Evaluation of the effects of deep brain neurostimulation in parkinson's patients, analysis of MRI images via artificial intelligence and its use for early diagnosis

Abstract

Aim: This study aimed to evaluate the early diagnostic potential of subthalamic nucleus (STN) radiomic features and machine learning algorithms to differentiate Parkinson's disease patients requiring Deep Brain Stimulation (DBS) from those managed solely with medical therapy. This approach seeks to provide a non-invasive imaging biomarker to assist in objective surgical decision-making. Methods: A retrospective analysis of 394 Parkinson’s disease patients (200 in the DBS surgical cohort and 194 in the medical treatment cohort) was conducted using preoperative MRI scans. Following meticulous manual segmentation of the STN, a comprehensive set of radiomic features—including shape, first-order statistics, and Gray Level Co-occurrence Matrix (GLCM) parameters—was extracted. A Random Forest classifier was trained on 70% of the dataset and validated on a 30% independent test set to classify the patients based on their microstructural STN signatures. Results: The Random Forest machine learning model achieved an exceptional overall classification accuracy of 93.10% on the test dataset. Receiver Operating Characteristic (ROC) analysis yielded an Area Under the Curve (AUC) of up to 0.938. The performance was perfectly unbiased, with precision, recall, and F1-scores reaching 0.93 for both the surgical and medical cohorts. Feature importance analysis revealed that morphological deformations, specifically STN flatness, and first-order signal intensity drops associated with iron deposition were the most highly discriminative biophysical biomarkers for identifying surgical candidates. Conclusion: The application of machine learning to STN radiomic features provides a highly accurate, non-invasive, and objective methodology to distinguish DBS candidates from medically managed Parkinson's patients. While these findings are promising, prospective and external validation is required before the model can be fully incorporated into routine clinical decision-making.