An Explainable Machine Learning Framework for Prediction of Recurrent Lumbar Disc Herniation

The term “Recurrent Lumbar Disc Herniation (rLDH)” describes the recurrence of a herniated disc following a prior surgical procedure in the lumbar (lower) area of the spine. Even after prior surgery, recurrent herniation can still happen and can affect the same or a different disc. Conservative measures and, in certain situations, revision surgery may be used to treat recurrent lumbar disc herniation. It is key to look for causes for the repeat occurrence and possibly take precautions in avoiding a rLDH. Artificial Intelligence (AI) methods are rapidly being adopted for risk assessment and prediction of outcomes of medical events in medicine. Improving the interpretability and transparency of machine learning (ML) models is a fundamental goal of applying artificial intelligence (AI) techniques in the healthcare sector. Health care looks for explanations and for the importance of each feature in the model’s decision-making process while dealing with medical issues. Explainable AI (XAI) is a phenomenon that arises from the use of AI in the healthcare business through several tactics such as interpretation, fairness, validation, and performance assessment of the model. In the current study, an XAI framework is attempted to identify risk factors and recommendations to assess the recurrence of LDH. The potential predictive value of whole lumbar spine (WLS) morphometry for rLDH has not been investigated to date according to the authors’ knowledge. The framework utilizes conformal analysis for prediction and performs combined pattern association analysis to extract the predictive factors that are significant against rLDH. In the ML analysis, extreme gradient boosting (XGB) performed best on the evaluation metrics of Accuracy of 0.95 and F-score of 0.75. Prediction further was proved significant with conformal analysis giving 85% accuracy over the prediction. Factors such as ‘age’, ‘modic end plate’, ‘saggital disc height’, and ‘disc level’ identified by combined pattern association mining are significant in identifying rLDH. Identification of practical and measurable factors by the ML framework compared to traditional models establishes the advantage and usability of ML models in healthcare scenarios. Thus results prove that the use of the proposed ML framework with careful analysis provides a clinically relevant and meaningful tool in the pathogenesis of rLDH.