Drawing-Aware Parkinson's Disease Detection Through Hierarchical Deep Learning Models

Parkinson's disease (PD) is a chronic neurological disorder that progresses slowly and shares symptoms with other diseases. Early detection and diagnosis are vital for appropriate treatment through medication and/or occupational therapy, ensuring patients can lead productive and healthy lives. Key symptoms of PD include tremors, muscle rigidity, slow movement, and balance issues, along with psychiatric ones. Handwriting (HW) dynamics have been a prominent tool for detecting and assessing PD-associated symptoms. Still, many handcrafted feature extraction techniques suffer from low accuracy, which is rather than optimal for diagnosing such a serious condition. To that end, various machine learning (ML) and deep learning (DL) approaches have been explored for early detection. Meanwhile, concerning the latter, large models that introduce complex and difficult-to-understand architectures reduce the system's recognition transparency and efficiency in terms of complexity and reliability. To tackle the above problem, an efficient hierarchical scheme based on simpler DL models is proposed for early PD detection. This way, we deliver a more transparent and efficient solution for PD detection from HW records. At the same time, we conclude that a careful implementation of each component of the introduced hierarchical pipeline enhances recognition rates. A rigorous 5-fold cross-validation strategy is adopted for evaluation, indicating our system's robust behavior under different testing scenarios. By directly comparing it against a similar end-to-end classifier, the benefits of our technique are clearly illustrated during experiments. Finally, its performance is compared against several state-of-the-art ML- and DL-based PD detection methods, demonstrating the method's supremacy.