AUTHOR=Peng Li , Lan Lan , Xiu Peng , Zhang Guangming , Hu Bowen , Yang Xi , Song Yueming , Yang Xiaoyan , Gu Yonghong , Yang Rui , Zhou Xiaobo TITLE=Prediction of Proximal Junctional Kyphosis After Posterior Scoliosis Surgery With Machine Learning in the Lenke 5 Adolescent Idiopathic Scoliosis Patient JOURNAL=Frontiers in Bioengineering and Biotechnology VOLUME=8 YEAR=2020 URL=https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.2020.559387 DOI=10.3389/fbioe.2020.559387 ISSN=2296-4185 ABSTRACT=Objective

To build a model for proximal junctional kyphosis (PJK) prognostication in Lenke 5 adolescent idiopathic scoliosis (AIS) patients undergoing long posterior instrumentation and fusion surgery by machine learning and analyze the risk factors for PJK.

Materials and Methods

In total, 44 AIS patients (female/male: 34/10; PJK/non-PJK: 34/10) who met the inclusion criteria between January 2013 and December 2018 were retrospectively recruited from West China Hospital. Thirty-seven clinical and radiological features were acquired by two independent investigators. Univariate analyses between PJK and non-PJK groups were carried out. Twelve models were built by using four types of machine learning algorithms in conjunction with two oversampling methods [the synthetic minority technique (SMOTE) and random oversampling]. Area under the receiver operating characteristic curve (AUC) was used for model discrimination, and the clinical utility was evaluated by using F1 score and accuracy. The risk factors were simultaneously analyzed by a Cox regression and machine learning.

Results

Statistical differences between PJK and non-PJK groups were as follows: gender (p = 0.001), preoperative factors [thoracic kyphosis (p = 0.03), T1 slope angle (T1S, p = 0.078)], and postoperative factors [T1S (p = 0.097), proximal junctional angle (p = 0.003), upper instrumented vertebra (UIV) – UIV + 1 (p = 0.001)]. Random forest using SMOTE achieved the best prediction performance with AUC = 0.944, accuracy = 0.909, and F1 score = 0.667 on independent testing dataset. Cox model revealed that male gender and larger preoperative T1S were independent prognostic factors of PJK (odds ratio = 10.701 and 57.074, respectively). Gender was also at the first place in the importance ranking of the model with best performance.

Conclusion

The random forest using SMOTE model has the great value for predicting the individual risk of developing PJK after long instrumentation and fusion surgery in Lenke 5 AIS patients. Moreover, the combination of the outcomes of a Cox model and the feature ranking extracted by machine learning is more valuable than any one alone, especially in the interpretation of risk factors.