Compressed Attentional Representation Learning For Disease Prediction

Parkinson’s disease (PD) is a neurodegenerative disorder of motor control with early vocal impairments, offering a critical pathway for early diagnosis by means of voice analysis. However, the diagnostic models proposed so far often result in low accuracy and precision, thus limiting their clinical utility. This study presents a cutting-edge framework integrating advanced feature extraction, compression, and attention mechanisms, achieving remarkable accuracy for Parkinson’s disease diagnosis.Our feature extraction process adopts various methods, including ing time-frequency analysis, Mel-frequency cepstral coefficients (MFCC), wavelet transforms, vocal fold features, and tunable Qfactor wavelet transform (TQWT), which capture comprehensive vocal characteristics indicative of PD. Optimal feature selection is done using the mRMR method and the features are compressed using autoencoders, which is further fed into an Multi-Head Attention (MHA) mechanism that learns complex dependencies within these compressed features. Of the four configurations evaluated, Scenario 3, where autoencoder compression is followed by an MHA mechanism operating on these compressed features, it yields the highest accuracy of 93% and precision of 97%, surpassing previous state-of-the-art methods. This integrated approach provides a new benchmark in the accuracy and precision of PD detection and may be considered as a promising tool for early diagnosis and follow-up in telemedicine applications.