Clinical analysis and artificial intelligence survival prediction of serous ovarian cancer based on preoperative circulating leukocytes

This study retrospectively analyzed patients with ovarian serous carcinoma who were initially treated at the Department of Obstetrics and Gynecology at Beijing Chaoyang Hospital, Capital Medical University, from July 2009 to December 2018. The case inclusion criteria were as follows: (1) surgical treatment performed at Beijing Chaoyang Hospital, (2) confirmation of ovarian serous carcinoma (serous cystadenocarcinoma or high-grade serous cystadenocarcinoma) by postoperative pathology, (3) standard platinum-based chemotherapy after the first tumor reduction surgery, and (4) complete preoperative routine blood and clinical data. The exclusion criteria were as follows: (1) presence of other types of benign and/or malignant ovarian tumors, (2) presence of primary malignant tumors of other organs, (3) no standardized chemotherapy after the first tumor reduction operation, and (4) incomplete routine blood and clinical data. The obtained data included age, BMI, childbirth history, menopause, neoadjuvant chemotherapy, surgical satisfaction, differentiation, stage based on the 2014 International Federation of Gynecology and Obstetrics (FIGO) staging system [25], ascites cytology, lymph node metastasis (LNM), recurrence, which is defined as the time from the first cytoreductive surgery to the time of ovarian cancer recurrence, death of disease (DOD) that is defined as the date from the first cytoreductive surgery to the date of the patient's death due to ovarian cancer, preoperative leukocyte count and proportion (within 90 days before the operation).

Recurrence was defined as the time from the first cytoreductive surgery to the time of ovarian cancer recurrence; death was defined as the date from the first cytoreductive surgery to the date of the patient's death due to ovarian cancer. The “normal”/control group selected and consisted of patients of a similar age who were diagnosed with uterine prolapse. The results of preoperative routine blood tests for these patients were also collected. Ethics approval for this research was provided by the Beijing Chaoyang Hospital, Capital Medical University (approval number 2021-ke-205, study number 2012DRF30490).

For survival prediction, we choose the machine learning-based decision tree algorithm. We divided the method into several steps, as shown in the flowchart (Fig. 1). We implemented the algorithm in Python 3.8 and the scikit-learn 0.24 package. In the preprocessing part, we first removed the patients’ identifying information. Second, we analyzed the distribution and removed the independent discrete points that were out of the value range of 5–95%. Third, we selected the features, such as stage, grade of differentiation and LNM, according to the National Comprehensive Cancer Network (NCCN) guideline [26]. Then, using tenfold cross-validation, we separated the data into two parts: 90% of the data was used for training and 10% of the data was used for testing. Imbalanced datasets are often handled well by decision tree classifiers [27], so we built a decision tree model and trained the model. Finally, to ensure the stability of the model, we used tenfold cross-validation to train and test the model. To reduce the influence of imbalanced data, we used the synthetic minority oversampling technique (SMOTE) method to oversample the training set, which is an improved scheme based on a random oversampling algorithm [28]. To prove the role of circulating leukocytes in survival prediction, we performed comparisons with the same model trained by the features without circulating leukocytes.

For the decision tree learning process, these patients (93 patients) who had same first chemotherapy regimen were included. The optimal feature was selected recursively, and the training data were segmented according to the feature so that each subdataset had the best classification process. This process corresponded to the division of the feature space and the construction of the decision tree. First, the root node was constructed, and all training data were placed in the root node. An optimal feature was chosen, and the training dataset was divided into subsets according to this feature so that each subset had the best classification under the current conditions. If these subsets could be relatively correctly classified, then the leaf nodes were constructed, and these subsets were divided into the corresponding leaf nodes; if there were still subsets that could not be relatively correctly classified, then these subsets selected the new optimal feature, continued to divide it, and constructed the corresponding node. This process proceeded recursively until all the training data subsets were basically correctly classified or there were no suitable features. Finally, each subset was assigned to the leaf nodes.

A total of 98 patients with ovarian serous carcinoma who were initially treated at the Department of Obstetrics and Gynecology at Beijing Chaoyang Hospital, Capital Medical University, from July 2009 to December 2018 were included in the analysis according to the inclusion and exclusion criteria. The first chemotherapy regimen after the first surgery for all selected patients was platinum-based treatment (93 patients received 6–8 cycles of paclitaxel and cisplatin (PT), 3 patients received 8 cycles of cisplatin + adriamycin + cyclophosphamide (PAC), 1 patient received 8 cycles of paclitaxel and carboplatin, and 1 patient received 4 cycles of PT and 2 cycles of cisplatin + etoposide + ifosfamide (PEI)). The average age was 57 years old, and the mean BMI was 24.3. The pathological results revealed that 88.60% of the patients had poorly differentiated tumors (G3), 79.60% had stage III disease, 66.3% had positive ascites cytology, and 43.2% had LNM. The recurrence and mortality rates were 55.3% and 29.7%, respectively, at the time of follow-up (28 July 2019).

The preoperative monocyte count and proportion in the serous ovarian cancer group (98 patients) were significantly higher than those in the control group (75 patients, p < 0.001 and p < 0.001, respectively, Table 1 and Fig. 2). The monocyte-to-lymphocyte (MO/LY) ratio in the serous ovarian cancer group was also significantly higher than that in the normal control group (p < 0.001). The number of white blood cells (WBCs, p = 0.022), lymphocytes (LYs, p < 0.001), neutrophils (NEs, p < 0.001), and eosinophils (EOs, p < 0.001) were also significantly different between the serous ovarian cancer group and the normal control group.

As shown in Table 2, the percentage of monocytes showed significant differences across the different disease stages (p = 0.046); the more advanced the stage was, the higher the average percentage. The monocyte counts also showed similar results, with patients with LNM having more monocytes (p = 0.05). The MO/LY ratio showed significant differences according to differentiation status (p = 0.029), stage (p = 0.007), LNM (p = 0.025), and recurrence (p = 0.036), with a higher ratio indicating a worse result, similar to the results for CA125. In addition, the number of NE (p = 0.049) and BA (p = 0.011) and the percentage of LY (LY%, p = 0.036) affected LNM. LY% (p = 0.048 and p = 0.015, respectively) and NE% (p = 0.027 and p = 0.028, respectively) were significantly correlated with positive ascites cytology and recurrence.

After dividing serous ovarian carcinoma patients into two groups based on the median value, OS and PFS decreased slightly faster in the group with a higher monocyte count (Kaplan–Meier analysis, Fig. 3). A higher MO/LY ratio was significantly correlated with shorter PFS (p = 0.001) and OS (p = 0.048), which was similar to the results for CA125 (p = 0.020 and < 0.001, respectively).

In addition, a higher NE (p = 0.029 and 0.014, respectively) and NE% (p = 0.011 and 0.004, respectively) significantly predicted shorter OS and PFS times. In contrast, the lower the LY% was (p = 0.021 and < 0.001, respectively), the worse the prognosis.

When assessing death and recurrence according to the tertiles of the MO/LY ratio cross-classified by the tertiles of the CA125 level, both the death rate and recurrence rate increased across the increasing tertiles of the MO/LY ratio for the first and second tertiles of the CA125 level (Fig. 4). All patients within the third tertile of the CA125 level belonged to the first tertile of the MO/LY ratio. Therefore, no cases showed a high MO/LY ratio and high CA125 level at the same time.

For the decision tree, CA125 was found to be the root node with the largest information gain by using the built-in method of sklearn (Fig. 5). The Gini coefficient reflects the measure of data uncertainty. The smaller the Gini value is, the higher the purity of the potential classes. In each node, the sample number shows the number of samples before being divided, and the value means the number belongs to each class. For example, in the root node, the total number of samples is 42, so the samples are 42. According to whether the CA125 attribute was less than or equal to 3726.05, the samples were split into two groups that contained 35 and 7 samples, respectively.

For prediction processing, at the root node, the sample was divided into two groups based on a CA125 value less than or equal to 3726.05. Then, the divided samples were judged by the second layer leaf node. In the second layer leaf nodes, the value WBC or LNM are the standards of classification. After that, the samples will go through into the third layer of leaf nodes until there is no leaf node left. Finally, when the decision reaches the last leaf node, the survival probability is the number of class samples divided by the total samples in the node. For example, at the leftmost leaf node, the probability of survival is 5/6, and the probability of death is 1/6.

The features involved in the decision tree were the MO/LY, differentiation status, CA125 level, NE, ascites cytology, LY% and age. The survival prediction AUC of the decision tree was 0.69 (95% CI: 0.67–0.70).Meanwhile, the survival prediction AUC of the logsitic regression (LR) was 0.55(95% CI: 0.53–0.57), which means that the performance of decision tree is much bether.The performance cooperation between the model trained by the features with (blue line) and without the routine blood test (RT) (red line) results is shown in Fig. 6. From Fig. 6, it is obvious that the model trained with RT has better performance than that without RT. The feature importance in the decision tree is shown in Fig. 7. The MO/LY, differentiation, CA125, NE, ascites cytology, LY% and age had a high impact, and WBC, MO and LNM had a low impact on the model.