Ovarian cancer (OC) is the deadliest gynecologic malignancy, characterized by nonspecific symptoms that often remain undetected until the disease has progressed [1,2]. The conventional diagnosis of OC principally depends on imaging techniques (encompassing ultrasound, computed tomography, and magnetic resonance imaging); serum biomarkers (such as cancer antigen 125, carcinoembryonic antigen, and human epididymis protein 4); along with the invasive procedure (histological biopsy) [3,4]. However, the sensitivity and specificity of imaging techniques and biomarkers are restricted [5]. Furthermore, the histopathological test is inherently invasive [3]. Therefore, there is an urgent need for more accurate, noninvasive, and reliable diagnostic methods.

Minimally invasive diagnostic procedures, particularly the use of blood samples, are among the foremost common methods of detection [6]. Moreover, patients are generally more willing to undergo blood tests, leading to higher compliance rates [7]. Blood contains a rich repertoire of biomolecules, including proteins, nucleic acids, and metabolites, which can potentially serve as indicators of OC diagnosis [8-11]. The development of omics has opened new doors for biomarker discovery. The genomics, proteomics, and metabolomics of blood samples can provide a wealth of information about the molecular changes in cancer [12]. For instance, Ke et al [13] systematically investigated OC metabolism through the metabolic profiling of 448 plasma samples. The analysis of dysregulated metabolic pathways extends our current understanding of OC metabolism. Similarly, Dhar et al [8] applied glycoproteomics to serum of women with OC or benign pelvic masses and healthy controls and analyzed glycosylation patterns in serum markers and supported the hypothesis that blood glycoproteomic profiling can be used for OC diagnosis and staging. Notably, the exponential growth of multiomics data has presented a major challenge that surpasses traditional analytic capabilities [14,15]. Fortunately, artificial intelligence (AI) algorithms offer a solution [16,17]. AI can manage complex datasets and spot hidden patterns and potential biomarkers, enabling more accurate OC diagnosis and personalized treatment.

AI, particularly machine learning (ML) and deep learning (DL), has attracted increasing attention in medical research due to its capability to analyze large biomedical datasets [18-23]. AI-driven models have emerged as a promising tool for developing predictive models for OC by analyzing complex and multidimensional datasets to uncover biomarkers. For instance, a multicenter retrospective study screened 52 features from laboratory tests in blood samples to build an ML model. Integrating 20 base AI models, it performed well internally and externally, outperforming the CA125 and HE4 biomarkers in identifying OC [24]. In addition, a blood-based metabolite panel demonstrated independent predictive ability and complemented the risk of ovarian malignancy algorithm for distinguishing early-stage OC from benign disease to better inform clinical decision-making [25]. Despite such encouraging results, the related results are scattered. Whether the application of AI in OC blood biomarker research can significantly improve diagnostic accuracy remains controversial.

Therefore, there is a crucial need for a high-quality synthesis of the available evidence. The purpose of this study is to provide a systematic overview of the diagnostic accuracy of AI techniques in identifying OC blood markers as well as to elucidate their applicability, potential, and limitations.

This meta-analysis was conducted in adherence to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) checklist (Multimedia Appendix 1) and MOOSE (Meta-Analysis of Observational Studies in Epidemiology) reporting guidelines [26,27]. The protocol was prospectively registered with the International Prospective Register of Systematic Reviews (PROSPERO; CRD42023481232).

A comprehensive search of the MEDLINE, Embase, IEEE Xplore, PubMed, Web of Science, and the Cochrane Library databases was carried out from inception to January 16, 2024. Detailed search strategies are summarized in Multimedia Appendix 2. Two independent investigators (MQJ and HLX) assessed the records after removing the duplicates at the title and abstract level, and finally at the full-text level, according to the inclusion and exclusion criteria (Textbox 1). Two investigators (MQJ and XYL) independently appraised the articles for eligibility. An inconsistency in selection was reconciled through discussion with a third independent investigator (HLX).

Data were independently extracted by 2 investigators (QPM and XYL) using a predefined data extraction sheet, with any discrepancies resolved through the adjudication of a third investigator (HLX). Necessary data of 2×2 contingency tables that included true positives (TP), false positives, true negatives (TN), and false negatives were extracted. The details of the data are presented in Multimedia Appendix 3 [8,24,25,28-64]. In cases where these values were not explicitly reported, values were calculated using descriptive statistics available in the study. For studies presenting multiple contingency tables, either for identical or disparate AI algorithms, each table was treated as an independent result [65].

The risk of bias and concerns about the applicability of all included studies were assessed by 2 independent investigators (MQJ and HLX), using the Quality Assessment of Diagnostic Accuracy Studies–Artificial Intelligence (QUADAS-AI) criteria [66]. Conflicts were discussed and solved with a third investigator (XYL). The risk of bias assessment included 4 domains: patient selection, index test, reference standard, and flow and timing. For assessing clinical applicability, only the first 3 domains were evaluated [66,67]. In addition, the median was used as the threshold for determining the risk level of bias, with studies classified as low risk if ≥2 domains were deemed as low risk and high risk if <2 domains were considered low risk [68].

We used the bivariate diagnostic random effects model to compute the summary receiver operating characteristics to determine summary estimates of the sensitivity, specificity, and area under the curve (AUC) with their respective 95% CIs [69]. Sensitivity was defined as the probability of a person with OC having a positive test result, indicating the capacity of the index test to identify patients, considered by the equation: sensitivity = TP / (TP +  false negatives). Specificity was the probability of a woman without OC having a negative test result, reflecting the test's ability to correctly identify OC-free individuals calculated by the equation: specificity = TN / (false positives + TN) [70,71]. The performance of the test can also be assessed using the AUC. This area may be interpreted as the probability that a random woman with OC has a higher value of the measurement than a random person without OC. In general, an AUC of >0.80 is considered good [72]. A perfect test would have an AUC of 1 and a useless test would have an AUC of 0.5 [73].

Study heterogeneity was determined using the I² statistic and the Q test [74]. When the same or different AI models were tested within the same article, the proposed model with the best accuracy was used for further meta-analysis [65]. Subgroup and regression analyses were performed to explore potential sources of heterogeneity. Sensitivity analysis was implemented via a qualitative systematic review of studies from which concatenated tables could not be extracted or constructed. Publication bias was assessed using funnel plot asymmetry test by Deeks et al [75].

Subgroup analyses were performed according to the following: (1) AI algorithms (ML or DL), (2) external validation (yes or no), (3) levels of risk of bias (low or high), (4) year of publication (after or before 2022), (5) geographical distribution (Asia, North America, or Europe), (6) sample size (>300 or ≤300 as median), (7) blood sample type (serum or plasma), (8) biomarkers type (protein or mixed), and (9) number of modeling biomarkers (>8 or ≤8 as median).

The variability in sensitivity and specificity estimates was graphically represented through a cross-hairs plot, generated using R software (version 4.2.1; R Foundation for Statistical Computing) [76]. All other statistical analyses were conducted in Stata (version 17.0; StataCorp). The statistical significance was defined as P<.05.

The database search identified 1566 records from which 604 duplicates were removed. We then performed the title and abstract screening of 962 records and subsequently a full-text evaluation of 55 records. Following the exclusion of 15 articles, as detailed in Multimedia Appendix 4 [77-91], a total of 40 studies were included in this meta-analysis (Figure 1).

Most of the studies were performed with retrospectively (21/40, 52%) and prospectively (18/40, 45%) collected data, and one study collected both retrospective and prospective data (Table 1). In total, 12% (5/40) of studies sourced their data from public databases. In terms of AI algorithm types, a total of 36 (90%) studies were classified as ML, whereas 10% (4/40) of studies were classified as DL. Most (36/40, 90%) of the studies validated their algorithms, while only some (7/40, 18%) studies carried out an external validation (Table 2). The blood samples were mainly serum (27/40, 68%), and the type of blood biomarkers was mainly protein (25/40, 62%; Table 3).

The quality of the included studies was appraised using the QUADAS-AI (Figures S1 and S2 in Multimedia Appendix 5). In detail, most of the studies were rated as having a high or unclear risk of bias based on patient selection (22/40, 55%) and index test (33/40, 82%) domains. These assessments might be attributed to the absence of explicit delineation of included patients, such as previous testing history and clinical setting, as well as to deficiencies in rigorous external validation of the AI models.

The summary receiver operating characteristics curves for the 40 included studies with 342 contingency tables are shown in Figure 2. The pooled sensitivity and specificity were 85% (95% CI 83%-87%) and 91% (95% CI 90%-92%), respectively, with an AUC of 0.95 (95 % CI 0.92-0.96) for all AI algorithms. Notably, when contingency tables with the highest accuracy were extracted from each study, the pooled sensitivity and specificity were 95% (95% CI 90%-97%) and 97% (95% CI 95%-98%), respectively, with an AUC of 0.99 (95% CI 0.98-1.00). Reported point estimates and CIs of all included studies are shown in a cross-hairs plot (Figure 3).

Results of the subgroup analyses revealed that acceptable diagnostic performance was observed in all subgroups, ranging from 74% to 98% for sensitivity and 85% to 96% for specificity. Detailed results are shown in Table 4, and the corresponding plots are presented in Figures S3-S21 in Multimedia Appendix 5. We divided the studies into subgroups according to the modalities of algorithms (Figures S3 and S13 in Multimedia Appendix 5), existence of external validation (Figures S4 and S14 in Multimedia Appendix 5), levels of risk of bias (Figures S5 and S15 in Multimedia Appendix 5), year of publication (Figures S6 and S16 in Multimedia Appendix 5), geographical distribution (Figures S7 and S17 in Multimedia Appendix 5), sample size (Figures S8 and S18 in Multimedia Appendix 5), blood sample type (Figures S9 and S19 in Multimedia Appendix 5), biomarkers type (Figures S10 and S20 in Multimedia Appendix 5), and number of modeling biomarkers (Figures S11 and S21 in Multimedia Appendix 5).

The meta-analysis uncovered substantial heterogeneity among studies, as evidenced by an I² of 97.54% (P<.05) for sensitivity and 99.15% (P<.05) for specificity (Figure S12 in Multimedia Appendix 5). To further explore the causes of study heterogeneity, a meta-regression analysis was conducted (Table 4). The results showed that both external validation and year of publication were significant factors that influenced study heterogeneity with regard to sensitivity and specificity. Subgroups based on levels of risk of bias, geographical distribution, sample size, and blood sample type showed intergroup heterogeneity in the sensitivity of prediction (P<.05). In terms of marker type, specificity presented significant heterogeneity between groups (P<.05).

A qualitative systematic review was performed for studies lacking directly or indirectly available contingency tables. The findings from this review were in alignment with the main analysis (Tables S1-S3 in Multimedia Appendix 6 [77-81]). In addition, the analysis did not reveal any publication bias in this meta-analysis (P=.72; Multimedia Appendix 7).

The burgeoning evolution of AI within the medical field has captured the attention of an increasing cadre of researchers, particularly in its applicability to disease diagnosis [92,93]. To the best of our knowledge, this meta-analysis is a pioneering effort specifically exploring the efficacy of AI in OC diagnosis via blood biomarkers. AI algorithms exhibited exceptional diagnostic capabilities for OC, boasting a pooled sensitivity of 85% (95% CI 83%-87%) and specificity of 91% (95% CI 90%-92%). Moreover, we identified substantial heterogeneity among the selected studies and determined the potential contributing factors through subgroup and meta-regression analyses. Overall, these results should be interpreted with caution as described by the constraints mentioned in subsequent sections.

Heterogeneity is an inevitable problem in meta-analyses [94]. Significant interstudy heterogeneity was noted in terms of sensitivity (I²=97.54%) and specificity (I²=99.15%) in this study. External validation emerged as a crucial variable influencing study heterogeneity. Studies without external validation might yield results that were hard to generalize owing to factors such as sample selection bias and the characteristics of the research setting [95]. To address this, future research should focus on standardizing and applying the validation procedures, thus getting closer to the goal of providing more accurate and reliable diagnostic tools for clinical practice [96]. Besides, several factors contributed to the heterogeneity observed in sensitivity in this study. Studies with a pronounced risk of bias were predisposed to introduce uncertainties. Such biases could stem from flaws in study design, improper data collection, or inappropriate statistical analyses, all of which might distort the true relationship between the biomarker and the condition under investigation [67,97]. On the other hand, geographic disparities might be attributed to a complex interplay of genetic polymorphisms and environmental factors that differentially modulate biomarker expression levels, which could result in significant variability in biomarker sensitivity [98,99]. In addition, larger sample sizes generally offer enhanced statistical power and precision, enabling more reliable estimations of biomarker performance [100]. Understanding and accounting for these factors comprehensively will help to further reduce heterogeneity and enhance the validity and clinical relevance of meta-analysis results, ultimately leading to more precise and useful diagnostic tools for clinical application.

Blood specimens are relatively stable and can be easily accessed. Therefore, blood-based biomarkers have been regarded as a minimally invasive method with great value for disease diagnosis [101,102]. Plasma and serum are rich sources of information regarding an individual’s health state and are the focus of this study’s investigation [103]. Particularly noteworthy is that serum samples showed higher sensitivity than plasma ones in this study, which can potentially be ascribed to multiple factors. First, the distinct sample preparation procedures for serum and plasma may lead to variations in the concentration and availability of biomarkers. Serum is obtained after blood clotting, during which certain intracellular components can be released, potentially augmenting the repertoire of biomarkers [104]. In contrast, the anticoagulants used in plasma collection may impede the integrity or accessibility of biomarkers [105]. Second, the microenvironment within serum and plasma varies. Serum contains a more intricate milieu of proteins, enzymes, and other biomolecules that can interact with cancer biomarkers in a manner that heightens their detectability [102]. For example, the presence of specific binding proteins or proteases in serum may modify the conformation of biomarkers, rendering them more amenable to detection by the analytic methods used [106]. Furthermore, the centrifugation processes involved in separating serum and plasma can differentially partition the biomarkers. The speed, duration, and temperature of centrifugation may cause certain biomarkers to be preferentially retained in the serum fraction, contributing to the observed higher sensitivity [107]. In addition, the limited number of included studies for plasma may contribute to this phenomenon to some extent. Therefore, it is of paramount importance to judiciously select the sample type in the context of developing and implementing blood-based biomarker assays for OC.

In subgroup analysis, ML surpassed DL in both sensitivity and specificity. This phenomenon warrants an in-depth exploration of the underlying reasons and improvement directions. The edge of ML likely stems from its algorithmic traits. For structured and well-defined data, traditional algorithms (eg, logistic regression and support vector machine) can adeptly capture biomarker-disease associations via mathematical and statistical tenets, yielding high diagnostic accuracy [108,109]. DL, empowered by its strong automatic feature extraction and complex architecture, can theoretically handle large data and extract deep patterns [110]. However, in this study, the number of studies included for DL was only 4, compared with 36 for ML, presumably constraining the exertion of DL’s advantages. The scant number of DL studies gives rise to data that are circumscribed in both sample variety and the expanse of feature distribution. To break through this dilemma, several optimization strategies can be considered. First, data augmentation, such as random rotations, scaling, flipping, and noise addition to the data can enhance dataset diversity, facilitating the DL model to learn more extensive features and patterns for better generalization [110-112]. Second, transfer learning is applicable. Using pretrained models from related medical or bioanalysis fields and fine-tuning with OC data, the DL model can draw on prior knowledge, accelerating training convergence and potentially improving performance [113,114]. In addition, model compression techniques, including pruning (ie, eliminating less important connections or neurons to maintain performance while reducing complexity) and quantization (ie, lowering parameter precision for faster inference and less memory use) can be used [115]. While these strategies hold promise, their implementation and efficacy in the context of OC diagnosis warrant further investigation and optimization.

At present, the data amassed for AI applications in OC diagnosis is circumscribed by the paucity of diverse external validation. Many studies rely on a single dataset for discovery, with cross-validation to estimate algorithm performance. Given the generalizable issues to unseen data, accuracy drops when tested on other research datasets, and substantially when tested on clinical data [116,117]. To address these challenges, several strategies can be implemented. First, multicenter collaborations should be actively pursued. Combining data from different medical institutions and regions to build a heterogeneous and comprehensive dataset and exposing algorithms to wider patient characteristics and biomarker profiles will enhance generalizability [118-120]. Second, standardized data collection and annotation protocols are crucial. They ensure data consistency and comparability among studies, minimizing variability from inconsistent methods [121]. This allows algorithms to be trained on more reliable and reproducible data, strengthening the foundation for AI application. Moreover, continuous evaluation and improvement of algorithms in clinical settings are essential. Prospective studies integrating the algorithms into routine practice and monitoring their performance can offer valuable feedback [122]. This iterative testing and refinement process helps algorithms adapt to clinical complexity, leading to more accurate OC diagnostic tools. Despite persisting challenges, we anticipate that these efforts will incrementally enhance diagnostic accuracy. Sustained refinement and collaboration are essential to exploiting the full potential of AI in OC diagnosis.

Beyond the previous mentions, the existing literature in the field exhibits certain areas of improvement to reduce the gap between research and deployment. First, one of the richest data sources of patient health and clinical history is embedded in the electronic health records of a patient but remains hugely underutilized. AI’s ability to integrate blood biomarkers with other clinically relevant nonblood biomarkers, such as age, cancer history, and family history of cancer, could potentially outpace current practices if trained on sufficiently extensive datasets [123]. Future research could explore the synergistic integration of AI tools with clinical expertise, echoing a more realistic clinical scenario. Second, the problem of explainability is the subject of intensive research and various initiatives. Although symbolic AI or simple ML models, such as decision trees or linear regression, are still fully understood by people, understanding becomes increasingly difficult with more advanced techniques and is now impossible with many DL models; this situation can lead to unexpected results and nondeterministic behavior [124,125]. Third, data privacy and patient consent are critical concerns that need to be addressed before adopting the use of AI in clinical practice [126]. The integration of AI into clinical workflows requires careful consideration of ethical, legal, and regulatory aspects [127,128]. Transparent guidelines and regulations should be established to govern the use of AI in health care and ensure its responsible and ethical implementation.

Our meta-analysis has several strengths. First, to the best of our knowledge, it represents a novel effort as the first systematic review and meta-analysis dedicated to evaluating the diagnostic performance of blood biomarker-based AI for OC. Our findings illuminate the considerable potential AI holds in this domain, while also highlighting the advantages of blood tests, such as their noninvasive nature, better patient compliance, and cost-effectiveness. Second, our comprehensive investigation included multiple subgroup analyses, all of which yielded acceptable diagnostic performance for the AI model. Third, the stringent quality assessment of all included studies was conducted using QUADAS-AI tool. In addition, the robustness of our meta-analysis was reinforced through a sensitivity analysis underpinned by a qualitative systematic review.

The results of our meta-analysis are likely to be overestimated or underestimated for some reasons. One limitation lies in the high heterogeneity of the studies included. Nonetheless, we thoroughly explored potential sources of between-study heterogeneity through meta-regression and subgroup analyses. Another limitation is that the contingency tables of 5 studies included in our systematic review were not directly or indirectly available. These studies provided only indicators, such as AUC, accuracy, and F₁-score, which did not allow for the construction of contingency tables [77-81]. Nevertheless, we conducted a qualitative systematic review of these 5 studies and discovered that the findings aligned with the main analysis. Moreover, although no publication bias was noticed, it is still highly likely that there is unpublished material for this topic from the ever-growing nature of the framework and the likelihood of undisclosed research for commercial development [129]. Moreover, available AI research tends to be skewed toward the publication of positive results, indicating a potential publication bias.

The findings of this study indicated that the use of AI for the analysis of noninvasive blood biomarkers in OC diagnostics holds substantial potential for achieving satisfactory predictive outcomes. Among the analyzed studies, those that used DL were notably fewer in number than those that used ML. This underscores a critical need for future research to prioritize the incorporation of DL methodologies. Furthermore, pursuing external validation datasets was a necessary avenue to optimize the performance and applicability of AI in this field.