Cardiovascular disease prediction demands accurate models alongside feature selection processes that are transparent, reproducible, and diagnostically interpretable. Existing wrapper-based metaheuristic selectors operate as black boxes, reporting only final accuracy without revealing how features are selected or whether subsets remain consistent across data samples. We propose WaOA-MRFO, an adaptive hybrid optimizer that combines Walrus Optimization Algorithm exploration with Manta Ray Foraging Optimization exploitation through stagnation-aware diversity control. A structured iteration-level diagnostic suite monitors population diversity, convergence behaviour, and feature selection stability, making the search dynamics directly observable rather than opaque. Under 5<InlineEquation ID="IEq1"><EquationSource Format="TEX">$\times $</EquationSource><EquationSource Format="MATHML"><math><mo>×</mo></math></EquationSource></InlineEquation>5 nested cross-validation on the Cleveland heart disease cohort (<InlineEquation ID="IEq2"><EquationSource Format="TEX">$n=261$</EquationSource><EquationSource Format="MATHML"><math><mrow><mi>n</mi><mo>=</mo><mn>261</mn></mrow></math></EquationSource></InlineEquation>), WaOA-MRFO achieves an AUC of <InlineEquation ID="IEq3"><EquationSource Format="TEX">$0.843 \pm 0.056$</EquationSource><EquationSource Format="MATHML"><math><mrow><mn>0.843</mn><mo>±</mo><mn>0.056</mn></mrow></math></EquationSource></InlineEquation> — statistically equivalent to the all-features baseline (AUC <InlineEquation ID="IEq4"><EquationSource Format="TEX">$=0.847$</EquationSource><EquationSource Format="MATHML"><math><mrow><mo>=</mo><mn>0.847</mn></mrow></math></EquationSource></InlineEquation>, <InlineEquation ID="IEq5"><EquationSource Format="TEX">$p=0.619$</EquationSource><EquationSource Format="MATHML"><math><mrow><mi>p</mi><mo>=</mo><mn>0.619</mn></mrow></math></EquationSource></InlineEquation>) — while reducing the feature set by <InlineEquation ID="IEq6"><EquationSource Format="TEX">$61\%$</EquationSource><EquationSource Format="MATHML"><math><mrow><mn>61</mn><mo>%</mo></mrow></math></EquationSource></InlineEquation>. The method matches LASSO and recursive feature elimination on identical splits and significantly outperforms uninformed selection at matched dimensionality (<InlineEquation ID="IEq7"><EquationSource Format="TEX">$\Delta \text {AUC}=+0.091$</EquationSource><EquationSource Format="MATHML"><math><mrow><mi mathvariant="normal">Δ</mi><mtext>AUC</mtext><mo>=</mo><mo>+</mo><mn>0.091</mn></mrow></math></EquationSource></InlineEquation>, <InlineEquation ID="IEq8"><EquationSource Format="TEX">$p<0.001$</EquationSource><EquationSource Format="MATHML"><math><mrow><mi>p</mi><mo><</mo><mn>0.001</mn></mrow></math></EquationSource></InlineEquation>). A stable four-feature core (exang, ca, thalch, oldpeak) emerges in more than <InlineEquation ID="IEq9"><EquationSource Format="TEX">$70\%$</EquationSource><EquationSource Format="MATHML"><math><mrow><mn>70</mn><mo>%</mo></mrow></math></EquationSource></InlineEquation> of folds, while the top two features are selected in every fold, indicating reproducible signal extraction. EvoMapX operator attribution analysis identifies the WaOA spiral operator as the primary convergence driver, accounting for <InlineEquation ID="IEq10"><EquationSource Format="TEX">$66.7\%$</EquationSource><EquationSource Format="MATHML"><math><mrow><mn>66.7</mn><mo>%</mo></mrow></math></EquationSource></InlineEquation> of improvement events, while the Population Evolution Graph reveals the progressive stabilisation of the informative feature core across iterations. Fixed-mask transfer evaluation on an independent three-site cohort (<InlineEquation ID="IEq11"><EquationSource Format="TEX">$n=616$</EquationSource><EquationSource Format="MATHML"><math><mrow><mi>n</mi><mo>=</mo><mn>616</mn></mrow></math></EquationSource></InlineEquation>) yields AUC <InlineEquation ID="IEq12"><EquationSource Format="TEX">$=0.811 \pm 0.041$</EquationSource><EquationSource Format="MATHML"><math><mrow><mo>=</mo><mn>0.811</mn><mo>±</mo><mn>0.041</mn></mrow></math></EquationSource></InlineEquation>, confirming cross-institutional transferability of the stable-tier features. Ablation analysis identifies bit-flip mutation as the sole statistically significant contributor to subset consistency (<InlineEquation ID="IEq13"><EquationSource Format="TEX">$p=0.028$</EquationSource><EquationSource Format="MATHML"><math><mrow><mi>p</mi><mo>=</mo><mn>0.028</mn></mrow></math></EquationSource></InlineEquation>, Cohen’s <InlineEquation ID="IEq14"><EquationSource Format="TEX">$d = 0.531$</EquationSource><EquationSource Format="MATHML"><math><mrow><mi>d</mi><mo>=</mo><mn>0.531</mn></mrow></math></EquationSource></InlineEquation>. These results establish that competitive predictive performance with feature sets reduced by <InlineEquation ID="IEq15"><EquationSource Format="TEX">$61\%$</EquationSource><EquationSource Format="MATHML"><math><mrow><mn>61</mn><mo>%</mo></mrow></math></EquationSource></InlineEquation> is achievable alongside transparency, cross-institutional transferability, and process-level interpretability.

Interpretable hybrid metaheuristic optimization with iteration-level behavior analysis for clinical feature selection in heart disease prediction

期刊论文
出版模式：

开放获取
发表日期：

2026年6月1日

查看全文

Shamsuddeen Adamu,
Hitham Alhussian,
Said Jadid Abdulkadir,
Majdy Mohamed Eltayeb Eltahir,
Khairy Sallam O. F.,
Gasim Hyder,
Fakhreldin Ali,
Umar Audi Ismaíla,
Shamsu Abdullahi,
Hussaini Mamman,
Mohammed Gamal Ragab

摘要

Cardiovascular disease prediction demands accurate models alongside feature selection processes that are transparent, reproducible, and diagnostically interpretable. Existing wrapper-based metaheuristic selectors operate as black boxes, reporting only final accuracy without revealing how features are selected or whether subsets remain consistent across data samples. We propose WaOA-MRFO, an adaptive hybrid optimizer that combines Walrus Optimization Algorithm exploration with Manta Ray Foraging Optimization exploitation through stagnation-aware diversity control. A structured iteration-level diagnostic suite monitors population diversity, convergence behaviour, and feature selection stability, making the search dynamics directly observable rather than opaque. Under 5$\times $ $\times$ 5 nested cross-validation on the Cleveland heart disease cohort ($n=261$ $n = 261$ ), WaOA-MRFO achieves an AUC of $0.843 \pm 0.056$ $0.843 \pm 0.056$ — statistically equivalent to the all-features baseline (AUC $=0.847$ $= 0.847$ , $p=0.619$ $p = 0.619$ ) — while reducing the feature set by $61\%$ $61 %$ . The method matches LASSO and recursive feature elimination on identical splits and significantly outperforms uninformed selection at matched dimensionality ($\Delta \text {AUC}=+0.091$ $Δ AUC = + 0.091$ , $p<0.001$ $p < 0.001$ ). A stable four-feature core (exang, ca, thalch, oldpeak) emerges in more than $70\%$ $70 %$ of folds, while the top two features are selected in every fold, indicating reproducible signal extraction. EvoMapX operator attribution analysis identifies the WaOA spiral operator as the primary convergence driver, accounting for $66.7\%$ $66.7 %$ of improvement events, while the Population Evolution Graph reveals the progressive stabilisation of the informative feature core across iterations. Fixed-mask transfer evaluation on an independent three-site cohort ($n=616$ $n = 616$ ) yields AUC $=0.811 \pm 0.041$ $= 0.811 \pm 0.041$ , confirming cross-institutional transferability of the stable-tier features. Ablation analysis identifies bit-flip mutation as the sole statistically significant contributor to subset consistency ($p=0.028$ $p = 0.028$ , Cohen’s $d = 0.531$ $d = 0.531$ . These results establish that competitive predictive performance with feature sets reduced by $61\%$ $61 %$ is achievable alongside transparency, cross-institutional transferability, and process-level interpretability.