpal osteoarthritis^[48,49]. These applications calculated the probability of functional improvement 6 months after CTS surgery and 12 months after first carpometacarpal joint surgery, based on preoperative patient-reported outcome measures (PROMs). While these ML algorithms are freely available, it is crucial to note that such online applications often lack external validation, which is essential for ML algorithms to be generalized to other patient cohorts.

Wound healing and burn surgery

While postoperative wound assessment and management are essential to ensure optimal treatment, no “gold standard” for scar evaluation currently exists^[50]. Challenges in scar evaluation may arise from the variability in clinical assessments and difficulty in achieving consistent accuracy in follow-up evaluations. A recent study by Kim et al. introduced DNN models to classify postoperative scars based on scar severity. This model was trained with both an image-based AI model and a model based on clinical variables related to postoperative scars, such as patient demographic data, scar age, and symptoms^[50]. Four scar severity groups were successfully classified using these image-based AI models at a performance level comparable to that of board-certified dermatologists, underscoring the efficacy of AI in clinical assessments^[50]. A pilot study by Squiers et al. separately demonstrated the utility of combining image-based analysis with ML risk factor assessment in predicting the healing outcomes of primary amputation wounds^[51]. The level of amputation was determined by the subject’s surgeon prior to imaging, and was based on clinical judgment such as patient history, physical exam, and any perfusion studies^[51]. Multispectral imaging of the subjects’ lower extremity planned for amputation was also conducted on postoperative day 30^[51]. Analysis of multispectral imaging demonstrated greater effectiveness in predicting primary amputation wound healing relative to surgeon judgment, with an 88% accuracy rate compared to 56%^[51]. If further evaluation and/or external validation confirm these findings, this type of ML tool may enhance the decision-making process in wound healing treatments.

Certain complex wounds, such as those from diabetes and burns, are particularly susceptible to complications and delayed healing due to impaired circulation and increased risk of infection. Recently, advancements in burn management have been integrated with AI tools to enhance the treatment of burn wounds^[19]. DL-based analysis models were able to identify the depth of early burn wounds using inputs based on clinical photographs of the wound. Furthermore, in a study by Robb et al., ANNs provided accurate diagnoses of burn injuries based on color attributes and successfully classified burns into standardized categories, achieving a diagnostic accuracy of approximately 80%^[52]. Apart from their beneficial role in burn assessment, AI algorithms have the potential to aid in clinical decision making by accurately predicting clinical outcomes in wound healing, such as the need for skin grafts or amputation^[52]. Diabetic wounds similarly pose significant challenges in the clinic due to their higher complications and slow healing, which may be addressed using AI techniques for modeling and therapeutic discovery^[53]. In one early example, Xue et al. used AI tools to identify a novel therapeutic agent for diabetic wound healing by predicting molecular interactions between the drug Trichostatin A and its receptors at the wound site^[53]. Ultimately, AI tools can support the personalized treatment of wounds and burns by integrating clinical data, patient-specific risk factors, biological modeling, and prediction of potential postoperative complications.

3D and predictive modeling

Given the unique size and geometry of each patient’s surgical site, 3D modeling has been utilized for patient-specific surgical planning^[42]. Similarly, the advent of 3D printing technologies has enabled the design of increasingly site- and patient-specific constructs for surgical implantation^[54]. Digital models of a surgical site can be reconstructed using traditional medical imaging techniques such as CT or magnetic resonance imaging (MRI)^[42]. In addition to supporting the preoperative planning of incisions, positioning, and other factors, these models can utilize AI methods to simulate patient-specific changes in tissue geometry (e.g., facial shape) that may result from a procedure^[42,55]. In one example, Knoops et al. utilized finite element modeling to develop patient-specific predictions of maxillofacial transformation based on preoperative and postoperative CT scans^[55]. The authors also utilized experimental design to infer the contributions of input parameters, including Young’s modulus and viscoelasticity, to soft tissue displacement^[55]. In other examples, surgeon-scientists have utilized generative AI to produce 3D models of facial shapes based on 2D medical images^[42,56-58]. For instance, a 3D morphable model trained on over 4,000 faces has been applied for the diagnosis, risk stratification, and treatment simulation of jaw surgery patients^[42]. Ultimately, ML algorithms have demonstrated the potential to help improve surgical outcomes and reduce medical costs by creating patient-specific predictive models prior to operation^[56].

3D printing for planning and implantation

In addition to 3D modeling, 3D printing may be utilized in combination with AI to generate increasingly patient-specific models and implants. On the printing and technological side, ML can be utilized to identify the optimal printing parameters to generate a desired shape and/or internal architecture^[59]. Hierarchical ML algorithms, for instance, have been used to identify optimal material formulations, process variables, and fiber geometries for the production of silicone implants and other constructs^[60,61]. With the support of AI methods, patient-specific models of the breast, vasculature, craniofacial tissue, and more have accordingly been fabricated by 3D printing for purposes of preoperative planning^[54]. Chae et al., for instance, used CT and MRI scans to visualize and print breast tissue models for mastectomies^[62]. In addition to supporting the planning process, “bedside” 3D printing has been explored in some early instances for the development of patient-specific implants^[54]. In one notable example, Lei et al. utilized ML models to design cochlear implants with optimal electro-anatomical properties given the patient’s specific inner ear geometry^[63]. Ultimately, 3D printing methods can produce more geometrically complex and site-specific constructs compared to traditional fabrication methods, particularly when paired with AI models.