Parkinson’s disease (PD) is a progressive neurodegenerative disorder characterized by motor symptoms, like tremors, rigidity and more affecting handwriting. This study explores the application of ensemble deep learning models to enhance early identification of PD through handwriting analysis. Archimedes spiral drawings from 54 PD patients and age and sex-matched healthy controls were used. Multiple Convolutional Neural Network architectures were considered: ResNet50, EfficientNet, DenseNet, and MobileNet, combined through an ensemble method to improve accuracy and robustness. The DenseNet architecture achieved the highest performance with an AUC of 0.9762 and an F1 score of 0.9756. Activation maps generated using the Grad-CAM mechanism provided visual explanations, enhancing the model’s interpretability and aligning with clinical observations. Statistical analysis of the activation maps images using Chi-Square, Shapiro-Wilk, and Mann–Whitney U tests confirmed significant differences between PD and healthy controls. The results demonstrate the potential of ensemble CNN models for non-invasive and effective identification of PD. This offers a reliable and interpretable tool for early diagnosis and patient follow-up.