In a constantly evolving digital landscape, object-oriented software plays a critical role but is frequently affected by structural and behavioral defects that undermine its maintainability. Existing detection techniques predominantly target structural flaws, often neglecting behavioral anomalies. This oversight limits their overall effectiveness in ensuring software quality. To overcome these limitations, this study proposes a novel approach capable of detecting both structural defects — namely Blob and Long Method — and a behavioral defect, Poltergeist. The proposed method is composed of two complementary modules. The first module, IsolationPlus, aims to extract true instances of the Poltergeist defect from an unlabeled dataset (Apache-Ant 1.6.2). It combines the Extended Isolation Forest (EIF) algorithm with expert feedback. This hybrid technique outperforms three benchmark anomaly detection methods by identifying 46 Poltergeist instances with an F1-score of 1.0, compared to 34 using Isolation Forest (IF), 23 using EIF and 12 using k-Nearest Neighbors (k-NNs) anomaly detection, which yielded F1-scores of 0.19, 0.17 and 0.13, respectively. These results highlight the effectiveness of integrating expert knowledge in behavioral defect detection. The second module, DetectDev, is a meta-model that leverages widely used ensemble learning techniques — bagging, stacking and boosting — alongside 10 base estimators, for a total of 1000 decision trees. Each sub-model specializes in detecting a specific type of defect. This targeted strategy significantly improved performance, achieving F1-scores of 0.72 for Blob, 0.73 for Long Method and 0.89 for Poltergeist, outperforming a stacking-only model, which obtained F1-scores of 0.71, 0.71 and 0.85, respectively.