PSI - Issue 78

Marco Martino Rosso et al. / Procedia Structural Integrity 78 (2026) 1382–1388

1387

Fig. 3. RF feature importance and explainability with SHAP.

4. Conclusions In this study, the authors proposed a data-driven machine learning (ML) framework for predicting post earthquake building usability classes using data from Italian AeDES forms of about 12k buildings struck by the 2016-2017 Central Italy earthquake events in the Abruzzi region (Italy). Specifically, the authors compared the binary classification metrics of a support vector machine (SVM) model and a random forest (RF) dealing with an imbalanced learning problem, merging the four AeDES building usability classes into two output classes: AB (almost immediately usable buildings) versus CE (unusable buildings). Although both methods showed almost the same balanced accuracy performances over the hold-out test set, the RF slightly outperformed the SVM in generalization capabilities, but the SVM provided higher recall on the minority class CE. Besides the classification results, the authors critically analyzed the ML explainable tools, focusing on the RF model only, in order to identify the feature importance ranking, and attempted to associate physical and engineering consistent explanations to the ML data-driven learned capabilities. Shapley additive explanations (SHAP) analysis revealed that older buildings, irregular masonry textures, and lack of curbs/chains increased the tendency to predict unusability scores, thus matching empirical damage patterns from post- event surveys. The models’ reliance on seismic intensity metrics underscores the need to incorporate ground-motion parameters in usability assessments since vertical peak ground acceleration (PGA) and horizontal pseudo-acceleration evidenced a relevant impact on usability scores. Although this kind of ML application may strongly support decision-makers, especially in immediate post-emergency conditions, future studies should address the current limitations. Indeed, despite merging classes to provide a binary classification scheme, minority-class performance could be improved with dedicated imbalanced learning procedures, such as advanced sampling techniques (e.g., SMOTE). Moreover, further validation is needed for other seismic regions and building typologies since the current available dataset was limited to masonry buildings.