Personalized Medicine Using Predictive Analytics: A Machine Learning-Based Prognostic Model for Patients Undergoing Hip Arthroscopy

Abstract

Background:

Personalized medicine models to predict outcomes of orthopaedic surgery are scarce. Many have required data that are only available postoperatively, mitigating their usefulness in preoperative decision making.

Purpose:

To establish a method for predictive modeling to enable individualized prognostication and shared decision making based on preoperative patient factors using data from a prospective hip preservation registry.

Study Design:

Cohort study (Prognosis); Level of evidence, 2.

Methods:

Preoperative data of 2415 patients undergoing hip arthroscopy for femoroacetabular impingement syndrome between February 2008 and November 2017 were retrospectively analyzed. Two machine-learning analyses were evaluated: Tree-structured survival analysis (TSSA) and Cox proportional hazards modeling for predicting time to event and for computing hazard ratios for survivorship after hip arthroscopy. The Fine-Gray model was similarly used for repeat hip arthroscopy. Variables were selected for inclusion using the minimum Akaike Information Criterion (AIC). The stepwise selection was used for the Cox and Fine-Gray models. A web-based calculator was created based on the final models.

Results:

Prognostic models were successfully created using Cox proportional hazards modeling and Fine-Gray models for survivorship and repeat hip arthroscopy, respectively. The Harrell C-statistics of the Cox model calculators for survivorship after hip arthroscopy and the Fine-Gray model for repeat hip arthroscopy were 0.848 and 0.662, respectively. Using the AIC, 13 preoperative variables were identified as predictors of survivorship, and 6 variables were identified as predictors for repeat hip arthroscopy. In contrast, the TSSA model performed poorly, resulting in a Harrell C-statistic <0.6, rendering it inaccurate and not interpretable. A web-based calculator was created based on these models.

Conclusion:

This study successfully created an institution-specific machine learning–based prognostic model for predictive analytics in patients undergoing hip arthroscopy. Prognostic models to predict survivorship and the need for repeat surgeries were both adapted into web-based tools to assist the physician with shared decision making. This prognostic model may be useful at other institutions after performing external validation. Additionally, this study may serve as proof of concept for a methodology for the development of patient-specific prognostic models. This methodology may be used to create other predictive analytics models in different realms of orthopaedic surgery, contributing to the evolution from evidence-based medicine to personalized medicine.