Artificial intelligence use in abdominal wall reconstruction: a systematic review

Predicting hernia occurrence

Two studies have evaluated the use of AI to predict incisional hernia. Talwar et al. aimed to establish a predictive framework using advanced image analysis and machine learning by analyzing CT scans from 212 adult patients who underwent elective colorectal surgery between 2005 and 2016^[17]. The images were analyzed for 279 features, including abdominopelvic volume, body wall skeletal musculature volume, and pelvic volume pressure ratio. The authors identified optimal predictive biomarkers for incisional hernia: volumes and ratios of visceral adipose tissue (VAT), inner and outer aspects of the body wall skeletal musculature. Machine learning models were then trained on these CT-based biomarkers to predict hernia occurrence, using ensemble boosting, random forest, and support vector machine (SVM) algorithms.

Ensemble boosting combines multiple algorithms to enhance predictive performance, often outperforming individual models^[18]. It involves stacking decision trees or neural networks to improve accuracy. The random forest method uses multiple decision trees run in parallel, each built on random data subsets, with final predictions derived from aggregating individual tree outputs^[19]. SVM algorithms are widely used for handling both linear and non-linear classification tasks^[20]. Among these, ensemble boosting was the best-performing classifier.

A second study also assessed CT imaging to predict incisional hernia following colorectal surgery^[21]. Using an SVM algorithm, this study evaluated the predictive accuracy of various features and found that widening of the rectus complex, increased visceral volume, and body wall skeletal muscle atrophy were all indicative of hernia risk. Both studies highlight increased intra-abdominal pressure - as measured by 3D imaging metrics such as abdominopelvic VAT volume, abdominopelvic inner abdominal musculature (IAM) volume, and the pelvic VAT to outer abdominal musculature (OAM) volume ratio - as key contributors to the pathophysiology of incisional hernia. The authors propose future clinical translation of these findings into a point-of-care tool for preoperative hernia risk assessment.

Preoperative planning

Only one study has evaluated the use of AI in preoperative planning for complex abdominal wall reconstruction. The authors utilized abdominal CT scans taken both at rest and during Valsalva maneuvers to generate models of tissue elasticity, mesh placement, and fixation point distribution. They then applied algorithms based on non-rigid B-spline registration combined with AI to integrate these results for surgical planning. This preoperative planning model was applied to 96 patients, resulting in reduced postoperative pain and no recurrences after one year^[22]. The authors also described their calculation of a gained resistance to impacts related to pressure (GRIP) value, considering mesh type, fixation method, positioning, overlap, native tissue elasticity, and peritoneal closure.

Non-rigid B-spline is a statistical deformation model commonly used in medical image processing^[23]. It allows flexible matching of two images and has been incorporated into radiation oncology for image matching during respiratory motion^[24]. Evidently, non-rigid B-splines have supported clinical practice with excellent outcomes in several hundred cases, demonstrating favorable three-year results.

Elhage et al. used deep learning models (DLMs) to analyze preoperative imaging with the goal of predicting operative and postoperative outcomes^[25]. They assessed the ability of DLMs to predict the need for advanced surgical techniques, such as component separation. When evaluating CT images, the DLMs outperformed expert surgeons in predicting surgical complexity and the necessity for component separation, achieving an accuracy of 81.3%, sensitivity of 88.9%, and specificity of 73.5%. In comparison, the surgeon group demonstrated 65% accuracy, 53.3% sensitivity, and 76.7% specificity. The authors conclude that further refinement of this technology can improve preoperative planning and enhance the informed consent process in surgical care.

Predicting postoperative complications

Licari et al. investigated the use of a kernel support vector machine (KSVM)-based intelligent data analysis technique to identify and validate known risk factors for hernia recurrence^[26]. Their machine learning approach focused on analyzing and mining data to confirm established risk factors such as age, sex, urgency of surgery, surgical technique, ASA class, BMI, comorbidities, mesh type and size, and postoperative infections. Their relapse classification model achieved an accuracy of 86.67%, sensitivity of 86.25%, and specificity of 87.14%.

In addition to predicting surgical complexity using DLMs, Elhage et al. developed models to predict surgical site infection and pulmonary complications. The model for surgical site infection achieved 85.2% accuracy, 78.5% sensitivity, and 87.4% specificity. However, their model for pulmonary complications was less effective, with 77.5% accuracy, 26.9% sensitivity, and 84.0% specificity. Notably, they used conventional 8-layer convolutional neural networks^[25].

Ayuso et al. also assessed hernia recurrence using CT imaging-based DLMs^[27]. They compared a conventional DLM trained on both positive and negative outcomes with models based on a generative adversarial network anomaly (GANomaly) framework, which is trained only on negative outcomes and detects positive cases as anomalies. Their findings showed that the conventional method had higher specificity, while the GANomaly framework demonstrated greater sensitivity for detecting mesh infection or pulmonary failure. Specifically, conventional DLM predicted mesh failure with 96% specificity but only 25% sensitivity, compared to 78% specificity and 68% sensitivity for the GANomaly framework. Similarly, for pulmonary failure, conventional DLM achieved 92% specificity vs. 67% for GANomaly, but only 27% sensitivity vs. 73%. The authors suggested that because GANomaly is trained only on negative cases (no mesh infection or no pulmonary failure), it is less affected by the limited number of positive outcomes. However, they concluded that the conventional DLM framework should not be discarded, given its superior specificity, but rather used alongside the GANomaly method.

While AI has proven useful in imaging evaluation and prediction, several studies have also incorporated clinical data into predictive models^[28-30]. Several risk assessment tools already exist, such as the Ventral Hernia Risk Score, Hernia Wound Risk Assessment Tool, and NSQIP Surgical Risk Calculator^[31-33]. However, these rely on linear regression models that assume risk factors behave in a linear and additive manner^[34]. O’Brien et al. applied a machine learning model to predict surgical site infection using clinical data extracted from medical records^[28], achieving accurate discrimination between infection and non-infection. Hassan et al. demonstrated that machine learning algorithms based on preoperative clinical data - such as patient comorbidities and medical history - effectively predicted recurrence, surgical site infections, and readmission^[29]. Wilson et al. evaluated whether adding clinical data to DLMs improved the prediction of hernia recurrence^[30]. Their results showed that models trained solely on clinical data (including age, sex, BMI, diabetes status, and tobacco use history) outperformed those trained on images only or on images combined with clinical data. This finding is notable, given that prior studies predicting initial hernia occurrence primarily relied on imaging data alone^[17,21].

Chat functions and patient education

The utility of AI chatbots for patient counseling and education has also been a topic of investigation. Li et al. evaluated ChatGPT’s ability to answer common patient questions regarding abdominoplasty^[35]. They provided the chatbot with a list of questions from the American Society of Plastic Surgeons related to abdominoplasty. Although the chatbot could answer frequently asked questions, some responses were outdated and lacked personalized recommendations. Lim et al. extended this work by comparing four popular chatbots - OpenAI’s ChatGPT, Anthropic’s Claude, Google Gemini, and Bing CoPilot^[36]. They found that while all platforms delivered instant and easily accessible responses, there was variability in language quality and supplementary information; similarly, none could offer personalized advice. Lima et al. also assessed the appropriateness of ChatGPT’s responses to questions about ventral hernia repair^[37]. While chatbots may serve as valuable initial tools for patient education, they should not replace surgical consultations.

Alternative uses and future directions of AI

The application of AI has been rapidly expanding across medical specialties, with the most frequent use in diagnostic fields^[38]. Machine learning, AI, and large language models have been used to tailor cardiac treatments and assist in clinical decision making. DLMs have been applied in precision medicine to predict the risk of coronary artery disease by analyzing risk factors such as smoking, diabetes, hypertriglyceridemia, and correlating these with electrocardiogram (EKG) or echocardiography findings^[9]. Additionally, integrating histopathology, genetic risk factors, and imaging modalities - such as cardiac MRI, echocardiography, and coronary angiography - has improved prediction of hospital admissions due to cardiac complications, including acute decompensated heart failure or stroke^[8].

Radiology remains the specialty with the highest volume of AI-related publications^[38]. AI has proven useful in enhancing radiographic diagnoses of conditions such as appendicitis, fractures, and pneumothorax, thereby facilitating faster emergency care. Studies report high accuracy in detecting stroke, intracranial hemorrhage, and suspicious breast lesions^[6]. Beyond diagnosis, AI has been creatively applied to workflow management by tracking imaging processing times and alerting providers to delays, thus expediting patient care^[7].

Minimally invasive surgical techniques hold promise for generating abundant new data. Most laparoscopic and robot-assisted surgical platforms already feature recording capabilities, and AI has shown effectiveness in image processing^{[6,7,17,22,25,27,39]}. With the ongoing digitization of medicine, intraoperative DLM-based assistance tools may soon become common. Such models could also shape the future of abdominal wall reconstruction surgery by integrating patient data - including age, BMI, tobacco use, occupation, and comorbidities - to predict hernia recurrence risk and identify potential postoperative complications.

Machine learning can further assist by analyzing patient imaging to assess muscle weakness, hernias, fat distribution, and other anatomical factors influencing surgical outcomes. AI systems may enable augmented reality via 3D simulations, helping surgeons prepare for procedures and providing real-time anatomical feedback tailored to the individual patient. Moreover, personalized treatment plans informed by AI can enhance postoperative care quality following hernia surgery by offering customized recommendations for nutrition, physical therapy, and recovery.

Limitations

This review included only publications written in English. Although efforts were made to minimize bias, this restriction may have led to the exclusion of relevant information, introducing some inherent bias. The use of AI in surgical specialties is an evolving field, as reflected by the recent publication dates of most included studies - all published within the last 5 years.

In conclusion, AI has significant potential to enhance both the efficacy and efficiency of medical practice. While machine learning models offer numerous possibilities to support clinical decision making, there is currently no consensus on optimal methodologies or applications. The diversity of models and variability in frameworks used across the included studies highlight this lack of standardization. Presently, AI applications in abdominal wall reconstruction primarily focus on predicting postoperative outcomes; however, there is considerable potential to expand its utility across all stages of reconstructive care.