AI Model Predicts Construction Injuries from Safety Reports with 79% Accuracy
A new study published in 2026 in the Internasional Journal of Integrative Sciences demonstrates that artificial intelligence can accurately predict workplace injury types by analyzing written accident reports. The research was conducted by Kimberly Long Holt, MS, of Health and Safety Concepts – Environmental Health & Safety. Using Natural Language Processing (NLP) and machine learning, Holt analyzed 16,878 construction accident records collected between 2018 and 2022. The findings show that AI can transform routine safety documentation into a proactive risk prediction tool, helping organizations prevent injuries before they occur.
The study, titled “Predictive Risk Modeling via Natural Language Processing of Industrial Safety Reports,” introduces a scalable and replicable method for converting unstructured safety narratives into predictive intelligence. The research is openly accessible under a Creative Commons license and carries the DOI: https://doi.org/10.55927/ijis.v5i2.8.
Why This Research Matters
Workplace injuries remain a major social and economic burden worldwide. Non-fatal occupational accidents lead to lost productivity, increased insurance costs, regulatory penalties, and long-term consequences for workers and families. In construction alone, injury rates remain consistently high.
Traditional safety management systems are largely reactive. Organizations analyze incidents after they happen. While structured data fields—such as date, time, and location—are routinely reviewed, the detailed narrative descriptions written by safety professionals are rarely analyzed systematically. These reports often contain rich descriptions of accident mechanisms, environmental conditions, worker actions, and equipment factors. Until recently, most of this information remained underused.
Advances in Natural Language Processing and machine learning now allow computers to extract patterns from large volumes of text. Holt’s research shows that safety narratives contain measurable predictive signals that can be used to forecast injury outcomes with high reliability.
How the Study Was Conducted
The research used a quantitative, non-experimental design based on secondary data. The dataset consisted of 16,878 documented construction accidents recorded between 2018 and 2022. Each record included both structured fields and detailed free-text descriptions.
The analysis followed four main steps:
1. Data Screening and CleaningRecords missing key injury classifications or containing fewer than 20 narrative words were removed.
The narratives were standardized using established NLP techniques:
- Tokenization (breaking sentences into words)
- Stop-word removal
- Stemming and lemmatization
- Text normalization
- Industry-specific terminology such as PPE and OSHA was preserved.
The cleaned text was converted into numerical data using Term Frequency–Inverse Document Frequency (TF-IDF).
Up to 5,000 relevant text features were extracted, along with 15 contextual metadata variables such as project type and shift timing.
Multiple algorithms were tested. The best-performing model was a Random Forest ensemble classifier, optimized using 5-fold cross-validation and grid search tuning.
The model was trained to predict three injury categories:
- Upper limb injuries
- Lower limb injuries
- Head and neck injuries
Key Findings
The Random Forest model delivered strong predictive performance:
- Accuracy: 79.3%
- Precision: 77.1%
- Recall: 78.0%
- F1 Score: 78.5%
- AUROC: 0.98
An AUROC of 0.98 indicates near-perfect ability to distinguish between injury types.
Most Important Predictors
Feature importance analysis revealed:
- Accident mechanism and nature of injury were the strongest predictors.
- Temporal factors (day of week, shift timing) had minimal predictive value.
- Economic variables such as project cost were also weak predictors.
These results indicate that the physical details described in accident narratives—rather than scheduling or financial context—carry the most predictive power.
Real-World Impact
The study demonstrates a shift from reactive to proactive safety management.
Traditional metrics such as Total Recordable Incident Rate measure past performance. Holt’s NLP-based system enables organizations to forecast injury risks based on narrative patterns found in previous reports.
Potential applications include:
- Identifying high-risk patterns in near-miss reports
- Prioritizing safety inspections based on predicted injury severity
- Allocating training resources to high-risk mechanisms
- Enhancing enterprise-wide risk intelligence systems
According to Kimberly Long Holt of Health and Safety Concepts, “Unstructured safety narratives contain actionable predictive signals that can be systematically extracted to support proactive risk intervention rather than reactive response.”
The system is designed to support—not replace—safety professionals. AI provides consistent, scalable analysis, while human experts apply contextual judgment and ethical decision-making.
Broader Implications for Industry
The methodology is replicable across sectors that rely on narrative incident reporting, including:
- Manufacturing
- Mining
- Oil and gas
- Healthcare
- Aviation
Because the model relies on widely available reporting systems, implementation costs remain manageable once data infrastructure is in place.
The study also emphasizes ethical governance. Predictive analytics must protect worker privacy and avoid punitive use. The goal is injury prevention, not surveillance.
Future Research Directions
Holt recommends expanding research in several areas:
- Developing industry-specific safety lexicons
- Predicting injury severity and lost work time
- Applying temporal sequence analysis to detect early intervention points
- Using transfer learning to adapt models across industries
- Conducting longitudinal studies to measure real-world injury reduction
Further validation could demonstrate measurable reductions in workplace injuries following AI deployment.
Author Profile
Kimberly Long Holt, MS Health and Safety Concepts – Environmental Health & Safety
Source
DOI: https://doi.org/10.55927/ijis.v5i2.8
Official URL : https://journalijis.my.id/index.php/ijis/index
0 Komentar