Investigating a real-time adaptive procedure system

Procedural issues are significant contributors to incidents in high-risk industries, yet current procedures fail to accommodate the gap between how work is prescribed and how it is actually performed. This study investigates the feasibility of developing a Real-Time Adaptive Procedure System that uses machine learning to understand procedural deviations and adapt procedures dynamically in high-risk industrial settings.Data were collected from eight workers performing three tasks at a high-fidelity petrochemical facility. Using the Skip-Order-Action framework, researchers documented procedural deviations by comparing observed behavior against prescribed procedures across 507 procedural steps. A logistic regression model was developed using features from the Multi-disciplinary Interactive Behavior Triad framework, including participant characteristics, step-level attributes derived from abstraction hierarchy levels and cognitive performance modeling, and real-time indicators such as timestamps and eye-tracking data.The model using step-level and participant-level features achieved relatively better performance (F1 = 0.49, AUC = 0.82), outperforming baseline models (F1 = 0.14-0.23). Adding real-time indicators reduced performance (F1 = 0.38), suggesting challenges in integrating real-time behavioral data.This proof-of-concept demonstrates that procedural deviations follow patterns based on task characteristics and worker attributes. The study establishes empirical foundations for machine learning-driven procedural adaptation and proposes a theoretical framework that bridges the gap between prescribed and actual work through adaptive systems rather than compliance monitoring, representing a paradigm shift toward learning-oriented procedural systems in high-risk industries.