High-Grade Serous Ovarian Cancer: Use of Machine Learning to Predict Abdominopelvic Recurrence on CT on the Basis of Serial Cancer Antigen 125 Levels

Abstract

Purpose

The aim of this study was to use machine learning to predict abdominal recurrence on CT on the basis of serial cancer antigen 125 (CA125) levels in patients with advanced high-grade serous ovarian cancer on surveillance.

Methods

This institutional review board–approved, HIPAA-compliant, retrospective, hypothesis-generating study included all 57 patients (mean age, 61 ± 11.2 years) with advanced high-grade serous ovarian cancer who underwent cytoreductive surgery from January to December 2012, followed by surveillance abdominopelvic CT and corresponding CA125 levels. A blinded radiologist reviewed abdominopelvic CT studies until recurrence was noted. Four measures of CA125 were assessed: actual CA125 levels at the time of CT, absolute change since prior CT, relative change since prior CT, and rate of change since prior CT. Using machine learning, support vector machine models were optimized and evaluated using 10-fold cross-validation to determine the CA125 measure most predictive of abdominal recurrence. The association of the most accurate CA125 measure was further analyzed using Cox proportional-hazards model along with age, tumor size, stage, and degree of cytoreduction.

Results

Rate of change in CA125 was most predictive of abdominal recurrence in a linear kernel support vector machine model and was significantly higher preceding CT studies showing abdominal recurrence (median 13.2 versus 0.6 units/month; P = .007). On multivariate analysis, a higher rate of CA125 increase was significantly associated with recurrence (hazard ratio, 1.02 per 10 units change; 95% confidence interval, 1.0006-1.04; P = .04).

Conclusion

A higher rate of CA125 increase is associated with abdominal recurrence. The rate of increase of CA125 may help in the selection of patients who are most likely to benefit from abdominopelvic CT in surveillance of ovarian cancer.

Introduction

Ovarian cancer is the most common gynecologic malignancy associated with patient mortality and the fifth most common cause of cancer mortality in women. An estimated 22,280 new cases and 14,240 deaths related to ovarian cancer occurred in 2016 in the United States alone [1]. The majority of patients with ovarian cancer present with advanced disease beyond the pelvis; at present, primary cytoreductive surgery is the preferred treatment offering the best chance of disease-free survival 2, 3, 4, 5, 6. Given advances in treatment, the median survival of patients with ovarian cancer has increased from approximately 20 months to more than 65 months over the past few decades [7]. As a result, these patients are followed for a longer duration, often for years.

National Comprehensive Care Network (NCCN) guidelines recommend that beyond initial therapy, these patients should have follow-up visits every 2 to 4 months for the first 2 years, then every 6 months for the next 3 years, and then annually for 5 years. At these visits, use of cancer antigen 125 (CA125) levels has been recommended, along with surveillance imaging studies using chest, abdominal, and pelvic CT, MRI, or PET/CT as clinically indicated [5]. Although CA125 may not be elevated in about 20% of patients with ovarian cancer and may occasionally be elevated with unrelated conditions such as endometriosis, liver cirrhosis, or peritonitis, it remains the most robust and widely used biomarker for ovarian cancer [8].

Although PET/CT is reportedly more accurate in the detection of metastatic disease in patients with ovarian cancer, CT is currently the most frequently used imaging modality for follow-up of these patients in our practice at a large academic cancer center. Imaging is also commonly used as a basis for treatment planning, inclusion of patients in clinical trials, and response assessment. Imaging surveillance allows the detection of asymptomatic disease recurrence, which has been reported to be associated with improved overall survival [9]. However, other studies have found variable (40%-93%) sensitivity of CT in the detection of disease recurrence 7, 10. Furthermore, because the NCCN guidelines recommend the use of imaging “as clinically indicated,” there is a high potential for inconsistent use of imaging for these patients, leading to unwarranted variability in patient care. Indeed, a large multicenter study by Esselen et al [11] reported that CA125 and CT are routinely used for surveillance of patients with ovarian cancer, adding significant costs without proven benefit. This study also found a high utilization rate and significant variation in the use of CT and indicated that this may be a reasonable target for improving value. Therefore, there is a clear need to improve imaging strategies to make them more streamlined, consistent, and high yield while avoiding overutilization and consequent costs and radiation burden.

Because CA125 is still frequently obtained in these patients in accordance with the NCCN guidelines 5, 11, it would be helpful to be able to make imaging-related decisions on the basis of CA125 levels. It is known that CA125 levels are frequently elevated before clinical detection of recurrence [12]; however, the exact relationship between CA125 levels and detection of disease on imaging is unknown. We hypothesize that CA125 levels can be used to select patients with ovarian cancer undergoing surveillance who are likely to benefit from abdominal CT. We have previously demonstrated that thoracic metastases typically occur in patients with existing abdominal disease 13, 14; therefore, we focused this study on abdominal disease. The purpose of this study was to use machine learning to predict abdominal recurrence on CT on the basis of serial CA125 levels in patients with advanced ovarian cancer on surveillance.