# Simulated depression risk classification from Parkinson’s voice features using a self-attention-enhanced MLP architecture

## TL;DR

This paper introduces a new model that uses voice features from Parkinson’s patients to predict depression risk with high accuracy.

## Contribution

A novel self-attention-enhanced MLP architecture is proposed for depression risk classification using vocal biomarkers in Parkinson’s disease.

## Key findings

- The model achieves 97% accuracy, 98% F1-score, 95% recall, and 100% specificity in depression risk classification.
- Voice features like Harmonic-to-Noise Ratio and Jitter are effective biomarkers for depression risk prediction in Parkinson’s patients.

## Abstract

Parkinson’s disease affects both motor and non-motor functions, including vocal features that may indicate underlying mental health conditions such as depression. This work proposes a novel framework for simulated depression risk classification using vocal biomarkers derived from the UCI Parkinson’s dataset. A Self-Attention-Enhanced Multilayer Perceptron-MLP architecture is used model interactions between key acoustic features, particularly Harmonic-to-Noise Ratio and Jitter, which serve as the basis for generating binary depression risk labels. The proposed model outperforming traditional and deep learning benchmarks including Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), TabNet, CNN-LSTM, Deep Neural Network (DNN), and Explainable Boosting Machine (EBM) with an accuracy of 97%, F1-score of 98%, recall of 95%, and specificity of 100%, While EBM offers strong interpretability, the attention-enhanced model demonstrates optimal predictive capability. These findings highlight the efficacy of voice-based features combined with attention mechanisms for early, non-invasive identification of depression risk in PD patients.