Diagnostic Accuracy of AI in Prediction and Assessment of Compromised Free Flaps: Systematic Review and Meta-Analysis Kuan-Chen Huang, Melanie J Wang, Yu-Ying Chu, Ren-Wen Huang, Yun-Jui Lu, Cheng-Hung Lin, Chung-Chen Hsu, Shih-Heng Chen, Yu-Te Lin, Che-Hsiung Lee, Stefano Brini Abstract AbstractBackgroundVascular compromise remains the leading cause of free-flap failure. AI-based monitoring and prediction tools have emerged as a promising adjunct for postoperative free flap monitoring and early detection of vascular compromise. Previous systematic reviews included limited evidence or broadly evaluated reconstructive outcomes.ObjectiveThis systematic review and meta-analysis assessed the diagnostic accuracy of AI for postoperative free flap monitoring and flap compromise detection.MethodsFollowing PRISMA-DTA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses of Diagnostic Test Accuracy Studies), PubMed, Embase, Cochrane, Web of Science, and Scopus were searched from inception to June 21, 2026. Studies developing or validating AI models for free flap monitoring or compromise prediction were included. Pooled sensitivity, specificity, area under the curve (AUC), and diagnostic odds ratio (DOR) were estimated using a hierarchical bivariate random-effects model. Risk-of-bias and applicability were assessed with Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2), and the certainty of evidence was evaluated using the GRADE (Grading of Recommendations Assessment, Development and Evaluation) approach. Prespecified subgroup analyses by modality, model-fit diagnostics, sensitivity analyses, and publication-bias testing with Deeks funnel plot were undertaken. The protocol was prospectively registered in PROSPERO (CRD420251175572).ResultsOf 2098 records identified, 18 studies met the inclusion criteria. Of these, 17 studies were included in the quantitative synthesis. Pooled analysis demonstrated an overall sensitivity of 0.83 (95% CI 0.70‐0.91; prediction interval [PI] 0.21‐0.99), specificity of 0.87 (95% CI 0.65‐0.96; PI 0.03‐1.00), AUC of 0.92 (95% CI 0.90‐0.94), and DOR 36.17 (95% CI 8.27‐158.26; PI 0.09‐14886.43). Image-based AI models demonstrated superior performance, with a sensitivity of 0.92 (95% CI 0.81‐0.97; PI 0.43‐0.99), specificity of 0.95 (95% CI 0.86‐0.98; PI 0.43‐1.00), and AUC of 0.98 (95% CI 0.96‐0.99). Nonimage-based models had sensitivity (0.69, 95% CI 0.45‐0.86; PI 0.11‐0.98) specificity (0.72, 95% CI 0.18‐0.97; PI 0.00‐1.00), and AUC (0.79, 95% CI 0.75‐0.82). For vascular compromise detection, pooled sensitivity was 0.92 (95% CI 0.81‐0.97; PI 0.44‐0.99) and specificity was 0.91 (95% CI 0.79‐0.97; PI 0.35‐1.00). Between-study heterogeneity was substantial, QUADAS-2 identified methodological concerns in several studies, and GRADE rated the certainty of evidence as low.ConclusionsAI demonstrated favorable diagnostic performance for detecting free flap vascular compromise, particularly with image-based models. These findings support the use of AI as an adjunct to conventional postoperative flap monitoring. Nevertheless, the low certainty of evidence, substantial heterogeneity, and limited external validation indicate that prospective multicenter validation and standardized reporting are required before routine clinical implementation. Introduction Microsurgical free tissue transfer has become a cornerstone of modern reconstructive surgery and is routinely performed in breast, head and neck, extremity, and other complex reconstructions [,]. With reported success rates of 94%‐99%, it offers durable and functional restoration for even the most challenging defects [-]. Despite continued technical refinements and attentive perioperative care, complications remain a persistent concern []. Among these, flap failure is particularly devastating. It is most often caused by vascular compromise, including arterial insufficiency, venous congestion, and ischemia-reperfusion injury, with incidence rates reported in 10%‐30% of cases []. These events frequently necessitate reoperation, with 2%‐5% of flaps requiring surgical exploration, thereby increasing morbidity, prolonging hospitalization, escalating healthcare costs, and imposing substantial psychological stress on patients [-]. Because salvage success is highly dependent on the interval between vascular compromise and reintervention, accurate risk identification, early recognition of compromised flaps, and timely response are critical []. Risk stratification and vigilant postoperative monitoring remain essential components of high-quality microsurgical care []. However, current assessment relies heavily on clinical observation, typically performed by staff and interpreted by experienced microsurgeons [,]. These approaches are labor-intensive, require continuous bedside monitoring, and are inherently subjective, often ambiguous and challenging to interpret. Identifying key predictive factors and prioritizing high-risk patients for intensified monitoring may reduce the risk of flap failure and improve microsurgical outcomes [-]. In recent years, advances in computing and data science have accelerated the integration of AI into clinical workflows []. AI systems are capable of emulating aspects of human perception and reasoning while reducing fatigue-related errors and certain systematic biases []. Among these, convolutional neural networks, which process images through hierarchically arranged filters analogous to the visual cortex, have demonstrated utility in medical imaging tasks. These include anatomical landmark detection, vascular segmentation, and perforator mapping, where high accuracy is essential for surgical planning [,]. Within AI, machine learning, a data-driven approach that enables pattern recognition and predictive modeling, has shown promise in various domains of plastic surgery, including wound analysis, surgical planning, dermatologic diagnosis, and postoperative outcome prediction [,-]. A growing number of studies have also investigated the use of AI in predicting and assessing flap-related complications []. While these tools hold considerable promise in augmenting clinical judgment and reducing human error, the diagnostic accuracy and clinical reliability in free flap monitoring remain insufficiently established [,]. This uncertainty poses a significant barrier to widespread clinical implementation. Two systematic reviews [,] have recently summarized the application of AI in free flap surgery. However, the available evidence remains incomplete in several respects. First, both reviews framed the outcome broadly as the prediction of postoperative complications rather than the diagnostic performance of AI for real-time flap monitoring and early recognition of vascular compromise, the specific scenario in which the interval to reexploration determines salvage []. Second, neither review distinguished image-based surveillance models from clinical variable-based risk-prediction models, although these approaches differ fundamentally in their inputs, intended point of care, and expected diagnostic behavior, and may therefore warrant separate synthesis []. Third, both searches predate a rapidly expanding body of image-based monitoring evidence, including camera-integrated deep-learning platforms, optical detection of venous congestion, and remote real-time surveillance systems [-]. These gaps underscore the need for an updated diagnostic test accuracy synthesis that reflects current clinical applications of AI and separately evaluates image-based and nonimage-based models according to their intended roles in postoperative free flap care. Therefore, we conducted a systematic review and meta-analysis to evaluate the diagnostic performance of AI-based approaches in predicting and detecting free flap compromise following microsurgical reconstruction. We also compared the diagnostic performance of image-based and nonimage-based AI models to better characterize their role in postoperative free flap monitoring. Methods Overview This systematic review and meta-analysis was registered with PROSPERO (International Prospective Register of Systematic Reviews; CRD420251175572) and conducted in accordance with the PRISMA-DTA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses of Diagnostic Test Accuracy Studies) guidelines [] and the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 statement []. The study followed Transparency in the reporting of AI (TITAN) recommendations []. The literature search was reported in accordance with the PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses literature search extension) extension to enhance the transparency and reproducibility of the search process. The complete search strategies for all databases and the completed PRISMA-S checklist are provided in and , respectively []. Search Strategies A comprehensive search was performed across PubMed, Embase, Cochrane, Web of Science, and Scopus databases for studies published up to June 21, 2026, without language restrictions (). Database searches were updated before manuscript revision to capture recently published studies. The search was designed to identify studies that developed and validated deep learning or machine learning algorithms for the diagnosis of free flap status following microsurgery, using either imaging or clinical variables. Keywords and MeSH terms included “Artificial Intelligence,” “Deep learning,” “Machine learning,” “Radiomic,” “Automated diagnosis,” “Free flap reconstruction,” “Flap monitoring,” “Flap viability,” “ischemic complications,” and “Flap failure” (). Search strategies were developed independently by the investigators and were not adapted from previous systematic reviews [,]. Backward citation searching was conducted by manually screening the reference lists of all included studies and relevant review papers. Forward citation searching was also performed to identify studies citing the included papers. No other supplementary search methods were used. Selection Criteria Eligibility assessment was conducted independently by two reviewers (KCH, CHL) who screened titles and abstracts of the search results, with discrepancies resolved through discussions. The inclusion criteria were: (1) original research paper evaluating the performance of AI in identifying or predicting flap compromise, regardless of model type; (2) cohort studies incorporating diagnostic analysis; (3) study populations consisting of patients who underwent microsurgical reconstruction; and (4) clearly defined diagnostic criteria for flap compromise as stated in the study with available and extractable data. Studies were excluded based on the following criteria: (1) nonoriginal study types, including systematic reviews, review papers, case reports, commentaries, and letters; (2) irrelevant research focus, such as basic science experiments; (3) studies lacking AI applications or relevance to flap-related outcomes; (4) incomplete diagnostic data, particularly when key metrics such as true positive (TP), true negative (TN), false positive (FP), and false negative (FN) could not be determined; (5) unavailable full-text articles; and (6) duplicate publications. There were no language restrictions applied in the search strategy. Data Extraction and Management The data extraction framework was developed with reference to the Cochrane Handbook guidance [,]. Data extraction was performed independently by two reviewers (KCH and YYC) using a predefined data extraction sheet. Extracted variables included study design, country, year of publication, authorship, patient population characteristics such as gender and age, and model-related characteristics including training and testing sample sizes, validation methods, reference standards, AI model type, and modality of flap assessment. Diagnostic performance metrics, including sensitivity, specificity, and the corresponding TP, FP, TN, and FN counts, were collected. When studies included both animal and human experiments, only human data were extracted for quantitative synthesis. Information regarding missing data handling and imputation methods was also extracted when reported. Study independence was confirmed based on author group, region, study period, and institution, with no evidence of overlapping cohorts, and analyses were conducted at the patient level with image- or pixel-level data aggregated accordingly to avoid distortion of diagnostic estimates []. When multiple models were reported within a study, the best-performing model identified by the original authors was selected. If no preferred model was specified, the algorithm with the highest area under the curve (AUC) was chosen, with priority given to higher sensitivity because of the clinical importance of early flap compromise detection. To avoid overrepresentation of individual cohorts, only one model from each study was included in the pooled analysis. Contingency tables were reconstructed directly from reported results or supplementary data or back calculated from available performance metrics. When necessary, corresponding authors were contacted for clarification or provision of additional data. Risk-of-Bias and Quality Assessment Methodological quality was assessed using the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) tool, which evaluates four domains: patient selection, index test, reference standard, and flow and timing. Each domain was rated independently by two reviewers (KCH, MJW) for risk of bias as low, high, or unclear, and the first 3 domains were also appraised for applicability concerns []. Discrepancies were adjudicated through discussion or consultation with a third senior reviewer (CHL). The certainty of evidence was assessed using the GRADE (Grading of Recommendations Assessment, Development and Evaluation) approach for diagnostic test accuracy studies []. Diagnostic Performance and Statistical Analysis Statistical analyses were performed using R (R Foundation for Statistical Computing) and Stata version 18.0 (StataCorp LLC) [-]. A bivariate random-effects model was applied to jointly synthesize sensitivity and specificity, generating pooled estimates with corresponding 95% CIs. Diagnostic odds ratios (DOR) were also computed. Overall diagnostic performance was visualized using summary receiver operating characteristic (SROC) curve, with the AUC reported as an indicator of discriminatory capacity, where values approaching 1.0 reflected superior accuracy []. For univariate random-effects analyses, the Sidik-Jonkman estimator was used to estimate between-study variance, and the Hartung-Knapp adjustment was applied to calculate the 95% CIs in R. Corresponding 95% prediction intervals (PIs) were calculated and displayed in the forest plots. Sensitivity and specificity were pooled as logit-transformed proportions, whereas DOR and likelihood ratios were pooled on the natural logarithmic scale. For studies containing at least one zero cell among the TP, FP, TN, or FN counts, a continuity correction of 0.5 was added to all four cells before calculation of the DOR, likelihood ratios, and their variances. No continuity correction was applied to studies without zero cells. These supplementary analyses did not replace the primary bivariate diagnostic meta-analysis []. Statistical significance was defined as P<.05. Clinical Value Verification The clinical utility of AI-based diagnostic approaches was further explored through Fagan plot analysis, which translated pooled diagnostic metrics into posttest probabilities, thereby enabling estimation of their clinical applicability in predicting flap compromise and related complications [,]. The pooled prevalence of flap compromise was estimated by dividing the total number of TP and FN cases by the overall sample size (TP+FN+FP+TN) across included studies. Publication bias was assessed using Deeks funnel plot asymmetry test, with significant asymmetry (P<.10) interpreted as suggestive of small-study effects or selective reporting. Subgroup Analysis To address potential sources of heterogeneity, prespecified subgroup analyses were conducted according to study methodology and diagnostic focus (, illustrating approaches from: first approach used perioperative clinical variables for risk prediction of flap-related outcomes, second approach applied postoperative image-based analysis to identify flap compromise through changes in flap color and surface characteristics). Studies were categorized as image-based models for real-time postoperative flap surveillance and early detection of flap compromise, or nonimage-based models using clinical and perioperative variables for risk stratification and prediction of flap perfusion-related complications. Additional subgroup analyses focused on models specifically designed to detect vascular compromise. Results were presented using SROC curves and forest plots to facilitate subgroup comparisons. Heterogeneity and Sensitivity Analysis Between-study heterogeneity was further quantified using the I2 statistic, with values greater than 50% considered to represent substantial heterogeneity. PIs were additionally calculated to provide a clinically relevant interpretation of between-study variability []. A bivariate box plot was also generated to visually inspect the distribution of sensitivity and specificity and to identify potential outliers. The robustness of the primary meta-analysis was interrogated through sensitivity analyses, which included assessments of goodness of fit via Q–Q plot of deviance residuals, evaluation of bivariate normality using a Mahalanobis distance-based Q–Q plot, identification of influential studies through Cook distance, and systematic detection of outliers. Outliers were defined as studies with standardized residuals exceeding 2.0 for sensitivity of FP rate, corresponding to a 99% CI (z score±2.0), or those demonstrating significant deviation from expected model fit. A conservative approach was employed to minimize inappropriate exclusions. Identified outliers were removed and the bivariate model was reestimated to determine their impact on pooled performance metrics and heterogeneity. Publication bias was assessed using Deeks funnel plot asymmetry test, with P<.10 considered suggestive of significant small-study effects. Ethical Considerations Ethical approval was not required because this study was a systematic review and meta-analysis based exclusively on published data. Results Literature Search The process of study acquisition is illustrated in the PRISMA flow chart (). A total of 2098 studies were initially retrieved, of which 792 studies were excluded as duplicates. After screening the remaining 1306 papers by title and abstract, 1177 were excluded, leaving 129 full text studies for detailed assessment. Following eligibility evaluation, 111 studies were excluded, resulting in 18 studies [,-,-] ultimately included in this meta-analysis. Seventeen studies included in the quantitative diagnostic accuracy meta-analysis are summarized in and . All evaluated AI for the prediction or detection of free flap compromise. One additional study [] met the eligibility criteria and is presented in the PRISMA flow diagram and ; however, it was excluded from the quantitative synthesis because insufficient data were available to reconstruct the 2×2 contingency table. Most included studies were retrospective single-center cohort studies, although several investigations incorporated multicenter or externally validated datasets. presents the basic study characteristics, including design, patient population, and clinical context, while details model specifications, validation methods, and diagnostic performance metrics. Most studies were retrospective (n=15) in design, with only two prospective investigations conducted by Huang et al [] and Kim et al []. External validation was reported in five studies [,,,,]. Eight studies [-,-,,] applied image-based analysis, using photographic data to evaluate flap viability primarily based on colorimetric features. The remaining 9 studies [,-,,] developed predictive models using diverse clinical variables to identify patients at risk for flap compromise. These models integrated preoperative (eg, laboratory values, neoadjuvant therapy, and medication history), intraoperative details (eg, operative duration, ischemic time, flap selection, and laterality), and postoperative parameters (eg, hypotensive episodes) details. Demographic and clinical factors, including age, BMI, comorbidities, and primary tumor characteristics such as metastatic status, were also incorporated. Although the included studies encompassed a variety of free flap types, the anterolateral thigh (ALT) flap for head and neck reconstruction was the most frequently investigated. Risk-of-Bias and Certainty of Evidence Risk-of-bias assessment using the QUADAS-2 tool is summarized in -. Most studies were rated as low risk across all domains; however, unclear or high risk of bias remained common, particularly in the index test domain, where 47.1% (8/17) and 23.5% (4/17) of studies were rated as unclear and high risk, respectively. High risk of bias was also identified in patient selection (5.8%; 1/17) and flow and timing (11.8%; 2/17). Applicability concerns were generally low, although isolated studies demonstrated high or unclear concern in patient selection, index test, and reference standard domains. GRADE assessment () rated the certainty of evidence for the pooled sensitivity and specificity as low, primarily because of study limitations identified by QUADAS-2 and substantial between-study heterogeneity. No serious concerns were identified regarding indirectness, imprecision, or publication bias. Diagnostic Performance Pooled diagnostic accuracy across the 17 studies [,,,-] demonstrated an overall sensitivity of 0.83 (95% CI 0.70‐0.91; PI 0.21‐0.99) and a specificity of 0.87 (95% CI 0.65‐0.96; PI 0.03‐1.00). The AUC was 0.92 (95% CI 0.90‐0.94), with a positive likelihood ratio (PLR) of 7.63 (95% CI 3.44‐16.93; PI 0.29‐197.90), a negative likelihood ratio (NLR) of 0.22 (95% CI 0.10‐0.47; PI 0.01‐5.11), and a DOR of 36.17 (95% CI 8.27‐158.26; PI 0.09‐14886.43; -). Study-level DORs varied widely, ranging from 0.02 to 5390.00, largely because DOR estimates are highly sensitive to very small FP or FN counts. The paired forest plot () showed that image-based models generally achieved higher sensitivity than models based on clinical variables, with 7 studies [,,-,,] reporting sensitivities exceeding 90%. Specificity estimates ranged from 0.00 to 1.00 across the included studies. Substantial between-study heterogeneity was observed for both sensitivity: (I2=98.1%) and specificity (I2=98.6%), with broad prediction intervals (sensitivity: 0.21‐0.99; specificity: 0.03‐1.00). These findings should be interpreted together with the substantial heterogeneity and the low certainty of evidence (). Clinical Value Analysis To assess clinical applicability, a Fagan nomogram (Figure S1 in ) was constructed using a pretest probability of 19%, reflecting the estimated prevalence of flap compromise in the included population. A positive test result increased the posttest probability of flap compromise from 19% to approximately 62%, while a negative test result reduced it to 4%. These findings highlight the strong rule-in and rule-out capabilities of AI-based models, underscoring their potential utility in clinical decision-making for flap monitoring. Subgroup Analysis Image-Based Assessment of Flap Viability Subgroup analyses revealed notable differences between methodological approaches. Pooled analysis of the 8 image-based studies [-,-,,] demonstrated a sensitivity of 0.92 (95% CI 0.81‐0.97; PI 0.43‐0.99) and a specificity of 0.95 (95% CI 0.86‐0.98; PI 0.43‐1.00), with an AUC of 0.98 (95% CI 0.96‐0.99) and a DOR of 206.05 (95% CI 40.65‐1044.49; PI: 2.29‐18579.34), confirming robust diagnostic accuracy (). Visual inspection of the forest plot demonstrated consistently high performance across most studies, with sensitivity and specificity estimates exceeding 0.90 (). Zhang et al [] reported the lowest sensitivity (0.60, 95% CI 0.15‐0.95), whereas Maktabi et al [] demonstrated the lowest specificity (0.75, 95% CI 0.60‐0.86). Performance of Clinical Variable-Based Models Pooled analysis of the 9 studies [,-,,] relying on clinical variables demonstrated more modest performance, with a sensitivity of 0.69 (95% CI 0.45‐0.86; PI 0.11‐0.98) and a specificity of 0.72 (95% CI 0.18‐0.97; PI 0.00‐1.00). The corresponding AUC was 0.79 (95% CI 0.75‐0.82) (), and the DOR was 7.64 (95% CI 1.01‐57.57; PI 0.02‐3623.04), indicating only moderate discriminative ability. Variability was evident in the paired forest plot (), with Wang et al [] reporting the very lowest sensitivity (0.24, 95% CI 0.22‐0.26), whereas Asaad et al [] demonstrated an extreme specificity of 0.00 (95% CI 0.00‐0.00). Prediction and Detection of Vascular Compromise The pooled analysis of 8 studies [-,-,,] specifically targeting the prediction and detection of vascular crisis demonstrated strong overall diagnostic accuracy. The model yielded a pooled sensitivity of 0.92 (95% CI 0.81‐0.97; PI 0.44‐0.99) and a specificity of 0.91 (95% CI 0.79‐0.97; PI 0.35‐1.00; and ). This robust performance was further reflected in an AUC of 0.97 (95% CI 0.95‐0.98; ) and a DOR of 124.23 (95% CI 20.76‐743.41), indicating satisfactory discriminatory ability for predicting vascular compromised flaps. The paired forest plot () demonstrated high heterogeneity in both sensitivity (I2=79.4%) and specificity (I2=95.4%). Zhang et al [] appeared to contribute disproportionately to the observed variability, reporting the lowest sensitivity and the widest CIs for both sensitivity and specificity, suggesting reduced estimate precision. Sensitivity Analysis To investigate the sources of heterogeneity and assess the robustness of the findings, a comprehensive sensitivity analysis was performed. Initial inspection of the bivariate boxplot suggested that 3 studies [,,] lay outside the main data cluster, raising concerns about its influence (Figure S2a in ). This prompted a more rigorous statistical evaluation using multiple diagnostic tools. Goodness of fit and bivariate normality were assessed using Q-Q plots based on deviance residuals and Mahalanobis distances, respectively (Figure S2b and S2c in ). Additional influence analysis using Cook distance (Figure S2d in ) and outlier detection based on standardized residuals from the bivariate Reitsma model (Figure S2e in ) were also undertaken. A conservative threshold of > |2.0| for standardized residuals was applied to identify outliers. These diagnostics identified 2 studies of concern, mainly based on the results of influence analysis. Asaad et al [] was identified as a statistical outlier (Figure S2e in ), while both Asaad et al [] and Wang et al [] were shown to exert a disproportionate influence on the pooled estimates (Figure S2d in ). To evaluate their impact, the sensitivity analysis was performed after excluding Wang et al [] and Asaad et al []. The exclusion did not materially change the summary estimates; both the I2 and the AUC of the SROC remained stable (-). Since no other studies exceeded the predefined thresholds, all remaining studies were retained in the final model. Collectively, these findings confirm that the overall results are robust and not unduly influenced by any single study, thereby reinforcing the validity of the pooled conclusions. Small-Study Effects Small study effects were assessed using Deeks funnel plot asymmetry test (Figure S3 in ). The analysis revealed no significant funnel plot asymmetry was observed (P=.15>.10), indicating no evidence of small-study effects. Discussion Principal Findings This systematic review and meta-analysis provides a comprehensive evaluation of AI–based algorithms for postoperative surveillance of microvascular free flaps. Overall, image-based approaches demonstrated favorable diagnostic performance and generally outperformed models based solely on clinical variables, supporting the potential role of AI-assisted monitoring in microsurgical care. Timely detection of vascular compromise remains fundamental to successful free flap salvage []. Although overall flap success exceeds 95%, vascular compromise remains one of the leading causes of flap loss and frequently necessitates urgent reexploration [,]. Current postoperative monitoring relies heavily on serial clinical examination, making assessment susceptible to interobserver variability, fatigue, and subjective interpretation [,]. Several adjunctive monitoring technologies, including implantable Doppler devices, tissue oximetry, near-infrared spectroscopy, and indocyanine green angiography, have improved flap surveillance, but remain limited by cost, workflow integration, equipment availability, and the absence of standardized protocols [,,,]. AI offers a complementary strategy by providing continuous and objective interpretation of clinical or imaging data []. In addition to postoperative monitoring, AI-based prediction models may facilitate preoperative risk stratification, enabling individualized perioperative planning and more efficient allocation of monitoring resources []. In this study, models incorporating photographic or image-based inputs consistently outperformed those based solely on clinical variables. This finding is biologically plausible because visual information directly reflects dynamic physiological changes, including alterations in flap color, congestion, and tissue perfusion, that may precede overt clinical deterioration and cannot be fully captured by conventional clinical variables alone []. Studies by Huang et al and Maktabi et al further support this concept using RGB color analysis and hyperspectral imaging–derived physiologic parameters [,]. To better reflect clinical applicability, data from Huang et al [] were extracted from an independent clinical validation cohort rather than the training cohort, thereby providing an estimate that more closely represents real-world diagnostic performance. Recent advances also illustrate the rapid evolution of this field. Huang et al [] employed a Random Forest model [], Kim et al [] implemented a DenseNet121 deep learning architecture using TensorFlow (Google LLC) in Python 3.7 (Python Software Foundation) [], Hsu et al [] introduced the smartphone-integrated “FLAPMATE” platform for real-time postoperative monitoring, and Huang et al [] presented the “DLscope” system, which integrates a camera device, data transmission, server, and smartphone application to support remote flap surveillance. Collectively, these developments highlight the increasing feasibility of incorporating AI-assisted assessment into routine microsurgical practice [,,]. In contrast, models that relied exclusively on clinical variables achieved more modest performance, likely reflecting both the multifactorial nature of flap compromise and heterogeneity in predictor selection across studies [,-,,]. However, these models may still provide practical value for risk stratification. By categorizing patients into high and low risk tiers, these models may help guide the intensity of postoperative surveillance and allocation of clinical resources. Frequently identified predictors, including smoking, female sex, and elevated BMI, are well-recognized clinical risk factors, suggesting that the strength of these models lies in integrating multiple modest predictors into an individualized estimate of postoperative flap-related risk [-]. Despite their promising diagnostic performance, several factors currently limit the translation of AI into routine microsurgical practice []. Image-based algorithms remain susceptible to variations in lighting conditions, camera systems, skin pigmentation, wound dressings, surgical drains, and postoperative edema, all of which may influence image quality and model performance [,]. In routine postoperative settings, image quality may also be influenced by edema, dried blood, dressings, surgical drains, or Doppler wires []. Moreover, clinically important findings such as early venous congestion may present with only subtle visual changes, making robust model generalization across institutions and patient populations particularly challenging []. These observations highlight the importance of prospective multicenter validation and standardized image acquisition before widespread clinical implementation []. Substantial clinical and methodological heterogeneity was observed across the included studies [,,,-]. Outcome definitions ranged from vascular compromise requiring reexploration to complete flap failure, while AI approaches varied from conventional machine-learning algorithms to deep learning-based image analysis. Considerable differences were also observed in predictor selection, sample size, event rates, and the timing of model application, with studies evaluating both preoperative risk prediction and postoperative flap surveillance []. These variations likely contributed to the substantial heterogeneity observed in the pooled analyses. The wide variation in study-level DORs should not be interpreted solely as indicating intrinsic superiority of certain AI models. Because DOR is highly sensitive to small FP or FN counts, studies with few misclassifications may produce very large estimates with wide CIs []. Differences in populations, outcome definitions, AI algorithms, decision thresholds, validation methods, and class distributions may have further contributed to this variability. Importantly, the pooled sensitivity, specificity, and DOR should be interpreted as estimates of the average diagnostic performance across the included studies. Although the CIs describe the uncertainty around these pooled average estimates, the wide PIs demonstrate that diagnostic performance may vary substantially across different clinical settings []. Therefore, the pooled estimates should not be interpreted as performance that can be uniformly expected in all practice environments. The methodological assessment also supports cautious interpretation of these findings. QUADAS-2 identified concerns regarding patient selection and index test methodology in several studies [], and the GRADE assessment rated the certainty of evidence as low because of study limitations and substantial between-study heterogeneity []. Although the pooled estimates suggest favorable diagnostic performance, the low certainty of evidence according to GRADE indicates that these findings should be interpreted cautiously and may not be readily generalizable across different clinical settings. Accordingly, the current evidence supports the potential of AI-assisted monitoring but remains insufficient to justify widespread clinical implementation without prospective multicenter validation and rigorous external evaluation. Another important methodological consideration relates to model development and reporting quality []. Flap compromise was an infrequent event in many included datasets, resulting in substantial class imbalance []. Under these conditions, models may achieve high specificity and AUC despite limited sensitivity for detecting uncommon adverse events []. This phenomenon was evident in several studies, particularly those with very low event rates, and may have contributed to optimistic estimates of diagnostic performance []. Although some investigators adopted imbalance-correction techniques, including data augmentation or synthetic oversampling, these approaches were inconsistently applied across studies and may have contributed to the observed between-study heterogeneity [,]. Several limitations should be acknowledged. Most included studies were retrospective, single-center investigations with limited external validation, increasing the risk of selection bias and model overfitting [,]. In addition, several studies demonstrated unclear or high risk of bias in the patient selection and index test domains, primarily related to limited reporting transparency and lack of external validation. Although sensitivity analyses demonstrated relative stability of pooled estimates, these methodological limitations should be considered when interpreting the findings. To ensure consistent quantitative synthesis, several methodological decisions were required during data extraction. When multiple AI algorithms were reported within a study, only the best-performing model was included because it most closely represents the model likely to be translated into clinical practice. Although this approach may slightly overestimate diagnostic performance, including multiple models derived from the same cohort would have disproportionately weighted individual datasets and violated the assumption of independent observations []. Furthermore, prioritizing higher sensitivity as a secondary tie-breaking criterion may increase FP alerts and clinical workload, but this situation occurred in only one included study because AUC alone determined model selection in the remaining studies []. Therefore, the impact of this selection strategy on the pooled estimates was likely limited. Outcome definitions also required harmonization across studies. For example, Kim et al [] originally classified postoperative outcomes as confirmed compromised, suspicious, and normal flaps. To maintain consistency with diagnostic test accuracy methodology, suspicious and normal flaps were combined into a single non-compromised category before reconstruction of the 2×2 contingency table. Similarly, several studies did not report TP, FP, TN, and FN directly, requiring reconstruction from published diagnostic performance measures [,,-]. Although this approach is widely accepted in diagnostic accuracy meta-analyses, it may have introduced additional uncertainty into the pooled estimates []. Crucially, the diagnostic fidelity of these AI systems remains inherently tethered to the “silver standard” of current microsurgical practice [,]. Given that the ground truth often relies on subjective clinical intuition, characterized by an inherent inter-observer variability, the reported diagnostic performance may represent a theoretical ceiling imposed by the noise within training labels []. Furthermore, the necessity of post hoc data reconstruction for 44% (8/18) of the included studies suggests a pervasive reporting gap in the microsurgical literature []. This review provides the most comprehensive synthesis of diagnostic test accuracy evidence currently available for AI in postoperative free flap monitoring [,]. Unlike previous systematic reviews, this review adopts a clinically oriented approach by focusing on postoperative free flap monitoring rather than postoperative complication prediction alone. It further distinguishes image-based surveillance from clinical variable-based risk stratification, providing evidence for their complementary roles in postoperative free flap care. By applying established diagnostic test accuracy methodology, this review provides a robust evaluation of AI-assisted detection of free flap compromise. It further demonstrates how different AI approaches may complement postoperative care, with image-based algorithms assisting diagnosis and clinical variable-based models enabling individualized risk stratification and more efficient allocation of monitoring resources. Collectively, these findings suggest that AI may contribute to postoperative free flap care in multiple ways. Image-based algorithms demonstrated strong diagnostic performance for early detection of vascular compromise and may enhance real-time postoperative surveillance through objective and reproducible assessment. In contrast, models based on clinical variables may provide additional value for perioperative risk stratification, enabling individualized surveillance strategies and more efficient allocation of monitoring resources. Together, these approaches have the potential to improve the consistency of flap monitoring, facilitate timely clinical decision-making, and complement rather than replace clinical judgment. Before routine clinical implementation, prospective multicenter studies with standardized imaging protocols, rigorous external validation, and transparent reporting in accordance with Standards for Reporting of Diagnostic Accuracy Studies-AI (STARD-AI) and decision support systems driven by artificial intelligence (DECIDE-AI) are needed to establish generalizability, support regulatory evaluation, and ensure the safe and responsible integration of AI into microsurgical practice [-]. Conclusion AI-driven flap surveillance, particularly through image-based deep learning, demonstrates strong diagnostic performance. Its application may facilitate earlier detection, reduce diagnostic errors, and support postoperative monitoring in both hospital and remote settings. However, these findings should be interpreted cautiously because several studies were based on relatively small retrospective datasets with limited external validation. Broader clinical adoption will require further high-quality prospective validation and standardized reporting, and these technologies should be integrated into clinical practice as adjuncts to, rather than replacements for, surgical judgment.