Four randomized trials concluded that overall survival for patients with advanced ovarian cancer is similar regardless of whether they receive neoadjuvant chemotherapy (NACT) or primary cytoreductive surgery [[1], [2], [3], [4], [5]]. However, some observational studies have found that NACT is associated with inferior survival [[6], [7], [8], [9], [10]]. Furthermore, patients who achieve complete cytoreduction before receiving any chemotherapy have a particularly favorable prognosis [[1], [2], [3], [4], [5], [6], [7], [8], [9], [10]].

The survival advantage observed among patients who undergo primary cytoreductive surgery in routine practice could be affected by baseline health status, which may influence both treatment selection and overall survival [11]. Comorbidities and poor functional status impart a poor prognosis to patients, irrespective of treatment strategy, and are often considered an indication for NACT. Studies of ovarian cancer based on administrative data sources, like health insurance claims or hospital discharge records, frequently adjust for confounding due to baseline health status using the Charlson Comorbidity Index (CCI) [12], a score calculated by identifying the presence of select comorbidities present before the cancer diagnosis.

For data sources based on International Classification of Disease (ICD) codes, the CCI considers only limited conditions that may appear in a patient's health claims history. If the diagnosis and procedure codes ignored by the CCI are informative about health status, using CCI, and similar methods [13] to control for health status may lead to residual confounding. Machine learning methods may allow for a richer representation of health status in observational studies by making more effective use of information encoded in patients' claims histories, and aid in comparative effectiveness studies in which health status is an important confounder.

In this study, we hypothesize that the CCI is a suboptimal method for representing baseline health status in studies that compare survival after NACT versus primary cytoreductive surgery among patients with advanced ovarian cancer. Specifically, we hypothesize that compared with the CCI, a machine learning-based comorbidity index will more accurately identify advanced ovarian cancer patients at risk for early mortality and will reduce bias in observational studies seeking to estimate the effectiveness of first-line treatments in this population. To this end, we have developed and validated a multidimensional comorbidity index (MCI) using partial least squares (PLS) regression, a supervised machine learning algorithm, and compared the predictive performance of this index with the CCI on 1-year mortality among patients with advanced ovarian cancer stage. Furthermore, we have evaluated the performance of the MCI in reducing confounding bias in a comparative effectiveness analysis comparing NACT with primary cytoreductive surgery on long-term all-cause mortality among Medicare beneficiaries with advanced ovarian cancer.