Using machine learning to estimate time-to-progression in ovarian cancer based on cancer registry and administrative health data.

BACKGROUND

Cancer progression and recurrence are not routinely registered in Belgium. A machine-learning algorithm was developed using administrative data to retrospectively estimate time-to-progression (including recurrence) of epithelial ovarian cancer. The algorithm was validated against a multicentric clinical reference dataset.

PATIENTS AND

METHODS

From the Belgian Cancer Registry, we extracted patients diagnosed with epithelial ovarian cancer during 2014-2018 and treated for their primary tumor in seven participating centers in Belgium. The reference dataset was constructed through a manual review of hospital records for the selected patients. Using a unique personal identification number, the cancer registry data and administrative records were linked, including reimbursed medical procedures and medication, vital status, and causes of death. The reference dataset was split into a training set and a test set. To estimate time-to-progression, we developed and validated a machine-learning algorithm (XGBoost classifier), framing the task as a discrete-time classification problem at a patient-month level.

RESULTS

Of the 588 patients in the reference dataset, 61.6% (stage IV: 73.2%) progressed during a median follow-up of 52.5 months (interquartile range: 36.3 months). The algorithm achieved excellent discrimination between pre- and postprogression months (area under the receiving operators curve: 0.99) in the held-out test set, translating into high patient-level classification performance for identifying progression [sensitivity: 98% (95% confidence interval: 96-100), specificity: 91% (84-97), positive predictive value: 95% (90-98)] and a mean absolute timing error of 1.5 months for true positive cases.

CONCLUSIONS

In the absence of systematic registration of progression in Belgium, automated data mining of population-level administrative records offers a valuable alternative to estimate the timing of ovarian cancer progression.