When Industry Data Meets Machine Learning: Systemic Barriers to Predictive Safety in the Construction Projects

Abstract

Machine learning (ML) has great potential in predictive safety management in the construction industry, which is among the most hazardous sectors worldwide. Although there has been an increasing interest in research and even organizational investment, there have been limited practical implementations of ML. Most of the current literature generates and tests algorithms based on curated or synthetic data, without answering a more fundamental question: why are the data systems that construction organizations operate on not supporting the needs of ML? Hence, this paper explores systematic data quality issues that arise when machine learning is applied to operational construction safety data and why these issues persist despite organizational investment in digitalization. Cross-case synthesis of three ML implementation projects within the DiSCo research program is carried out, which have variation in data type: structured relational databases (Case 1: 2.1 million worker records, 929 projects), semi-structured HSE records (Case 2: 778,000 quality deviations, 63,000 observations), and unstructured JSA texts (Case 3: 3,425 documents). This can be explained by three incompatibility types, namely structural (82% loss of project-level data as a result of unlinkable database tables), semantic (78-99% loss of contextual data as a result of data entry done under compliance requirements), and epistemic (tacit expert knowledge not encoded in text, demonstrated by quantitative features performing better than NLP). All types are reproduced structurally by data storage, compliance-based incentive structures, and professional boundary processes. The research provides a conceptual typology of data-model discrepancies to evaluate ML preparedness before algorithm construction. The key outcome is that predictive safety is not a machine learning problem in the first place but an organizational design problem, the solution of which lies in data governance, incentive redesign, and knowledge externalization.