Using a machine learning-based risk prediction model to analyze the coronary artery calcification score and predict coronary heart disease and risk assessment

Improving Heart Disease Probability Prediction Sensitivity with a Grow Network Model

“…Explainable AI (XAI) such as LIME [17] [18] was used to interpret the model, however, no initiative was conducted to show feature importance hueing the predictive classes. An approach to predict coronary heart disease and risk assessment was conducted by Huang et al [12] using a collected dataset. The obtained accuracy, specificity, sensitivity, and AUC were 0.7896, 0.5113, 0.9386, and 0.8375 respectively using a Random Forest (RF) model.…”

“…The obtained accuracy, specificity, sensitivity, precision, recall, f1 score, and AUC were 0.9878, 0.971, 0.9791, 0.9807, 0.9531, 0.9789, and 0.96 respectively using a Learning Vector Quantization (LVQ) model. In the above studies [4] [5] [10] [12] [13] [14] there was mention of much more important metrics rather than accuracy to rely on the outcomes obtained by these studies. However, these studies [4] [5] [10] [12] [13] [14] were evaluated across a limited and balanced number of records.…”

“…Class imbalances are common among medical survey data [1] [6]. It is challenging to achieve high accuracy for the minority class while remaining well-balanced [1] [6] [12]. Sensitivity and specificity analyses are therefore vital for demonstrating such reliability and ensuring balanced outcomes.…”

“…Sensitivity and specificity analyses are therefore vital for demonstrating such reliability and ensuring balanced outcomes. Prior studies that attempted such additional analyses found that the sensitivity and specificity were eventually unbalanced [1] [6] [12], resulting in biased results - an important problem that the proposed strategy in this work addresses. Prior studies that showed feature significance in predicting heart disease failed to provide precise information [11].…”

“…The proposed lightweight GrowNet model obtained high recall (0.81) for the minority class (heart disease) while maintaining a balanced recall (0.74) for the majority class (healthy) across heavily unbalanced data. This was a main challenge in heart disease prediction that many prior studies [1] [6] [12] struggled to ensure.…”