Ovarian cancer is easy to metastasize, has a poor prognosis, and is the most lethal malignant tumor of the female reproductive system, posing a serious threat to women's life and health. However, due to the lack of effective biomarkers, it is difficult to diagnose ovarian cancer in time. Laboratory tests have been widely used in routine health check-ups, and some of them have been shown to be significantly related to the diagnosis and prognosis of ovarian cancer.
Although routine laboratory tests have the potential to be biomarkers for ovarian cancer, none of them currently provide sufficient sensitivity or specificity for individual ovarian cancer prediction. Excitingly, emerging artificial intelligence (AI) technologies can integrate data from multiple inspections to aid in diagnosis. In January of this year, The Lancet Digital Health published online a multicenter, retrospective cohort study in China that aims to systematically evaluate the value of routine laboratory tests for ovarian cancer diagnosis and to develop a stable and generalizable ensemble AI model to assist in the diagnosis of ovarian cancer [1]. This article summarizes the main points of the study for the benefit of readers.
Research Methods:
This retrospective, multicenter study collected 98 laboratory tests and clinical features in women diagnosed with ovarian cancer (including primary malignant ovarian, fallopian tube, or peritoneal cancer) and uterine adnexal benign lesions/normal physical examination between January 2012 and April 2021 in three hospitals. Subjects from Tongji Hospital affiliated to Tongji Medical College of Huazhong University of Science and Technology (TJ cohort) were selected as the training set, and the training set was cross-validated with five folds. The two external validation sets were from the Obstetrics and Gynecology Hospital of Zhejiang University School of Medicine (WHZJU cohort) and Sun Yat-sen University Cancer Center (SYSUCC cohort). Those with a history of other malignancies or precancerous lesions, those who have been pregnant or infected with HIV in the past 6 months, those who are not newly diagnosed in these three hospitals, and those who do not have available laboratory results are excluded. A total of 10,992 participants were included in the final study (Figure 1).
Fig.1 Flow chart of the study population selection
After data preprocessing, the researchers constructed a prediction model for ovarian cancer based on artificial intelligence fusion method, a Classification Fusion (MCF) model based on multi-criteria decision-making. The algorithm integrates 20 AI-based classification models (Fig. 2) and evaluates the performance of the models in an internal validation set (3007 people) and two independent external validation sets (5641 and 2344 people), respectively. The main outcome of the study was the predictive accuracy of ovarian cancer of the MCF model, which was quantified by the area under the receiver operating characteristic curve (AUC), accuracy, specificity, sensitivity, positive predictive value, negative predictive value and F1 score.
Figure 2 Schematic diagram of the MCF construction and verification process
Findings:
The baseline characteristics of the study population are shown in Table 1.
Table 1 Baseline characteristics of the study population
The MCF model for ovarian cancer diagnosis finally included 52 features (51 laboratory tests and age), and the importance ranking of these 52 features is shown in Figure 3. Its performance on the internal and two external validation sets is shown in Table 2, with AUCs of 0.949 (95% CI 0.948-0.950), 0.882 (95% CI 0.880-0.885), and 0.884 (95% CI 0.882-0.887), respectively.
Figure 3 Ranking of feature importance
Table 2 Manifestations of MCF model in the diagnosis of ovarian cancer
Compared with CA125, HE4 and their combinations, the MCF model (MCFall) showed significantly higher AUC and sensitivity in identifying patients with ovarian cancer (Fig. 4A-C), especially those with early-stage ovarian cancer (Fig. 4D-F). Even in the absence of CA125 and other tumor markers, the MCF model can still accurately predict the risk of ovarian cancer by using incomplete conventional laboratory tests and age, and the prediction effect is better than that of the currently reported ovarian cancer prediction ensemble model.
Fig.4 Comparison of the performance of different models
Conclusions of the study
The MCF model for ovarian cancer diagnosis based on routine laboratory tests has achieved satisfactory and stable performance in the three validation sets, providing a low-cost, easily accessible and accurate auxiliary diagnostic tool for ovarian cancer. Not only CA125, which ranked the highest importance in the diagnosis of ovarian cancer, but also the test indicators included in the model contributed to the identification of ovarian cancer patients.
The MCF model developed in this study has been packaged into an open-source ovarian cancer prediction tool, and the risk value of ovarian cancer can be calculated by entering the corresponding laboratory test data and age as prompted.
Bibliography:
1. Cai G ,Huang F ,Gao Y , et al. Artificial intelligence-based models enabling accurate diagnosis of ovarian cancer using laboratory tests in China: a multicentre, retrospective cohort study. [J]. The Lancet. Digital health,2024,6(3):e176-e186.
Disclaimer: This article is published with the support of AstraZeneca and is intended for healthcare professionals only
审批编号：CN-133822
Valid until: 2025-4-17