[Paper Review] PCA-RF: An Efficient Parkinson's Disease Prediction Model based on Random Forest Classification

Ishu Gupta, Vartika Sharma+2 more|arXiv (Cornell University)|Mar 21, 2022

Parkinson's Disease Mechanisms and TreatmentsMedicine30 citations

TL;DR

The paper proposes PCA-RF, a Parkinson’s disease prediction model that uses PCA for feature reduction followed by a Random Forest classifier, achieving up to ~90% accuracy without PCA and ~77% with PCA on the UCI Parkinson’s dataset.

ABSTRACT

In this modern era of overpopulation disease prediction is a crucial step in diagnosing various diseases at an early stage. With the advancement of various machine learning algorithms, the prediction has become quite easy. However, the complex and the selection of an optimal machine learning technique for the given dataset greatly affects the accuracy of the model. A large amount of datasets exists globally but there is no effective use of it due to its unstructured format. Hence, a lot of different techniques are available to extract something useful for the real world to implement. Therefore, accuracy becomes a major metric in evaluating the model. In this paper, a disease prediction approach is proposed that implements a random forest classifier on Parkinson's disease. We compared the accuracy of this model with the Principal Component Analysis (PCA) applied Artificial Neural Network (ANN) model and captured a visible difference. The model secured a significant accuracy of up to 90%.

Motivation & Objective

• Motivate early prediction of Parkinson's disease to reduce healthcare burden and enable timely intervention.
• Develop a high-accuracy, scalable classifier for high-dimensional medical data using PCA for feature reduction.
• Compare PCA-RF against PCA+ANN baselines to evaluate the impact of feature reduction on classifier performance.
• Leverage PySpark for handling high-dimensional data and evaluate standard performance metrics.

Proposed method

• Apply PCA to extract a reduced feature set from a 754-feature Parkinson’s dataset.
• Train a Random Forest classifier on the PCA-reduced features to predict disease status.
• Use Gini impurity to guide decision tree splits within the Random Forest.
• Evaluate model performance using accuracy, sensitivity, specificity, precision, F1 score, and ROC analysis.
• Implement the pipeline on PySpark and compare with PCA+ANN baselines.

Experimental results

Research questions

• RQ1 Does applying PCA before Random Forest improve or degrade Parkinson’s disease prediction on the given dataset?
• RQ2 How does PCA-RF compare to PCA+ANN in terms of accuracy and other performance metrics?
• RQ3 What is the impact of PCA on model performance without compromising key metrics like sensitivity and specificity?
• RQ4 Is the PCA-RF approach scalable for high-dimensional medical datasets?

Key findings

• Without PCA, PCA-RF achieves 89.867% accuracy, 70.175% sensitivity, 96.470% specificity, 70.175% precision, and 77.669% F1 score.
• With PCA, PCA-RF achieves 76.651% accuracy, 55.555% sensitivity, 80.628% specificity, 35.087% precision, and 43.010% F1 score.
• The ROC analysis shows a high true positive rate for the proposed model, indicating good discriminative ability.
• Compared to PCA+ANN, PCA-RF performs better in the non-PCA setting but performs worse when PCA is applied, highlighting the interaction between feature reduction and classifier choice.
• The study notes a notable accuracy difference between the PCA-RF and PCA+ANN baselines across reported metrics.