Machine learning for automated differentiation of parkinson’s disease and its mimics using ¹²³I-mIBG scintigraphy: insights from a multicentre real-world cohort (ITA-mIBG study)

Purpose: To assess machine learning (ML) classifiers trained on harmonised multicentre ¹²³I-mIBG planar scintigraphy for differentiating Parkinson’s disease (PD) from non-PD parkinsonian syndromes and to determine whether early imaging alone may ensure accurate discrimination. Methods: This retrospective study included patients with suspected PD who underwent early (~ 15 min) and delayed (~ 240 min) imaging and received a definitive diagnosis after ≥ 12 months. Harmonised region of interest (ROI) placement and ComBat correction were applied. Early and late heart-to-mediastinum (H/M) ratios and washout rate (WR) were calculated. Differences were tested by Mann-Whitney U test, and cut-points identified by ROC analysis. Logistic regression, Gaussian naïve Bayes, and support vector machine were trained on these features with Z-score normalisation and synthetic minority oversampling technique (SMOTE). Results: 127 patients were analysed (85 PD, 42 non-PD). Early and late H/M ratios were significantly lower in PD than non-PD (early H/M 1.45 ± 0.20 vs. 1.80 ± 0.20; late H/M 1.33 ± 0.22 vs. 1.68 ± 0.21; both p < 0.001). WR was modestly higher in PD (8.74 ± 5.76% vs. 6.49 ± 6.19%, p = 0.024). Optimal cut-points for PD were: early H/M ≤ 1.62 (accuracy 80.3%, sensitivity 83.3%, specificity 78.8%, and AUC 0.878), late H/M ≤ 1.52 (83.5%, 83.3%, 83.5% and 0.866) and WR ≥ 6.03% (70.1%, 70.6%, 69.0% and 0.645). ML achieved mean accuracy 78.9–80.7%, sensitivity 81.9–84.0%, specificity 68.6–78.0%, and AUC 0.850–0.875. Conclusion: Classifiers trained on ¹²³I-mIBG semi-quantitative indices accurately distinguished PD from non-PD. Early H/M ratio alone provided excellent discrimination, supporting early-imaging; prospective validation is warranted.