Prognosis of hip osteonecrosis after cell therapy with a calculator and artificial intelligence: ten year collapse-free survival prediction on three thousand and twenty one hips.

PURPOSE

Several reports have identified prognostic factors for hip osteonecrosis treated with cell therapy, but no study investigated the accuracy of artificial intelligence method such as machine learning and artificial neural network (ANN) to predict the efficiency of the treatment. We determined the benefit of cell therapy compared with core decompression or natural evolution, and developed machine-learning algorithms for predicting ten year collapse-free survival in hip osteonecrosis treated with cell therapy. Using the best algorithm, we propose a calculator for "prognosis hip osteonecrosis cell therapy (PHOCT)" accessible for clinical use.

METHODS

A total of 3145 patients with 5261 osteonecroses without collapses were included in this study, comprising 1321 (42%) men and 1824 (58%) women, with a median age of 34 (12-62) years. Cell therapy was the treatment for 3021 hips, core decompression alone for 1374 hips, while absence of treatment was the control group of 764 hips. First, logistic regression and binary logistic regression analysis were performed to compare results of the three groups at ten years. Then an artificial neural network model was developed for ten year collapse-free survival after cell therapy. The models' performances were compared. The algorithms were assessed by calibration, and performance, and with c-statistic as measure of discrimination. It ranges from 0.5 to 1.0, with 1.0 being perfect discrimination and 0.5 poor (no better than chance at making a prediction).

RESULTS

Among the 3021 hips with cell therapy, 1964 hips (65%) were collapse-free survival at ten years, versus 453 (33%) among those 1374 treated with core decompression alone, and versus 115 (15%) among 764 hips with natural evolution. We analyzed factors influencing the prediction of collapse-free period with classical statistics and artificial intelligence among hips with cell therapy. After selecting variables, a machine learning algorithm created a prognosis osteonecrosis cell therapy calculator (POCT). This calculator proved to have good accuracy on validation in these series of 3021 hip osteonecroses treated with cell therapy. The algorithm had a c-statistic of 0.871 suggesting good-to-excellent discrimination when all the osteonecroses were mixed. The c-statistics were calculated separately for subpopulations of categorical osteonecroses. It retained good accuracy, but underestimated ten year survival in some subgroups, suggesting that specific calculators could be useful for some subgroups. This study highlights the importance of multimodal evaluation of patient parameters and shows the degree to which the outcome is modified by some decisions that are within a surgeon's control, as the number of cells to aspirate, the choice of injecting in both the osteonecrosis and the healthy bone, the choice between unilateral or bilateral injection, and the possibility to do a repeat injection.

CONCLUSION

Many disease conditions and the heterogeneities of patients are causes of variation of outcome after cell therapy for osteonecrosis. Predicting therapeutic effectiveness with a calculator allows a good discrimination to target patients who are most likely to benefit from this intervention.