Analysis of Occupational Fatalities Across U.S. Industries¶

Eunsoo Suk

Table of Contents¶

1. Introduction

2. Data Collection

3. Data Cleaning

   • 3.1 Initial inspection
   • 3.2 Standardizing industry names
   • 3.3 Converting accident-type columns
   • 3.4 Recomputing total fatalities
   • 3.5 Cleaning and imputing NAICS codes
   • 3.6 Normalizing industry names and manual mapping
   • 3.7 Removing summary rows and final dataset

4. Exploratory Data Analysis (EDA)

   • 4.1 Dataset overview
   • 4.2 Distribution of total fatalities
   • 4.3 Yearly trend of total fatalities
   • 4.4 Top industries by cumulative fatalities
   • 4.5 Accident-type composition
   • 4.6 Industry-year heatmap
   • 4.7 NAICS-level analysis

5. Industry Clustering (Autoencoder + KMeans)

   • 5.1 Feature aggregation
   • 5.2 Scaling
   • 5.3 Autoencoder architecture
   • 5.4 Training the autoencoder
   • 5.5 Latent embedding extraction
   • 5.6 KMeans clustering
   • 5.7 Descriptive cluster labels
   • 5.8 Latent-space visualization
   • 5.9 Cluster-level accident-type profiles
   • 5.10 Representative industries

6. Fatal Injury Forecasting

   • 6.1 Annual time-series preparation
   • 6.2 Holt’s Linear Trend model
   • 6.3 Visualization of Holt’s forecast
   • 6.4 Linear Regression trend model
   • 6.5 Forecast comparison
   • 6.6 Interpretation of forecasting results
   • 6.7 Accident-Type Forecasting (Multivariate Perspective)

7. Statistical Validation and Robustness Checks

   • 7.1 Bootstrap confidence intervals
   • 7.2 Levene’s test for variance shift
   • 7.3 ANOVA across clusters

8. Clustering Robustness and Validation

   • 8.1 Silhouette Score and Davies–Bouldin Index
   • 8.2 Hierarchical Clustering of Top 40 High-Fatality Industries

9. Key Findings and Discussion

10. Limitations and Future Work

11. Policy Interpretation / Actionable Insights

1. Introduction¶

Understanding workplace fatalities is essential for identifying high-risk industries, improving safety standards, and informing future prevention strategies. Each year, the U.S. Bureau of Labor Statistics (BLS) publishes detailed fatal-injury tables that summarize how and where these incidents occur—categorized by industry, accident type, and year. Although these tables offer valuable insights, they are separated across multiple files and vary in format, terminology, and data quality. As a result, meaningful analysis requires careful integration and preprocessing before any conclusions can be drawn.

This project brings together nearly a decade of BLS fatal-injury data and transforms it into a unified, analyzable dataset. After cleaning inconsistencies in industry names, correcting missing or ambiguous NAICS codes, and standardizing accident-type counts, the data is used to explore major patterns in occupational fatalities. Through visualization, clustering, and trend analysis, the project highlights which industries are most affected by specific accident types and how conditions have changed over time.

In addition to descriptive analysis, the project incorporates forecasting models to estimate future fatal-injury trends for industries with sufficient historical data. By combining data engineering, statistical modeling, and machine-learning techniques, the study provides an interpretable, data-driven overview of workplace safety in the United States—offering insights that can support both research and real-world safety decision-making.

2. Data Collection¶

The dataset used in this project comes from the U.S. Bureau of Labor Statistics (BLS), specifically the Census of Fatal Occupational Injuries (CFOI) program.
Each year, the BLS publishes a detailed table titled:

“Table A-1. Fatal occupational injuries by industry and event or exposure”

This table reports the total number of fatal workplace injuries across all major U.S. industries, broken down by accident types including:

• Transportation incidents
• Violence and other injuries by persons or animals
• Contact with objects and equipment
• Falls, slips, and trips
• Exposure to harmful substances or environments
• Fires and explosions

To conduct a consistent 10-year analysis, we collected and standardized Table A-1 data for 2014–2023, then merged all years into a single dataset ready for downstream analysis (trend analysis, clustering, and forecasting).

2.1 Source Files and Naming¶

Most of the CFOI A-1 tables are provided directly as downloadable Excel workbooks on the BLS website, while the 2014 table is only available as a PDF.
To ensure consistency across all years, we harmonized these formats into a single standardized structure.

Excel files (2015–2023)¶

On the BLS website, each yearly table appears under titles such as:

• “Industry by event or exposure, 2022 (XLSX)”
• “Industry by event or exposure, 2021 (XLSX)”

When downloaded, these files follow a consistent standardized naming pattern:

• fatal-occupational-injuries-table-a-1-2014.pdf
• fatal-occupational-injuries-table-a-1-2015.xlsx
• fatal-occupational-injuries-table-a-1-2016.xlsx
• …
• fatal-occupational-injuries-table-a-1-2023.xlsx

These Excel tables share a common layout and can be loaded directly using pandas.read_excel().

PDF file (2014)¶

The 2014 A-1 table is distributed only as a PDF (fatal-occupational-injuries-table-a-1-2014.pdf).
To maintain a fully reproducible workflow, the PDF was parsed programmatically and converted into a dataframe with the same schema as the Excel files.

Data sources¶

All files were obtained from the official BLS CFOI database:

https://www.bls.gov/iif/fatal-injuries-tables.htm
https://www.bls.gov/iif/fatal-injuries-tables/archive.htm#RATES

Unified directory¶

All downloaded files (2014 PDF + 2015–2023 Excel workbooks) were stored in a single directory, enabling a unified loading loop that iterates through years and builds a fully consolidated dataset.

2.2 Parsing the 2014 PDF¶

Because the 2014 table is only available as a PDF, we implemented a small line-level parser using pdfplumber and regular expressions:

1. Read each page of the PDF and split it into text lines.
2. For each line, detect whether it represents a valid industry row by checking:
   • the last seven tokens are numeric (or --) and correspond to accident-type counts,
   • the remaining tokens contain the industry name and, optionally, a NAICS code.
3. Split the line into:
   • Industry name,
   • NAICS code (when present),
   • seven accident-type values.
4. Assemble these into a dictionary and collect them as rows in a dataframe.

This produces a 2014 dataframe that already matches the standard column layout used for the later Excel files.

2.3 Standard Table Schema and Basic Cleaning¶

To be able to merge all years into a single dataset, we enforce a common column schema and apply a small amount of structural cleaning.

Standard column order¶

We define a fixed column order that is used across all years:

• Industry
• NAICS code
• Total fatal injuries
• Violence and other injuries by persons or animals
• Transportation incidents
• Fires and explosions
• Falls, slips, trips
• Exposure to harmful substances or environments
• Contact with object and equipment

Removing non-data rows¶

The original tables include footnotes, notes, and source lines (e.g., “Footnotes”, “Source: U.S. Department of Labor”, or rows starting with (1)).
These are removed using a simple text filter which checks the Industry column and drops rows that correspond to notes rather than actual industries.

Minimal cleaning steps¶

For each year’s table we:

1. Force column names to match the standard schema (truncating if necessary).
2. Drop rows that are completely empty.
3. Remove footnotes / notes / header explanation rows.
4. Keep only rows where Industry is a valid string and strip extra spaces.
5. Drop exact duplicate rows.

The helper functions below implement this logic.

2.4 Loading and Merging All Years (2014–2023)¶

Before loading the datasets, the notebook automatically downloads all A-1 tables directly from the BLS website.
To ensure full reproducibility on any machine, the script first detects the user’s Desktop directory dynamically:

from pathlib import Path

desktop_dir = Path.home() / "Desktop"
save_dir = desktop_dir / "Data"

Once all files have been downloaded, we load and merge them into a unified dataframe:

1. Identify the directory containing yearly A-1 files (~/Desktop/Data).
2. 2014: parse the PDF version using pdf_2014_to_df().
3. 2015–2023: read Excel versions using pandas.read_excel().
4. Normalize raw column names (trim and collapse whitespace).
5. Apply a common cleaning function clean_df() to each table.
6. Append a Year column to every dataframe.
7. Concatenate all annual tables into a single dataset df_raw.

The final merged dataframe contains standardized columns and one row per industry–year combination, enabling consistent multi-year analysis of occupational fatalities.

Summary¶

After running the pipeline above:

• All 10 years (2014–2023) of Table A-1 data are combined into a single dataframe df_raw.
• The 2014 PDF is parsed into the same column layout as the Excel years.
• Non-data rows such as footnotes and notes are removed.
• Each record includes:
  • Industry,
  • NAICS code,
  • total fatal injuries,
  • seven accident-type counts,
  • and the corresponding Year.

This cleaned, unified dataset forms the basis for the subsequent analyses:

• NAICS-level trend analysis,
• accident-type distributions,
• industry clustering, and
• time-series forecasting of fatal injuries.

3. Data Cleaning¶

This section describes the full cleaning pipeline applied to the raw dataset after merging all yearly A-1 tables. The cleaning process includes:

• Standardizing text fields (e.g., Industry)
• Converting accident-type columns to numeric
• Recomputing totals where needed
• Cleaning and imputing NAICS codes
• Removing summary rows and zero-fatality entries

The final cleaned dataset is stored as df_clean.

3.2 Clean Industry Names (whitespace, line breaks, special spaces)¶

Industry names often contain non-breaking spaces, inconsistent whitespace, or line breaks.
These issues prevent proper grouping and merging across years.
We apply a text-normalization function to standardize all names.

3.3 Converting accident-type columns to numeric values¶

Accident-type fields sometimes contain placeholders such as “--” or include comma formatting.
We convert all injury fields to numeric and drop rows where all numeric components are missing.

3.4 Recomputing and validating total fatalities¶

Some rows contain totals that do not match the sum of accident-type components.
We recompute the total values and replace inconsistent entries.
We also remove rows where total fatalities are zero.

3.5 Cleaning NAICS codes and filling missing values¶

The NAICS field contains missing values and inconsistent formatting (e.g., decimals like "92216.0").
We standardize the formatting and fill missing NAICS codes using the most common code for each Industry.

3.6 Normalizing Industry names and applying high-level NAICS mapping¶

For some broad industry categories, NAICS codes are not included in the A-1 tables.
We normalize Industry names (letters only, lowercase) and apply a manual mapping to fill high-level groups.

3.7 Removing summary rows and finalizing the clean dataset¶

The A-1 tables include broad summary rows (e.g., "Total," "Goods producing"). These rows are removed because they do not represent individual industries. We produce the final cleaned dataset as df_clean.

4. Exploratory Data Analysis (EDA)¶

After cleaning the dataset, we perform exploratory data analysis to understand the distribution of fatal injuries, industry-level patterns, yearly trends, and accident-type characteristics. All analyses in this section use the final cleaned dataset df_clean.

4.1 Dataset overview¶

We begin with a structural overview of the cleaned dataset including dimensionality, column types, and missing values.

4.2 Distribution of total fatalities¶

We first examine the distribution of fatal injury counts across all Industry–Year entries in the dataset.

4.3 Yearly trend of total occupational fatalities¶

We aggregate total fatal injuries by year to visualize long-term trends from 2014 to 2023.

4.4 Top industries by cumulative fatalities¶

We aggregate fatalities across all years (2014–2023) and identify the 20 industries with the highest cumulative fatality counts.

4.5 Accident-type composition¶

We analyze the share of each accident type aggregated across all years.

4.6 Heatmap of fatalities by industry and year¶

This heatmap visualizes how fatal injury counts vary across industries and over time.

4.7 NAICS-level analysis¶

Using the cleaned NAICS codes, we can aggregate fatal injuries by NAICS sector to identify the most hazardous sectors.

5. Industry Clustering (Autoencoder + KMeans)¶

To identify broader patterns among industries, we apply unsupervised learning to group industries based on their fatal-injury profiles. Because the dataset contains multiple accident-type variables that may exhibit nonlinear relationships, we use a shallow autoencoder to learn a low-dimensional (2-dimensional) latent representation. Clustering is then performed in this latent space using KMeans.

This approach provides an interpretable structure of U.S. industries based on their accident-type risk profiles over the last decade.

5.1 Construct industry-level feature matrix¶

We aggregate all accident types across all years for each industry. This produces a compact view of accident profiles for clustering.

5.2 Scaling the features¶

Autoencoders and KMeans both require numerical features on comparable scales. We apply StandardScaler to normalize accident-type totals.

5.3 Autoencoder architecture¶

We use a shallow autoencoder with:

• 16-unit hidden layer (ReLU)
• 2-unit latent layer (linear)
• 16-unit decoder layer
• Output layer matching the original feature dimension

The model is trained to reconstruct the original accident-type profile.

5.4 Training the model¶

The autoencoder is trained using mean squared error (MSE) loss until reconstruction stabilizes.

5.5 Extracting 2D latent embeddings¶

Each industry is now represented as a 2-dimensional embedding capturing its overall accident-type risk profile.

5.6 KMeans clustering on the latent space¶

We apply KMeans with k=4, producing four interpretable clusters that reflect different industrial risk patterns.

5.7 Assigning descriptive cluster labels¶

We assign meaningful labels to each cluster based on observed profiles and industry characteristics.

5.8 Latent-space visualization¶

The 2-dimensional latent space reveals four clearly separated groups that correspond to:

• high-volume transport/public sector industries (Cluster 0)
• low-risk services and building/office operations (Cluster 1)
• construction and heavy specialty trades (Cluster 2)
• mixed production, resource extraction, and energy-related sectors (Cluster 3)

The clean boundaries confirm that the autoencoder successfully learned meaningful representations of industry-level accident patterns

5.9 Cluster-level accident-type profiles¶

We compute the average accident-type counts within each cluster to interpret their risk characteristics.

5.10 Representative industries per cluster¶

For interpretability, we list the top industries (by total fatalities) within each cluster.

6. Fatal Injury Forecasting¶

To understand long-term safety trends and anticipate future fatality levels, we perform time-series forecasting using annual total fatalities from 2014–2023. Two predictive approaches are evaluated:

1. Holt’s Linear Trend (Exponential Smoothing)
2. Linear Regression Trend Model

These models provide simple but interpretable projections for the next year (e.g., 2024), based on the decade-long historical trend.

6.1 Preparing the annual time-series¶

We aggregate total fatalities per year using the cleaned dataset df_clean.
This produces a 10-year time-series suitable for regression-based forecasting.

6.2 Holt’s Linear Trend Model¶

Holt’s method extends exponential smoothing by adding a trend component. We fit the model on the 2014–2023 data and forecast the next year (2024).

6.3 Visualization of Holt’s forecast¶

6.4 Linear Regression Trend Model¶

To compare with Holt’s model, we also fit a simple linear regression model to capture the overall upward/downward trend of annual fatalities.

6.5 Visualization of linear regression trend forecast¶

6.6 Comparing the forecast results¶

We summarize the predicted fatality levels from each model.
Both methods yield almost identical forecasts for 2024, which increases confidence in the overall trend estimate.

Interpretation¶

• Holt’s method adapts more flexibly to recent changes in trend.
• Linear regression captures the overall long-term trend but assumes linearity across all years.

Together, these models provide a reasonable initial projection of future occupational fatality levels.
More advanced forecasting (e.g., ARIMA, Prophet, LSTM) could be explored if longer time-series data becomes available.

6.7 Accident-Type Forecasting (Multivariate Perspective)¶

Forecasting total fatalities provides a high-level outlook but cannot reveal whether specific accident categories are becoming more or less prominent over time. Because prevention strategies are often category-specific, we extend the forecasting framework to a multivariate setting that analyzes accident-type–level trends separately.

To capture these category-level dynamics, we construct a 10-year time series (2014–2023) for each accident type, fit the same interpretable models used earlier (Holt’s linear trend and linear regression), and generate next-year forecasts for each series. The results are summarized in both a forecast table and a combined historical-plus-forecast heatmap.

This multivariate perspective offers richer and more actionable insight by highlighting which accident types are expected to remain dominant or rise in the near future, rather than treating all fatalities as a single aggregated count.

Yearly time-series per accident type

Forecasting next-year fatalities by accident type

Heatmap: historical + forecasted accident-type trends

Interpretation¶

Linking the accident-type forecasts back to the cluster structure reveals several actionable implications. Falls, slips, and trips—one of the dominant hazards in construction-heavy industries—remain at a high and persistent level rather than showing a sharp decline, indicating that fall-prevention programs will continue to be critical for construction-oriented clusters.

Overall, the multivariate forecasts show that transportation incidents and contact with objects or equipment are expected to remain the leading contributors to occupational fatalities, whereas categories such as fires and explosions continue to account for relatively small volumes. By combining the 10-year historical trends with next-year projections in a unified heatmap, we gain a compact and forward-looking view of how the risk landscape may evolve across accident categories. This perspective supports more targeted prevention strategies—for example, prioritizing transportation safety initiatives even in industries where aggregate fatality counts appear stable.

7.1 Bootstrap Confidence Interval for Forecasts¶

To quantify the uncertainty in our forecasting results, we estimate 95% confidence intervals using bootstrap resampling. Rather than relying on model assumptions, the bootstrap directly resamples historical annual fatality values and fits the forecasting model repeatedly to generate a distribution of predictions. This provides a non-parametric and robust measure of predictive uncertainty.

Bootstrap CI for Holt’s Model

Bootstrap CI for Linear Regression

Add CI to Forecast Comparison Table

7.2 Levene’s Test for Variance Shift (Before vs After 2019)¶

To evaluate whether the variability in annual fatality counts has changed over time, we perform Levene’s test comparing:

• Period A: 2014–2018
• Period B: 2019–2023

Levene’s test is robust to non-normality and tests whether two groups have significantly different variances.

Interpretation rule

p < 0.05 → variance changed significantly after 2019
p ≥ 0.05 → no statistically significant change in variance

7.3 ANOVA Across Clusters¶

To test whether the clusters identified in Section 5 truly differ in accident-type composition,
we conduct ANOVA for each accident type across the four clusters.

A statistically significant ANOVA (p < 0.05) supports the interpretation that clusters represent meaningful structural differences.

8. Clustering Robustness and Validation¶

While the autoencoder-based KMeans clustering in Section 5 produced a meaningful representation of industry-level accident-type profiles, clustering outcomes can vary depending on initialization, dimensionality-reduction method, and data noise.
To ensure that the discovered cluster structure is stable and statistically sound, we assess clustering robustness using three complementary approaches:

1. Silhouette Score and Davies–Bouldin Index
2. Hierarchical Clustering (Top 40 Highest-Fatality Industries)

These analyses validate the reliability of the clustering results and provide additional insight into the structural similarity among industries.

8.1 Silhouette Score and Davies–Bouldin Index¶

Silhouette Score evaluates how well-separated and cohesive the clusters are (closer to 1 = better clustering quality).
Davies–Bouldin Index measures cluster compactness and separation (lower values indicate clearer cluster boundaries).

For the 2-dimensional autoencoder latent space:

• Silhouette Score ≈ 0.938
• Davies–Bouldin Index ≈ 0.632

These metrics indicate strong intra-cluster cohesion and large inter-cluster separation, supporting the validity of using k = 4 clusters.

8.2 Hierarchical Clustering of Top 40 High-Fatality Industries¶

To further examine whether the four-cluster structure aligns with natural groupings in the data, we apply Ward’s hierarchical clustering to the top 40 industries ranked by total fatalities. This provides a structural view of how industries merge based on accident-type similarity.

The dendrogram reveals:

• Clear separation between construction, heavy manufacturing, and transportation sectors
• Consistent grouping patterns that closely match the autoencoder-based clusters
• Meaningful substructures within service-oriented and administrative industries

The hierarchical organization confirms that the four-cluster solution reflects the underlying geometry of the accident-type data.

9. Key Findings and Discussion¶

This study analyzed ten years of BLS occupational fatality data (2014–2023) to identify structural patterns across industries, accident types, and temporal trends. The results reveal where occupational risks are concentrated and how these risks are likely to evolve.

9.1 Transportation incidents remain the dominant source of fatalities¶

Across the decade, transportation incidents consistently accounted for the largest proportion of fatal injuries.
This trend was robust across industries of varying sizes and remained visible even after cluster-based segmentation.

Other major contributors included:

• contact with objects/equipment
• falls, slips, and trips
• exposure to harmful substances

Fires and explosions were relatively rare.

The accident-type forecasting results reinforce this pattern, indicating that transportation-related fatalities are projected to remain elevated in the upcoming year. This highlights a persistent structural risk associated with vehicle and equipment movement across multiple sectors.

9.2 Four distinct industry “risk archetypes” emerged from clustering¶

The clustering analysis reveals four distinct “risk archetypes” across U.S. industries:

• Archetype 1 – Transport/Public/High-count

Large public-facing sectors and transportation-intensive industries consistently exhibit high fatality levels. Their exposure profile is dominated by transportation incidents, contact injuries, and large workforce size.

• Archetype 2 – Services/Office/Building

Administrative, business services, and building-operations industries form the lowest-severity group. They involve structured work environments, relatively fewer hazardous exposures, and stable year-over-year fatality counts.

• Archetype 3 – Construction/Heavy

Construction and specialty trades represent the most hazard-intensive cluster. Falls, trips, contact injuries, and equipment-related events are defining characteristics of this group. This aligns with long-standing regulatory emphasis on construction safety.

• Archetype 4 – Production/Energy/Mixed

Agriculture, mining, utilities, and mixed production sectors share high physical-hazard exposure and environmental risk factors. Their risk patterns combine transportation, machinery, chemical, and environmental hazards.

Together, these archetypes offer a structured lens for understanding how fatal-injury patterns concentrate across the U.S. industrial landscape.

9.3 Historical patterns reveal persistent high-risk sectors¶

Long-term heatmaps show recurring concentrations of fatalities in:

• Trade, transportation, and utilities
• Construction
• Truck transportation
• Natural resources and mining

These sectors demonstrate both high annual totals and long-term persistence, suggesting that their elevated risk levels are structural rather than temporary fluctuations.

9.4 Overall fatality levels show limited signs of decline¶

Forecasting using Holt’s trend model and linear regression suggests that total occupational fatalities are likely to remain near current levels.
Both Holt’s trend model and the linear regression model converge to a forecast of roughly 17,750 fatalities in 2024, very close to the recent historical average.

The close agreement between models indicates a stable underlying trend with no strong evidence of long-term decline.

9.5 Accident-type forecasting provides deeper insight than aggregate totals¶

Extending the forecasting process to individual accident types reveals more nuanced trends:

• Transportation incidents and contact-with-object injuries remain the highest-impact categories.
• Falls, harmful substance exposures, and violence-related incidents remain relatively stable but meaningful contributors.
• Fires and explosions show volatility due to small annual counts.

By combining historical and forecasted accident-type patterns, organizations can tailor preventive strategies to the specific categories most likely to influence future outcomes.

10. Limitations and Future Work¶

Several limitations in the dataset and methodology should be acknowledged, along with opportunities for further research.
These considerations help clarify the boundaries of the current analysis and suggest directions for enhancing future occupational safety studies.

10.1 Limited temporal granularity¶

The CFOI Table A-1 dataset reports only annual fatality counts.
More granular data—such as monthly, weekly, or event-level records—would enable richer time-series analyses and stronger causal inference.

Future work: Incorporate complementary sources such as OSHA incident logs or employer-reported injury datasets to build higher-resolution time-series models.

10.2 Ambiguity in NAICS code assignment¶

Some industries in the raw dataset lacked NAICS codes or included only high-level categories.
Imputation based on majority mappings reduces missingness but cannot fully eliminate structural ambiguity.

Future work: Combine CFOI tables with NAICS-coded OSHA inspection records or employment datasets to improve industry-level classification accuracy.

10.3 Forecasting is constrained by the short 10-year time window¶

A decade of data limits the use of more advanced time-series techniques.
Models such as ARIMA, Prophet, or multivariate state-space approaches typically require longer historical sequences for stable parameter estimation.

Future work: Extend the dataset using earlier CFOI archives or additional historical sources to support more robust long-term forecasting.

10.4 Accident-type forecasting is sensitive to small sample sizes¶

Certain accident categories—especially fires and explosions—have low annual totals, resulting in noisier forecasts.
Small sample sizes reduce the stability and reliability of Holt’s trend and linear regression models.

Future work: Apply smoothing techniques, combine related event categories, or use Bayesian hierarchical models to better handle low-frequency data.

10.5 Clustering captures only fatality patterns¶

The clustering analysis focuses exclusively on fatal incidents and does not incorporate non-fatal injuries, exposure data, workforce size, or incident severity.
Fatality-only patterns may not fully represent the broader occupational risk landscape.

Future work: Integrate non-fatal injury data, fatality rates (per 100,000 workers), and exposure metrics to build a more comprehensive understanding of industry risk.

11. Policy Interpretation / Actionable Insights¶

The results of the clustering and forecasting analyses offer several practical implications for industry stakeholders, policymakers, and safety professionals.
These insights highlight where targeted interventions, technological investment, and regulatory focus can meaningfully reduce occupational fatality risks.

11.1 Strengthen transportation-related safety across multiple industries¶

Transportation incidents consistently represent the largest share of occupational fatalities.
Improving transportation safety is therefore relevant across nearly all industry groups.

Effective measures include:

• enhanced driver training programs
• adoption of telematics and GPS-based monitoring
• fatigue detection and management systems
• implementation of advanced driver-assistance technologies (ADAS)
• safer vehicle and equipment operation protocols

Because transportation risks cut across many sectors, interventions in this area have broad potential impact.

11.2 Develop targeted safety programs aligned with industry risk clusters¶

Different industry clusters show distinct accident-type profiles.
Targeting safety practices to match these risk structures improves the relevance and efficiency of interventions.

Examples:

• Construction-heavy sectors: emphasize fall prevention, equipment operation safety, scaffold/ladder protocols, and routine site hazard assessments.
• Transport and logistics sectors: focus on roadway hazards, cargo handling, vehicle inspection routines, and warehouse traffic management.
• Manufacturing sectors: prioritize machine guarding, lockout/tagout (LOTO) compliance, and ergonomic hazard reduction.

Cluster-based guidance enables organizations to design safety programs that closely match their risk environments.

11.3 Allocate regulatory and inspection resources based on projected accident-type trends¶

Accident-type–level forecasting highlights which categories are likely to remain major contributors in the coming years.
Regulators can use these projections to direct inspection resources and compliance audits toward:

• high-burden accident types
• industries with persistent elevated risk
• emergent categories showing upward trends

This approach improves the impact of regulatory oversight and supports more strategic allocation of limited resources.

11.4 Establish ongoing monitoring systems integrating heatmaps and forecasting¶

The combined use of historical heatmaps and next-year forecasting creates a useful foundation for early-warning systems.
By updating these models annually with new CFOI data, safety practitioners can:

• identify rising risks at an earlier stage
• monitor whether interventions are having measurable effects
• detect deviations from long-term trends
• make data-informed adjustments to safety strategies

This allows for proactive rather than reactive safety management.

11.5 Encourage technological adoption in high-risk clusters¶

Several technologies offer strong potential to reduce fatalities, especially in the highest-risk clusters (construction, logistics, and manufacturing):

• automated machine monitoring
• sensor-based hazard detection
• wearable safety devices (fatigue, location, fall detection)
• real-time situational awareness systems
• robotics and automation for high-exposure tasks

Encouraging adoption through incentives, standards, or voluntary programs can produce measurable reductions in fatality risks.