A hybrid approach with metaheuristic optimization and random forest in improving heart disease prediction

Performance measure

The effectiveness of the proposed method is evaluated based on the selected features and the accuracy of classification. The performance is assessed using key metrics, including accuracy, sensitivity, precision, and recall. The formulas used for calculating specificity, F-score, precision, recall, and accuracy are given in Table 3.

Design setup for experiment

The proposed model is evaluated using optimization techniques and machine learning models. The process begins with data collection, followed by preprocessing to manage any missing values. Next, the dataset undergoes optimization, and the selected features are utilized for classifying heart disease patients using a machine learning model. Various evaluation metrics are applied to assess performance. The experiment is conducted on Anaconda Jupyter Notebook 6.4.8, which includes built-in machine learning packages, and runs on an Intel(R) Core (TM) i7-7600U CPU @ 2.80GHz–2.90GHz.

Feature selection results

This work investigates feature selection to optimize the accuracy of ML algorithms for heart disease prediction.  The authors employed the SelectKBest method with three filter methods: Chi-square, Mutual Information, and F-statistics. Each feature’s rank was determined by each filter, and an overall rank was calculated by combining these individual ranks mentioned in Table 4. The lowest-ranked features were removed, and the remaining features were fed into the classification algorithms. Using all features, Random Forest achieved the highest accuracy of 90.16% from Table 4. After applying feature selection, Random Forest experienced a slight accuracy drop. However, Logistic Regression performance improved, reaching 90.16% accuracy Table 6.

This section explores three metaheuristic algorithms GA ACO, PSO for feature selection in predicting heart disease.These algorithms aim to identify a subset of important features from the original dataset, potentially enhancing the accuracy of the prediction model. All three algorithms successfully reduced the initial 13 features to a smaller set. Genetic Algorithm and Particle Swarm Optimization selected 10 features each. Ant Colony Optimization selected 9 features shown in Table 7. This study will compare the performance of these optimized feature sets with existing models utilizing all 13 features. Each optimized feature set will be fed into a Random Forest classifier to assess its impact on prediction accuracy. The results will be presented, comparing the accuracy achieved by different feature selection methods and traditional approaches using all features. In medical predictions, accuracy is paramount. By identifying the most relevant features, these algorithms have the potential to improve the model’s ability to correctly diagnose heart disease, potentially leading to better patient outcomes.

The parameters considered for GAO are.

Mutation Rate: Varied (1000, 100, 5).

Crossover Rate: Implicitly 100%.

Crossover Operator: Single-Point Crossover.

Number of Generations: 50 and 100 in different runs.

Population Size: 20.

Stopping Criteria: Number of generations.

Selection Operator: Fitness-proportional selection from a randomly chosen subset of 5 individuals.

Selection Pressure: Implicit in the selection operator.

Fitness Function: Based on the inverse of the root mean squared error from a Random Forest Regressor.

The parameters considered for PSO are.

Features = 13.

C1 = 0.5 (Cognitive factor influencing individual particle movement).

C2 = 0.5 (Social factor guiding particle interaction within the swarm).

W = 0.9 (Inertia weight controlling the influence of previous velocity).

K = 30 (Number of top-performing particles or neighboring particles considered).

P = 2 (Parameter defining the search space dynamics).

Iterations = 1000.

The parameters considered for ACO are.

iteration = 100.

n_ants = 5 (Total number of ants in the system).

n_citys = 5 (Total number of cities or nodes to be visited).

e = 0.5  (Evaporation rate, determining pheromone decay over time).

alpha = 1 (pheromone influence factor, controlling the impact of pheromone trails).

beta = 2 (visibility factor, emphasizing the importance of heuristic information in decision-making).

Based on the parameters used by each optimization algorithm, the features are selected and are listed in Table 7.

RF algorithm is performed for the features selected by GA, ACO, PSO. Table 8 records the performance metrics achieved after the feature selection by metaheuristics. Random forest with Genetic algorithm achieves 92% accuracy, precision, and recall with 94.1%, specificity of 92.59%, and Area under the curve of 95%.

Random forest with PSO achieves 91.8% accuracy, 94% recall 91.43% precision, specificity 88.89%, and Area under curve 93%. Random forest with ACO achieves only 86.88% accuracy, 91.18% recall, 86.11% precision, and specificity with 81.48% and 92% area under the curve. From Table 8, it is clear that GAORF performs better compared with other techniques and achieved 92%. The same is visualized with the help of a bar graph shown in Fig. 5.

Comparing the accuracy of the optimization techniques along with RF for the Cleveland dataset.

GAORF Achieves Highest Accuracy in Heart Disease Prediction. Accuracy is a key metric in evaluating the effectiveness of prediction models. In this study, Fig. 6 visualizes the accuracy achieved by different algorithms applied to the Cleveland dataset. The proposed Genetic Algorithm-optimized Random Forest (GAORF) model outperformed all other techniques with a remarkable 92% accuracy. This result underscores the efficiency of GAORF in both feature selection and classification for heart disease prediction. Table 9 summarizes the performance of different machine learning algorithms. Random Forest (RF) alone achieved 90.16% accuracy, demonstrating its strong potential. Other methods like LR (85.25%), KNN (68.85%), SVM (70.49%), and NB (85.25%) displayed lower accuracy. Notably, Grid Search optimization slightly improved RF accuracy to 86.89%, but remained below GAORF.

Comparing the accuracy of the proposed model along with the ML classification algorithm for the Cleveland dataset.

The different performance metrics for the proposed model GAORF are compared with varied Machine learning classification techniques. The performance is visualized with the help of Table 9. Figure 7 explains the accuracy variation between features selected using filter methods and optimization techniques. The line graph clearly explains that optimization techniques perform better compared with filter methods. Among the optimization techniques, GA achieved an accuracy of 92% in predicting heart disease. The performance metrics values are mentioned in Table 10.

Discussion

While the proposed approach using GAORF achieved promising results for heart disease prediction, it’s important to acknowledge the inherent limitations associated with machine learning and metaheuristic optimization techniques. The following Table 11 highlights some key limitations of the methods explored in this paper.

This study seeks to enhance heart disease prediction accuracy by integrating feature selection methods with machine learning algorithms. The work consists of 2 phases. Phase 1 is Feature Selection. The SelectKBest method with three filter methods (Chi-square, Mutual Information, and F-statistics) is employed to identify the most relevant features from the dataset. This reduced the initial feature set while potentially maintaining or even improving prediction accuracy. Initially, Random Forest performed best with 90.16% accuracy using all features. However, after applying feature selection, Logistic Regression showed better performance (90.16%). Phase 2 is applying Metaheuristic Optimization for Feature Selection. Observing the impact of feature selection on different algorithms, we explored metaheuristic techniques for further optimization. The study utilized GAO, PSO, and ACO to select even more impactful feature subsets. Each optimized feature set was then fed into a Random Forest classifier to assess its effectiveness. The combination of GAO and Random Forest (GAORF) achieved the highest accuracy of 92%, surpassing all other techniques. This demonstrates the effectiveness of GAORF in selecting optimal features for heart disease prediction. By combining metaheuristic optimization with machine learning, we were able to significantly improve prediction accuracy compared to using all features or basic feature selection methods. The existing model with the proposed model comparison is shown in Table 12.