Can We Predict Capsular Contracture? Exploring the Answer with Machine Learning

Goals/Purpose: Capsular contracture (CC) is a frequent and distressing complication of breast augmentation and reconstruction, yet its risk factors and mechanism are poorly understood. Predicting which patients develop CC is difficult because it is dependent on a variety of factors that differ from patient to patient. Therefore, this study aims to determine which patients undergoing breast reconstruction will develop CC by evaluating patient specific features with a modern machine learning algorithm.

Methods/Technique: Patients were queried from the 2003-2017 IBM MarketScan Databases with Common Procedural Terminology (CPT) codes for implant-based breast reconstruction and augmentation. International Classification of Disease (ICD) and CPT codes were then used to identify all events and conditions of a patient’s medical and procedural history, as well as whether each patient suffered CC of any severity. Next, a series of hyperparameter-tuned random forests was fit. The random forest machine learning algorithm leverages ensembles of classification trees. It is a sophisticated statistical model commonly employed to select several important single nucleotide polymorphisms out of several million in large genome wide association studies. Random forests have successfully been employed to determine single nucleotide polymorphisms that increase risk for several complex diseases. The out of bag classification accuracy, permutation importance, and metrics of significance were computed for the final model. To create a semi-parsimonious model, only diagnoses and procedures occurring in at least 1,000 patients were excluded. Finally, a multivariate logistic regression model with features of previous capsular contracture and prior irradiation was developed to validate known risk factors for the development of CC that have been established in previous literature.

Results/Complications: A total of 112,489 patients were included, with 4,825 common conditions and procedures captured in the model. The rate of CC in the study cohort was 9.55%. The random forest’s five-fold cross validated error rate was 9.54%. Therefore, an algorithm with knowledge of the entire patient past medical and procedural history only improved prediction capacity by 0.01%. An example of an effective predictive model is compared the model developed in this study in Table One. Prior diagnosis of CC was the most important variable in the model. When removed, model accuracy decreased by 0.0011% (P<0.001). Removing prior irradiation as a feature decreased model accuracy by 0.00014% (P<0.001). In line with previous literature, prior CC (OR=2.47, P<0.001) and irradiation (OR=1.16, P<0.001) were significantly associated with CC development in a multivariate logistic regression model.

Conclusion: This study aimed to answer if CC can be predicted. The findings suggest that the medical history, as encoded by ICD and CPT codes, is not a valuable informant of CC development. While previous associations between CC and aspects of the patient history were confirmed, their clinical utility is debatable. These associations do not seem to improve a clinician’s ability to reliably predict whether a patient is at increased risk for CC development or will develop CC. Future analysis will evaluate the influence of antibiotic irrigation, acellular dermal matric placement, and leukotriene inhibitor administration on capsular contracture development in this cohort. Genetic factors should also be explored to determine any possible links.