PSI - Issue 78

Franco Ciminelli et al. / Procedia Structural Integrity 78 (2026) 921–928

927

Experimental validation on a sample of 100 real bridges confirmed the robustness of the approach, yielding highly satisfactory results: the model achieved an overall accuracy of 79.0% , with a weighted precision of 80.4% and a weighted F1-score of 78.3% . These outcomes, obtained without using the subjective parameter Level of Defectiveness (LoD), highlight the potential of this approach to support seismic risk management in a rapid, cost-effective, and objective manner. The proposed framework can serve as a valuable tool for infrastructure managers, enabling more efficient prioritization of inspections and planning of interventions — particularly in contexts with limited resources. Among the main limitations, it is noted that while the accuracy on real cases is good, it could be further improved by expanding the validation sample, integrating data from heterogeneous sources, and refining input parameter calibration to better capture real-world variability. Finally, the same framework can be extended in the future to predict the remaining three Classes of Attention defined by the Guidelines — structural and foundational, landslide, and hydraulic — thus contributing to an integrated, automated, and multi-risk management of national infrastructure assets.

Fig. 3. Confusion matrices for the prediction of the CoA-S: (a) results obtained on synthetic data; (b) results on real bridges.

Table 2. Classification report for CoA-S on synthetic data.

Table 3. Classification report for CoA-S on real data.

Class High

precision recall

f1-score support

Class High

precision recall

f1-score support

0,995

0,991 0,974 0,957 0,880 0,643 0,972 0,889 0,972

0,993 0,967 0,960 0,878 0,750 0,972 0,910 0,972

3039 2233 1989

0,643

0,643 0,722 0,971 0,500 1,000 0,790 0,767 0,790

0,643 0,776 0,868 0,667 0,857 0,790 0,762 0,783

14 36 34 10

Medium-High 0,961

Medium-High 0,839

Medium

0,964

Medium

0,786

Medium-Low 0,876

258

Medium-Low 1,000

Low

0,900 0,972 0,939

28

Low

0,750 0,790 0,803

6

accuracy macro avg

7547 7547 7547

accuracy macro avg

100 100 100

weighted avg 0,972

weighted avg 0,804

References Abdallah, A. M., Atadero, R. A., & Ozbek, M. E. (2022). A State-of-the-Art Review of Bridge Inspection Planning: Current Situation and Future Needs. Journal of Bridge Engineering , 27 (2). https://doi.org/10.1061/(asce)be.1943-5592.0001812 Alogdianakis, F., Dimitriou, L., & Charmpis, D. C. (2022). Data-driven recognition and modelling of deterioration patterns in the US National Bridge Inventory: A genetic algorithm-artificial neural network framework.