This article describes a predictive injury prevention concept currently under development, not a finished or commercially available system. The work reflects an active effort to design, test, and refine an approach that could move occupational safety from reactive analysis toward real-time risk anticipation. The next step for this concept is a pilot phase, to be pursued through collaboration with a qualified technology partner capable of helping translate theory and design into a working implementation.

For most of its history, occupational safety has depended on learning from what has already gone wrong. Injuries occur, investigations follow, and controls are strengthened in hopes of preventing recurrence. While this approach has delivered meaningful progress, it leaves a persistent gap: the period of time when risk is forming but no one is yet hurt. Advances in predictive analytics and artificial intelligence now make it possible—at least in concept—to close that gap by identifying emerging risk conditions and intervening earlier than traditional systems allow.

More than a decade ago, I argued that the EHS profession needed to prepare for a fundamental shift in how risk would be identified and controlled—one driven by emerging digital and analytical capabilities that were only beginning to take shape. My vision in 2014 was that EHS could, and should, move beyond static indicators and retrospective analysis toward systems capable of continuously sensing conditions, integrating diverse data streams, and seeing risk before it hurts. While the term “AI” was not yet common in professional safety conversations, the intent was clear: use advanced analytics to proactively manage risk as a dynamic system rather than react to its failures. Today, the convergence of computer vision, machine learning, and causal modeling makes it possible to actively pursue that vision, translating early foresight into a concrete design effort aimed at redefining how safety risk is recognized, understood, and acted upon in real time.

The following example outlines how such a predictive approach could function in a manufacturing environment. It is intentionally presented as a design framework rather than a finished solution, with the goal of encouraging discussion, critique, and collaboration across the EHS and technology communities. The emphasis is on how modern data integration and causal analytics might be applied to injury prevention, and what new capabilities could emerge if these tools are implemented thoughtfully and responsibly.

The Limits of Reactive Safety Systems

Traditional injury prevention systems are inherently retrospective. Even many “leading indicators” are signals that something did exist rather than confirmation that it does exist right now. Audits, observations, and lagging metrics provide valuable insight, but they are episodic and often disconnected from the moment-to-moment realities of work.

As manufacturing systems become more complex, tightly coupled, and sensitive to production pressure, risk increasingly emerges dynamically. Unsafe behaviors, degraded equipment condition, environmental stressors, and organizational demands can align quickly, creating exposure that may not be visible through conventional reporting cycles.

A Shift Toward Real-Time Risk Awareness

Recent advances in artificial intelligence enable a fundamentally different approach. Instead of relying solely on periodic reviews, safety systems can now maintain continuous awareness of operating conditions. Rather than asking what went wrong, they can ask what is happening right now—and what combination of factors makes an injury more likely in this moment.

The predictive injury prevention concept described here is built around that shift. Its purpose is not to replace existing EHS processes, but to augment them with a real-time layer of risk intelligence that operates continuously alongside traditional systems.

Integrating Disparate Data Streams

One of the greatest challenges in advancing predictive safety is not the lack of data, but the fragmentation of it. In most manufacturing organizations, information relevant to injury risk exists across multiple systems that were never designed to work together. Video feeds, EHS management systems, Wearable devices, maintenance platforms, employee feedback/reporting systems, and production databases each capture a partial view of reality, often using different structures, time scales, and levels of data quality.

The predictive injury prevention concept addresses this challenge by using artificial intelligence not just as an analytical engine, but as an integration and data-preparation layer. Before any causal modeling occurs, AI is used to harmonize these disparate inputs into a form that can meaningfully support structural equation analysis.

From Raw Signals to Comparable Inputs

The first role of AI in the system is signal normalization. Video-based AI generates high-frequency observations—counts or rates of unsafe behaviors and conditions detected in specific zones. EHS systems produce lower-frequency, event-driven data such as hazard reports, near misses, and corrective action updates. Operational systems generate continuous performance data tied to production cycles, shifts, or equipment states.

Machine learning algorithms are used to align these inputs onto a common analytical timeline and spatial context. This includes time-window aggregation (for example, rolling 5–15 minute intervals), zone-level mapping, and shift-based normalization. The goal is to ensure that data describing behavior, system condition, and operational pressure are comparable and synchronized, rather than evaluated in isolation.

Data Quality, Noise Reduction, and Contextual Weighting

Raw data—particularly from video analytics—can be noisy. AI plays a critical role in filtering false positives, de-duplicating repeated observations, and weighting signals based on confidence and relevance. For example, repeated detections of the same behavior by the same individual in a short period are treated differently than multiple independent detections across a work group.

Natural language processing is applied to free-text fields in hazard reports, near-miss narratives, and employee concerns. These narratives are classified, clustered, and scored for relevance to specific risk drivers, allowing qualitative inputs to be translated into structured indicators without losing nuance.

Operational data are similarly contextualized. Production rates are evaluated relative to historical baselines rather than absolute values, distinguishing normal high output from abnormal stress. Maintenance indicators are adjusted for asset criticality and operating mode. In this way, AI ensures that the data feeding the model reflect meaningful deviations, not background variation.

Constructing Latent Variables for Structural Equation Modeling

Once data are cleaned, aligned, and contextualized, AI assists in feature construction—the process of grouping related observable indicators into candidate latent variables suitable for structural equation modeling. This step is critical, as SEM depends on theoretically sound groupings that reflect real-world risk mechanisms.

For example, AI-driven clustering and correlation analysis may confirm that PPE violations, line-of-fire exposure, and unsafe lifting consistently co-occur under similar conditions, supporting their use as indicators of a latent “Unsafe Acts” construct. Similarly, delayed preventive maintenance, rising vibration levels, and increased breakdown frequency may form a coherent “System Condition” construct.

Importantly, this process is guided by safety theory and professional judgment, not automated pattern recognition alone. AI accelerates discovery and validation, but human oversight ensures that constructs remain interpretable, defensible, and aligned with how work is actually performed.

Preparing Data for Causal Analysis

Before structural equation modeling is executed, AI-driven preprocessing ensures that the data meet the assumptions required for stable causal analysis. This includes handling missing data, standardizing variables, identifying outliers that represent true signals rather than errors, and testing for temporal stability.

The system also evaluates whether relationships between variables remain consistent over time or vary under different operating conditions. Where appropriate, models are adapted to account for site-specific or process-specific differences, allowing the causal structure to remain valid without forcing uniformity where it does not exist.

Creating a Living Risk Model

This integration process is not a one-time exercise. As new data are collected and conditions change, the AI layer continuously re-evaluates indicator performance, latent construct validity, and model fit. When new patterns emerge—such as a shift in how production stress influences behavior—the system flags these changes for review and model refinement.

The result is a living risk model: one that evolves with the operation, improves with experience, and maintains alignment between data, theory, and practice. This model also operates in real time, continuously integrating all data as it is recieved.

By using AI to integrate and prepare data for structural equation modeling, the system transforms disconnected signals into a coherent representation of risk. This foundation is what enables predictive analytics to move beyond correlation, providing reliable, explainable insight into how and why injury risk is forming in real time.

Understanding Risk Through Causal Modeling

Most safety analytics struggle with the same fundamental limitation: they treat risk factors as independent signals. An unsafe behavior is counted, a maintenance backlog is tracked, a production rate is monitored—each measured, trended, and reviewed largely on its own. While this provides visibility, it does not explain how these factors interact to create injury risk, nor does it help leaders understand which combinations of conditions matter most in a given moment.

Predictive injury prevention requires a different analytical approach—one that is explicitly designed to model cause-and-effect relationships in complex systems. This is where structural equation modeling (SEM) becomes a critical enabling technology.

SEM allows multiple observable signals to be grouped into broader, underlying risk drivers—often referred to as latent variables. These latent variables represent conditions that cannot be measured directly but are inferred from patterns in real-world data. For example, repeated PPE violations, frequent line-of-fire exposure, and unsafe lifting behaviors may collectively indicate an underlying behavioral risk state. Similarly, missed preventive maintenance, increasing breakdown frequency, and abnormal vibration levels may indicate system degradation that increases exposure even when work practices appear unchanged.

The power of SEM lies in its ability to model how these latent risk drivers influence one another and contribute—individually and in combination—to overall injury risk. Rather than assuming that all unsafe acts carry equal weight at all times, the model estimates how strongly each driver contributes to risk under current conditions, and how those contributions change as the system evolves.

In the predictive injury prevention concept, this modeling approach enables the calculation of an instantaneous risk level for a specific operating area. That risk level is not simply the sum of recent events. It reflects the structure of the system: how production pressure amplifies behavioral risk, how degraded equipment condition increases the consequence of minor errors, and how effective (or ineffective) corrective action processes dampen or accelerate exposure.

An Example Structural Equation Model

To make this more concrete, consider a simplified example of how risk might be modeled using SEM in a manufacturing environment.

First, observable data are grouped into latent drivers:

• Unsafe Acts (UA):
  Indicated by PPE violations, line-of-fire exposure, unsafe lifting, and bypassed guards.
• System Condition (SC):
  Indicated by preventive maintenance compliance, breakdown frequency, and equipment reliability measures.
• Operational Stress (OS):
  Indicated by production rate deviation, unplanned changeovers, and yield instability.
• Safety Response Capability (SRC):
  Indicated by hazard reporting rate, corrective action timeliness, and near-miss follow-up quality.

These latent variables are then related to overall injury risk through a structural equation such as:

Instantaneous Risk (IR) =
0.45 × Unsafe Acts

• 0.30 × Operational Stress
  − 0.25 × System Condition
  − 0.20 × Safety Response Capability

In this example, unsafe acts have the strongest direct influence on risk, but their effect is moderated by operational stress and system condition. High production pressure increases the impact of unsafe behaviors, while strong maintenance performance and timely corrective actions reduce overall risk, even when behaviors are not perfect.

The model can also include interaction effects, such as:

IR = … + 0.10 × (Unsafe Acts × Operational Stress)

This term reflects a reality familiar to most practitioners: the same behavior that is tolerated under stable conditions becomes far more dangerous when the system is under stress.

Importantly, these coefficients are not assumed—they are estimated from actual site data and recalibrated as conditions change. As new hazards are reported, corrective actions are completed, or production stabilizes, the relationships update, allowing the model to distinguish between short-term noise and meaningful shifts in risk.

Why This Matters in Practice

This causal structure is what allows the system to move beyond alerting and into guidance. When a monitored area transitions from green to yellow or red, the system does not simply state that risk is high. It can explain why—for example, that rising production pressure combined with delayed maintenance has amplified the impact of recurring unsafe lifting behaviors—and point to corrective actions most likely to reduce risk quickly.

Just as importantly, the model can confirm when interventions work. If a targeted maintenance action or workflow adjustment reduces the contribution of a specific risk driver, the overall risk score reflects that improvement in near real time. Over time, this creates a learning loop in which the organization gains insight not only into where risk exists, but into which controls are most effective under which conditions.

By modeling safety as a dynamic system rather than a static checklist, structural equation modeling provides the analytical backbone that makes predictive injury prevention both credible and actionable. It allows EHS professionals and operations leaders to see risk forming, understand its drivers, and intervene with precision—before someone gets hurt.

From Analytics to Action

To be usable at the front line, the system translates complex analytics into a simple visual signal: green, yellow, or red. A green state indicates stable conditions with effective controls. Yellow signals elevated risk requiring timely attention and local correction. Red indicates a critical condition where immediate intervention is warranted.

Behind each color is a clear explanation of the dominant risk drivers and a prioritized set of suggested corrective actions. This allows supervisors and EHS professionals to respond quickly and decisively, without sorting through dashboards or debating which metric matters most.

Explainability and Learning by Design

A key design principle of the system is explainability. The model is intended to support professional judgment, not replace it. Users can see which factors are driving risk, how those factors have changed over time, and whether previous interventions successfully reduced exposure.

As new data is collected, the system recalibrates. When corrective actions reduce risk, the model learns from that success. When new patterns emerge, it adapts. Over time, this creates a feedback loop that strengthens both predictive accuracy and organizational learning.

Supporting the Human Side of Safety

Equally important is how the system interacts with people. The intent is not surveillance, but early recognition and prevention. Feedback mechanisms allow users to validate or challenge AI observations, improving trust and model accuracy simultaneously.

When elevated risk is detected, the system can also trigger targeted coaching prompts or short, task-specific learning reminders. In this way, technology reinforces safe behavior at the moment it matters most, supporting—not replacing—the conversations that are central to effective safety leadership.

The Potential Benefits of Predictive Injury Prevention

When implemented effectively, the benefits of this approach would significant. Organizations will gain continuous risk awareness instead of periodic snapshots, enabling earlier and more precise intervention. Injuries can be reduced by addressing exposure before harm occurs rather than after.

Safety, maintenance, and operations gain a shared, data-driven view of system health, improving coordination and reducing friction between priorities. Leaders gain predictive insight instead of retrospective explanation, allowing resources to be focused where they will have the greatest impact. Most importantly, employees benefit from safer, more stable work environments where risks are recognized and controlled before someone gets hurt.

Predictive analytics do not replace the fundamentals of EHS practice—they strengthen them. By combining modern analytical tools with established safety principles, this approach offers a practical path toward fewer surprises, faster learning, and more reliable protection of people in complex manufacturing systems.