Source-linked AI summary
Analyzing the impact of feature selection on the accuracy of heart disease prediction
Muhammad Salman Pathan, Avishek Nag, Muhammad Mohisn Pathan, Soumyabrata Dev
TL;DR
High-dimensional medical data makes heart-disease prediction and feature relationships difficult to analyze. This study identifies relevant risk factors using feature-selection analysis across two datasets and benchmarked classifiers. Reduced feature sets improved classification accuracy over full feature sets, with reported gains from 0.73 to 0.75 for CVD and from 0.66 to 0.71 for the Framingham dataset.
Problem
High-dimensional medical datasets complicate heart-disease prediction and make it difficult to identify important feature relationships.
Method
The study analyzes feature importance and applies ANOVA-based feature selection across two heart-disease datasets, then compares classifiers trained on full and reduced feature sets.
Results
0.75 versus 0.73 accuracy for CVD and 0.71 versus 0.66 for the Framingham dataset after reducing the feature set.
Takeaways & Limitations
Reduced feature sets produced better classification performance than full feature sets in the evaluated models.
Takeaways & Limitations
The authors identify using more than one feature-selection technique as future work.
Abstract
from arXiv · showhide
Heart Disease has become one of the most serious diseases that has a significant impact on human life. It has emerged as one of the leading causes of mortality among the people across the globe during the last decade. In order to prevent patients from further damage, an accurate diagnosis of heart disease on time is an essential factor. Recently we have seen the usage of non-invasive medical procedures, such as artificial intelligence-based techniques in the field of medical. Specially machine learning employs several algorithms and techniques that are widely used and are highly useful in accurately diagnosing the heart disease with less amount of time. However, the prediction of heart disease is not an easy task. The increasing size of medical datasets has made it a complicated task for practitioners to understand the complex feature relations and make disease predictions. Accordingly, the aim of this research is to identify the most important risk-factors from a highly dimensional dataset which helps in the accurate classification of heart disease with less complications. For a broader analysis, we have used two heart disease datasets with various medical features. The classification results of the benchmarked models proved that there is a high impact of relevant features on the classification accuracy. Even with a reduced number of features, the performance of the classification models improved significantly with a reduced training time as compared with models trained on full feature set.
1. Introduction
Heart disease creates a strong need for timely, accurate diagnosis, while high-dimensional medical data complicates prediction. The study examines whether selecting relevant features improves machine-learning classification while reducing complexity.
- 17.90 million people died from heart disease in 2016, underscoring the need for accurate and timely detection.
- Machine-learning models can support heart-disease diagnosis by processing medical data and producing disease predictions.
- Redundant and irrelevant features can add noise, increase model complexity, slow training, and contribute to prediction errors.
- Selecting features closely related to the target can reduce dimensionality and improve prediction accuracy.
- The study analyzes two patient datasets, feature interdependence, filter-based selection, multiple classifiers, and performance on full versus reduced feature sets.
- Classification models including Logistic Regression, Decision Tree, Naive Bayes, Random Forest, and Multi Layer Perceptron are evaluated on full and reduced feature subsets.
2. Related Work
Prior heart-disease studies use diverse feature-selection strategies to reduce input dimensionality and improve classification. However, existing work provides limited systematic guidance on feature choice and feature interdependence.
- Reported methods identified clinically relevant variables, including blood pressure, age, ALP, LDH, Serum Creatinine, and Ejection Fraction.
- Earlier studies combined filter, wrapper, embedded, optimization, and dimensionality-reduction methods with classifiers including SVM, MLP, NB, and XGBoost.
- Selected feature subsets often matched or improved predictive performance, including 97.7% stroke accuracy and 88.34% versus 83.34% heart-disease accuracy.
- Feature-selection techniques were also associated with improved accuracy and computation time, with Logistic Regression reaching 89% using RelieF.
- Existing research often lacks systematic guidance for selecting input features and pays limited attention to feature correlation and individual importance.
- This study addresses that gap using two datasets, interdependence and importance analysis, and comparisons between full and selected feature sets.
3. Proposed Methodology
The methodology analyzes two heart disease datasets, preprocesses and balances them, examines feature relationships, and selects important predictors using ANOVA-F scores. It identifies age, hypertension, glucose, and blood-pressure-related variables as influential risk factors, while several other features show little correlation or importance.
- Datasets: The study uses CVD and Framingham datasets to examine medical features and develop machine-learning systems for heart disease detection.The datasets were selected for their feature variety, data availability, volume, and similarity of features.
- Feature Correlation Analysis: CVD features age, hypertension, heart_disease, and avg_glucose_lvl correlate positively with stroke at 0.57, 0.24, 0.27, and 0.2, respectively.Framingham features age, sysBP, prevalentHyp, diabBP, and glucose also show positive correlations with TenYearCHD.
- Feature Correlation Analysis: Gender, bmi, heart rate, smoking habits, education, social status, and living standards show little or no correlation with the output features.The paper reports that age, hypertension, and glucose are common features closely related to outcomes in both datasets.
- Feature Selection: ANOVA-F feature importance identifies age, hypertension, heart_disease, and avg_glucose_lvl as important for stroke, while age, prevalentHyp, diabetes, sysBP, diaBP, and glucose rank highest for TenYearCHD.The importance results show similarity with the correlation analysis, particularly for age, hypertension, glucose, and blood pressure.
4. Evaluation Matrices
The study evaluates classifiers using Accuracy, F1-score, and ROC-related measures, with confusion-matrix outcomes providing the basis for these metrics. These measures capture overall correctness, balance between precision and recall, and classification capability.
- Accuracy, F1-score, and ROC were used to evaluate classifier performance.F1-score is defined as the harmonic mean of Precision and Recall, while ROC evaluates true-positive and false-positive rates.
- A confusion matrix describes classifier performance through four outcomes: true positives, true negatives, false negatives, and false positives.These outcomes distinguish correct predictions from errors in identifying disease presence or absence.
- False negatives are considered the most harmful predictions in the medical setting described.A false negative incorrectly classifies a patient with heart disease as disease-free.
5. Results and Discussions
The experiments compare machine-learning classifiers trained with full and ANOVA-F-selected feature sets across CVD and Framingham datasets. Reduced feature subsets generally improved classification results while lowering computational cost, with the strongest CVD result reaching 0.74 across Accuracy, F1-Score, and ROC using four features.
- Experimental design: Models were evaluated on both datasets using full features and selected subsets to measure the effect of feature selection.Performance was checked with Accuracy, F1-score, and ROC.
- Full feature set: 0.73 Accuracy, 0.74 ROC, and 0.73 F1-Score were achieved by MLP on the CVD dataset using the full feature set.Other classifiers, including LR, SVC, and RF, also produced reasonable results, whereas the dummy classifier reached only 0.46 Accuracy.
- Full feature set: 0.66 Accuracy, 0.67 ROC, and 0.66 F1-score were the strongest reported Framingham full-feature results.LDA, LR, and ridge classifier produced similar performance, while feature-value ranges were suggested as a possible reason for weaker results.
- Feature selection: ANOVA-F scores were used to identify the most outcome-relevant features in each dataset.Four of 11 CVD features and five of 15 Framingham features were selected.
- Selected feature set: 0.74 Accuracy, 0.74 F1-Score, and 0.74 ROC were achieved by SVC on the CVD dataset using only 4 input features.The selected CVD features were age, hypertension, heart_disease, and avg_glucose_lvl.
- Selected feature set: 0.71 Accuracy was achieved on the Framingham dataset with a reduced feature set, exceeding all reported full-feature accuracies for that dataset.The selected features were age, prevalentHyp, sysBp, diaBp, and glucose.
- Computational cost: Reduced-feature models used 3.86 iterations per second on CVD and 15.52 iterations per second on Framingham during training.The reported full-feature training rates were 10.98 iterations per second for CVD and 24.20 iterations per second for Framingham.
- Overall findings: The experiments report that relevant-feature selection improved model performance while reducing the feature-space size and computational iterations.Comparisons with other published proposals using the same datasets produced accuracy results less than or equal to those obtained with reduced feature sets.
6. Conclusion and Future Works
The study evaluated feature selection for heart disease prediction using CVD and Framingham datasets, identifying important features and comparing full and reduced feature sets. Reduced feature sets improved classification accuracy while lowering training time and computational complexity; future work will expand datasets, models, and selection techniques.
- The study analyzed CVD and Framingham heart disease datasets to assess how feature selection affects machine-learning prediction performance.
- ANOVA-F feature selection identified age, hypertension, glucose, previous heart disease, and blood pressure among the most important risk factors.
- 0.75 and 0.71 accuracy were achieved with reduced features for the CVD and Framingham datasets, compared with 0.73 and 0.66 using all features.
- Reduced feature sets produced better model performance than full feature sets while using fewer features and less training time.
- Selecting only important features reduced computational complexity and improved prediction-model accuracy.
- Future Works: Future work will evaluate additional datasets, combinations of machine-learning and deep-learning models, and multiple feature-selection techniques.