A multi-layered stacked classifier developed for diagnosis of Parkinson’s disease and SWEDD patients using fusion of multimodal data

Early detection of Parkinson’s disease (PD) is difficult due to overlapping with the common symptoms of other neuro-disorders. One of the most prominent of these related diseases is SWEDD (scans without evidence of dopamine deficit), which is considered clinically similar to PD and also has normal dopamine transporter scans. Therefore, there is a pressing need for a reliable method for distinguishing PD from SWEDD and related disorders. To handle this problem, the association between PD and SWEDD has been explored using the fusion of features based on multimodal data (biological, clinical, and imaging). First, the data is normalized by implementing the min–max normalization. Subsequently, feature selection and data-balancing strategies are applied to select optimal features and overcome the data imbalance issue. In addition, a multi-layered stacking (MULS) classifier of three layers is proposed for classification. Also, Bayesian optimization and 5-fold nested stratified cross-validation for hyperparameter tuning are applied on each layer of the MULS classifier. The performance of the developed classifier is estimated using the best feature set against three binary classifications. From the outcomes, it has been observed that the MULS classifier achieved better results for classification between PD and SWEDD compared to the methods in the literature. The results yielded are 97.38% accuracy, 96.21% f1-score, 98.78% sensitivity, 98.47% precision, and 98.21% area under the curve. Furthermore, the impact of multimodal fusion features is analyzed, and also the proposed model is validated with the independent datasets. Hence, the suggested method is believed to help healthcare professionals analyze diseases early.