This review explores the application of machine learning (ML) for predicting safety sign comprehensibility based on socio-demographic factors and cognitive sign features. We compare the efficacy of various supervised ML algorithms in identifying crucial predictors for enhancing safety sign design.
Introduction
Safety signs are essential for risk communication in diverse work environments. Understanding which factors influence sign comprehension is critical for effective safety management. Traditional statistical methods have limitations in analyzing complex datasets. This review investigates the potential of ML, a data-driven approach utilizing algorithms to learn patterns and make predictions, for addressing this challenge.
Methodology
Data Collection and Participants
Data was collected from 2310 male construction workers in Tehran, Iran, using a stratified sampling method. Participants were tested on their comprehension of 20 ISO-standardized safety signs across five categories: mandatory, prohibition, emergency, warning, and firefighting. They also rated each sign’s familiarity, concreteness, simplicity, meaningfulness, and semantic closeness.
Feature Selection and Classification
Eight features (age, experience, education level, and the five cognitive sign features) were initially considered. Filter methods (Information Gain, Pearson’s correlation, etc.) and wrapper methods (Correlation Feature Selection) were employed to identify the most relevant features. Fourteen supervised ML classifiers (Logistic Regression, SVM, K-NN, etc.) were then trained and evaluated using 10-fold cross-validation. Performance was assessed using accuracy, sensitivity, specificity, precision, F1-score, and AUC.
Results
Comprehension Scores and User Factors
Comprehension scores varied significantly across signs and sign categories. Prohibition and firefighting signs generally exhibited higher comprehension rates. Age and work experience significantly impacted comprehension across most sign categories. Education level showed a significant effect only on firefighting sign comprehension.
Feature Selection and Classification Performance
Feature selection techniques consistently identified familiarity, simplicity, and meaningfulness as the most relevant predictors of sign comprehension. Classifiers trained on these three features outperformed those trained on all eight features. The K-NN classifier achieved the highest accuracy (94.369%) using 10-fold cross-validation, demonstrating the effectiveness of learning from examples with these specific features.
Discussion
This comparative review highlights the potential of ML for predicting safety sign comprehensibility. The findings corroborate previous research emphasizing the importance of familiarity, simplicity, and meaningfulness in sign design. The superior performance of K-NN suggests its suitability for this specific learning task. The use of ML allows for a data-driven approach to optimize sign design, potentially leading to improved safety outcomes.
Conclusion
This study demonstrates the efficacy of using ML to learn from examples and predict safety sign comprehensibility. Focusing on familiarity, simplicity, and meaningfulness in sign design can significantly enhance comprehension rates. Future research should explore the application of deep learning techniques and larger datasets to further refine predictive models and improve safety sign design guidelines. This data-driven approach promises significant advancements in occupational safety and risk communication.
