[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Previous Article in Journal
Development and Application of an Integrated Index for Occupational Safety Evaluation
You seem to have javascript disabled. Please note that many of the page functionalities won't work as expected without javascript enabled.
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Safety
Volume 11
Issue 1
10.3390/safety11010001
safety-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Investigating Injury Outcomes of Horse-and-Buggy Crashes in Rural Michigan by Mining Crash Reports Using NLP and CNN Algorithms

by
Baraah Qawasmeh
*,
Jun-Seok Oh
and
Valerian Kwigizile
Department of Civil and Construction Engineering, Western Michigan University, Kalamazoo, MI 49008, USA
*
Author to whom correspondence should be addressed.
Safety 2025, 11(1), 1; https://doi.org/10.3390/safety11010001
Submission received: 22 October 2024 / Revised: 20 December 2024 / Accepted: 27 December 2024 / Published: 30 December 2024
Browse Figures
Figure 1
<p>Data extraction framework.</p> "> Figure 2
<p>Crash narratives analysis flow.</p> "> Figure 3
<p>Unigram displayed as word clouds. (<b>a</b>) for minor/no injury; (<b>b</b>) for fatal/serious injury.</p> "> Figure 4
<p>(<b>a</b>) Trigram word frequency for fatal/serious injury outcome for horse-and-buggy crashes. (<b>b</b>) Trigram word frequency for minor/no injury outcome for horse-and-buggy crashes.</p> "> Figure 5
<p>The confusion matrix of the logistic regression model.</p> ">
Versions Notes

Abstract

:
Horse-and-buggy transportation, vital for many rural communities and the Amish population, has been largely overlooked in safety research. This study examines the characteristics and injury severity of horse-and-buggy roadway crashes in Michigan’s rural areas. Detailed crash data are essential for safety studies, as crash scene descriptions are mainly found in narratives and diagrams. However, extracting and utilizing this information from traffic reports is challenging. This research tackles these challenges using image-processing and text-mining techniques to analyze crash diagrams and narratives. The study employs the AlexNet convolutional neural network (CNN) to identify and extract horse-and-buggy crashes, analyzing (2020–2023) Michigan UD-10 rural crash reports. Natural Language Processing (NLP) techniques also identified primary risk factors from crash narratives, analyzing single-word patterns (“unigrams”) and sequences of three consecutive words (“trigrams”). The findings emphasize the risks involved in horse-and-buggy interactions on rural roadways and highlight various contributing factors to the severity of these crashes, including distracted or careless actions by motorists, nighttime visibility issues, and failure to yield, especially by elderly drivers. This study suggests prioritizing horse-and-buggy riders in road safety and public health programs and recommends comprehensive measures that could significantly reduce crash incidence and severity, improving overall safety in Michigan’s rural areas, including better signage, driver education, and community outreach. Also, the study highlights the potential of advanced image-processing techniques in traffic safety research that could lead to more precise and actionable findings, enhancing road safety for all users.

1. Introduction

In the United States, farming and ranching are among the most hazardous occupations [1]. However, agricultural injury surveillance, while a persistent challenge, is manageable. Significant gaps characterize it, but numerous opportunities for improvement and development exist. A unique area of concern within agricultural communities, particularly among the Amish and other groups relying on horse-and-buggy transportation, is at high risk of transportation crashes. These crashes pose a significant risk to the individuals involved and underscore these communities’ unique safety challenges [2]. Addressing these challenges through targeted surveillance and tailored safety interventions is not just essential; it is a beacon of hope in mitigating the risks associated with horse-and-buggy transportation.
Horse-and-buggy transportation is not just a mode of travel for many rural communities; it is a lifeline, particularly for the Amish population. This traditional form of transport is not just a part of their cultural heritage; it is their cultural heritage [3]. It represents a significant aspect of their daily lives, and its safety challenges should not be overlooked in contemporary traffic safety research.
The interaction between modern motor vehicles and horse-drawn buggies on public roadways poses unique risks and hazards [1]. Differences in speed, vehicle size, and visibility exacerbate these risks. As rural roadways are shared by diverse types of vehicles, understanding the specific characteristics and dynamics of horse-and-buggy crashes is crucial for developing effective safety interventions [2].
Addressing the safety challenges associated with horse-and-buggy transportation requires a multifaceted approach. Through dedicated research, targeted interventions, and community engagement, it is possible to reduce the risks and enhance the safety of these traditional forms of transport, thereby preserving the cultural heritage of the communities that rely on them [1]. However, the consequences of not acting are severe. Without comprehensive safety measures, the high incidence of transportation crashes could continue, leading to more injuries and fatalities and potentially threatening the survival of these traditional forms of transport.
Detailed crash data are essential for such safety studies, as comprehensive descriptions of crash scenes and riders’ behavior are predominantly available within crash narratives and diagrams [4,5]. However, extracting and applying this information from traffic crash reports poses significant challenges [6,7]. Various studies have employed diverse methodologies to analyze the severity of crash injuries. A long-standing practice in crash severity analysis, statistical modeling is a conventional method that yields dependable insights into the probability of crashes and produces straightforward results. Despite this, statistical modeling necessitates predefined relationships between dependent and independent variables and specific assumptions regarding underlying data distribution [8].
On the contrary, machine learning methodologies are increasingly embraced in this field due to their independence from presumptive associations among variables [8,9,10]. Consistently, recent advancements in text mining methodologies have shown promise in extracting valuable insights from crash narratives [11,12,13]. By employing natural language processing (NLP) techniques, researchers can analyze textual data from crash reports to identify patterns, factors, and contributing circumstances related to crashes (e.g., [6,14]). These text-mining approaches offer a complementary perspective to traditional statistical modeling, allowing for a more nuanced understanding of the complex dynamics involved in such incidents. However, challenges such as data preprocessing, standardization of terminology, and ensuring the accuracy of extracted information remain areas of ongoing research and development [15]. Integrating text mining methodologies with existing analytical frameworks can enhance the comprehensiveness and effectiveness of safety studies in this domain, ultimately contributing to developing targeted interventions and policies to reduce the risks associated with horse-and-buggy transportation.
Computer vision disciplines are the source of the computer vision methodologies utilized in safety-related research [16]. As part of the conventional computer vision process, feature extraction continues to be used for object detection and classification. A suitable feature model is necessary to implement precise object detection and classification [17,18]. This methodology will automatically extract an extensive set of image characteristics, which will subsequently be utilized in image classification. Convolutional neural networks (CNNs) and histogram of oriented gradients (HOGs) are a few examples. The capability of these algorithms to self-learn from a provided dataset is an advantage [18]. Pre-performing feature extraction is not required due to the end-to-end nature of the deep neural network process. The deep neural network has emerged as the preeminent technology for image processing and resolving computer vision-related challenges, coinciding with the progress made in computer hardware and software [18].
Meanwhile, Michigan police crash reports (UD-10) do not include a variable accounting for crashes caused by horse-drawn carriages in a structured format. This study represents the first utilization of CNNs to recognize such crashes from crash diagrams and fill the existing transportation safety research gap by investigating the characteristics associated with horse and buggy roadway crashes in Michigan’s rural areas. Detailed crash data, essential for such safety studies, are often embedded within crash narratives and diagrams, which are complex and require sophisticated techniques for extraction and analysis. The narratives are often lengthy and contain a mix of unstructured data, while the diagrams are graphical representations that need to be processed into a format suitable for analysis. This research employs cutting-edge image-processing techniques integrated with text-mining methodologies to tackle these challenges. The power of the AlexNet CNN architecture was utilized as a highly skilled investigator capable of discerning intricate patterns within photographs [19,20]. It is exceptionally well-suited for identifying various features from crash diagrams. The feature classification technology provided by AlexNet was employed to extract horse and buggy-related crashes in a tabulated format, facilitating the development of an automated system for safety data extraction based on crash diagrams. Furthermore, this study aims to evaluate the effectiveness of AlexNet CNN architecture in classifying this type of crash using crash report diagrams.
This study offers three primary contributions to the field of traffic safety research. First, it develops a novel approach using AlexNet CNN architecture to automatically extract and classify horse-and-buggy-related crashes from crash diagrams in Michigan’s rural crash reports (2020–2023). Second, it applies NLP techniques to analyze crash narratives, identifying primary risk factors and patterns using advanced text-mining methodologies. Finally, the study provides actionable insights into the unique risks associated with horse-and-buggy transportation, highlighting contributing factors and proposing targeted interventions to improve the safety of rural roadway users.
This paper’s structure is as follows: Section 2 discusses the crash database and outlines the research methodology. Section 3 presents the results and discusses the findings. Finally, Section 4 concludes the research with conclusions and suggestions for future work.

2. Materials and Methods

2.1. Crash Data

The present study employs UD-10 crash data gathered by the Michigan Safety Police (MSP) agency as the primary source of information regarding traffic crashes. The UD-10 crash data serve as a unique approach to record and systematically document many facts about traffic crashes. These details encompass driver characteristics, vehicle information, environmental conditions, and contributing factors. UD-10 crash data offer valuable insights into the causes and consequences of traffic crashes due to its extensive utilization and comprehensive nature in traffic safety analysis. We aim to discern patterns, trends, and risk factors linked to horse and buggy crashes by examining the (2020–2023) UD-10 crash reports of Michigan’s rural areas.
Upon reviewing the UD-10 rural crash reports, it was observed that such incidents were categorized as pedestrian crashes. Therefore, the initial step involved extracting pedestrian-related crashes from the rural crash database. During the study period, 634 pedestrian crash reports were identified and downloaded in Portable Document Format (PDF) from the Michigan Traffic Crash Facts website [21]. These reports included crash diagrams and narratives. The crash narratives, diagrams, and crash IDs were obtained iteratively from the texts and images of the crash reports using a Python script designed to extract the information in a tabulated format. Subsequently, the crash narratives and diagrams were merged with the crash metadata using the crash IDs. A comprehensive debugging process removed inconsistent or incomplete entries to assure data quality. As a result, the final database comprised 609 diverse pedestrian crash diagrams.

2.2. Methodology

This study employs convolutional neural network (CNN) architectures and k-fold cross-validation to extract horse-and-buggy information from pedestrian crash diagrams accurately. The procedure for data collection and preparation is methodical. Subsequent subsections (Section 2.2.1, Section 2.2.2 and Section 2.2.3) explain each stage, highlighting the strategies and tools to ensure consistent and reliable outcomes. The data extraction framework is illustrated in Figure 1.

2.2.1. Crash Diagrams Preprocessing

The crash diagrams database underwent adjustments during preprocessing to enhance the model’s performance and ensure uniformity. An initial resizing was performed using the “squish” method to maintain consistent resolution across all diagrams [22]. While this compromised the aspect ratio of the original images, it ensured uniform dimensions across the entire collection. Subsequently, normalizing picture pixel values is essential for training deep learning models. Normalization usually sets pixel values to {0–1}. These preprocessing methods minimized image dimension differences to optimize CNN convergence during training [23]. Ultimately, these efforts aimed to achieve more accurate and robust results in recognizing the horse and buggy-related crashes.

2.2.2. K-Fold Cross-Validation

This research involves a systematic data collection and preparation process, which employs the k-fold cross-validation and AlexNet CNN architecture to recognize horse and buggy crashes accurately. Figure 1 illustrates the data extraction framework. The extracted information regarding horse-and-buggy crashes was also organized into structured data formats and merged with the original metadata. This facilitated the analysis of patterns influencing injury severity outcomes.
We used k-fold cross-validation with k = 5 to evaluate our model’s ability to process novel data. K = 5 was selected as it balances computational efficiency and reliable performance evaluation, a widely accepted standard in machine learning research. This method splits our dataset into five equal-sized “folds” [24]. For each iteration, one-fold is used for validation and the other four for model training. This procedure is repeated five times, with each fold serving as the validation set precisely once. K-fold cross-validation provides the advantage of undertaking a comprehensive evaluation of our model across multiple subsets of the data, thereby reducing the impact of data variability and establishing a more dependable assessment of its performance [24].

2.2.3. AlexNet Convolutional Neural Network Architecture

This study employs AlexNet CNN architecture as a robust data extraction methodology to classify horse-and-buggy crashes using crash diagrams. Inspired by the human brain, this architecture exhibits exceptional intelligence, akin to highly skilled investigators capable of discerning intricate patterns within photographs [19,20]. This entity demonstrates remarkable proficiency in differentiating complex components, boundaries, and shapes depicted in images, making it particularly well suited for identifying various features from crash diagrams. Despite being a mature architecture, AlexNet is particularly suited for this application due to its simplicity, efficiency, and compatibility with moderate-sized datasets. Its proven ability to extract intricate patterns and features from images makes it ideal for analyzing detailed and complex crash diagrams in Michigan’s rural crash reports.
Convolution operations are generally executed using two-dimensional convolution layers that utilize trainable kernels or filters. An optional trainable bias for each kernel may be incorporated into these layers. The kernels scan the input in “strides” during these operations. The stride, a crucial parameter, directly affects the output size and total convolutions. As the stride increases, the kernels skip more space between each convolution, leading to a reduction in the output size and total convolutions [25]. The bias is added to the result after a multiplication operation between the input section and the kernel at each position. This process generates a feature map representing the convolved outcome. Typically, the feature maps are then passed through an activation function to serve as input for the subsequent layer [25]. Furthermore, two-dimensional subsampling layers use non-trainable kernels or windows to reduce input features. This process often reduces the number of features and decreases the network’s dependence on position [26]. The two most commonly used subsampling types are max pooling and average pooling. Both methods compute the maximum or average values within each kernel to be included in the resulting feature map. The feature map size for subsampling layers is determined in a similar manner as for convolution layers, ensuring a consistent approach to feature map creation. Some implementations even introduce specific trainable parameters into these layers to facilitate the learning process [25].
CNNs are the sole example of fully connected layers. These layers are typically located in the final few layers of most CNNs, following numerous convolutions and subsampling processes. A fully connected layer consists of a loss function, an activation function, an output layer, and several hidden layers. Their operations involve multiplying inputs by trainable weight vectors with a trainable bias and aggregating the results [25]. The output of these layers is transmitted via activation functions that resemble those in convolution layers. Dropout regularization, a technique used by CNNs to prevent overfitting, randomly assigns a value of 0 to a neuron during the backpropagation and forward learning stages. This simple technique ensures the model’s capacity for generalization by preventing the neural network from overfitting.
AlexNet, an innovative CNN architecture, has made significant advancements in deep learning and image recognition [27]. Its remarkable performance was demonstrated when it won the 2012 ImageNet Large Scale Visual Recognition Challenge [28]. This achievement marked a pivotal moment in neural networks’ ability to identify objects within images accurately. AlexNet comprises eight layers: five convolutional layers and three fully connected layers. One significant breakthrough was the use of rectified linear units (ReLU) as activation functions, which significantly sped up both training and convergence. Furthermore, adding dropout regularization improved the network’s performance by preventing overfitting. AlexNet’s sophisticated architecture and ground-breaking accomplishments provided the impetus for later advances in deep learning, especially in image classification problems [28]. The architecture of AlexNet is depicted in Table 1.
The AlexNet CNN framework, used for extracting horse-and-buggy crashes from crash diagrams as outlined in Section 2.2.1, Section 2.2.2 and Section 2.2.3 and depicted in Figure 1, represents one pillar of this study. In the following subsections, the focus shifts to crash narratives. Section 2.2.4 and Section 2.2.5 describe the integration of NLP techniques and logistic regression for analyzing textual data, offering complementary insights into the factors contributing to horse-and-buggy crashes.

2.2.4. Natural Language Processing (NLP) Techniques

Many textual databases are available within transportation safety, encompassing crash data narratives containing descriptions written by police officers and consumer complaints detailing vehicle defects. Despite the wealth of valuable information contained in crash narratives, only a limited number of researchers have integrated this type of data into their studies (e.g., [6,11,13,15,29]). However, the potential of NLP and text mining techniques in analyzing unstructured data is immense. These techniques, including thematic analysis, content analysis, supervised modeling, unsupervised modeling, and NLP [30], have found widespread application in the construction and building industry, where the primary goal is to program computers to process, understand, interpret, and manipulate human language [31,32]. NLP finds utility in diverse tasks, fields, and industries, such as sentiment analysis, text classification, and question answering. By leveraging these techniques, we can automate the content analysis of a large volume of crash narratives in the field of traffic safety.
As increasingly sophisticated qualitative analysis procedures for unstructured data have rapidly evolved [33], it becomes evident that unstructured data often exist in disorganized formats and offers minimal insight unless indexed and stored systematically. By systematically analyzing crash narratives related to horse-and-buggy incidents, new knowledge can be derived to investigate associations between their key contributing factors. Consequently, a comprehensive understanding of crash scenarios can be attained by elucidating the associations between various factors pertinent to horse-and-buggy crashes, a feat challenging to accomplish solely with structured crash data.
The modeling process commences with compiling a list of significant words from the narratives. Before this compilation, the narratives undergo preprocessing, as is customary in NLP. This preprocessing involves several steps: converting all text to lowercase, eliminating specific words such as commonly occurring acronyms in crash databases, and creating a personalized list of “stop words”, including standard terms, punctuation marks, and extraneous characters that may not contribute meaningful information. These stop-word lists are available in various Python toolkits, and in this study, the stop-word list obtained from the Natural Language Toolkit (NLTK) was adapted and utilized. Importantly, lemmatization is also a part of this process. It involves identifying the base word or grouping similar terms, thus ensuring related words are not considered separately, a crucial aspect in NLP modeling. Furthermore, analyzing trigrams in narratives associated with horse-and-buggy crashes offers additional insights. Trigrams analyze sequences of three words that commonly occur together. The Python NLTK collocations toolkit was employed to conduct the trigram analysis [34].
A critical data analysis phase is “feature extraction”, which involves reducing the original data’s dimensionality to a more manageable set of features while preserving essential information [35,36]. This investigation implements two methodologies, n-gram and feature selection, during the feature extraction phase. N-gram analysis seeks to identify sequential patterns in text data by examining sequences of “n” consecutive words or characters. By extracting these linguistic units, which range from individual words (unigrams) to pairings of words (bigrams) and onward, n-gram analysis increases the model’s ability to capture semantic meaning and context accurately [37]. The objective of feature selection is to identify and retain the most informative and discriminative features from the original dataset while simultaneously discarding redundant or extraneous ones [35]. Feature selection enhances the interpretability of the model and streamlines the analysis process by prioritizing features that contribute the most to the model’s predictive performance. These methods, with their primary aim to enhance the data analysis’s accuracy and efficacy, work towards identifying and collecting the dataset’s most pertinent and significant features [35,36].
The NLP techniques employed in this study go beyond standard applications by being specifically tailored to analyze crash narratives related to horse-and-buggy crashes. The preprocessing steps were customized to address the unique language structure of crash narratives, incorporating case folding, tokenization, stop-word removal, and stemming strategies. Furthermore, integrating NLP with AlexNet-based image processing of crash diagrams introduces a novel methodological approach, combining textual and visual data to understand crash scenarios comprehensively. This integrated framework provides a new perspective on identifying and analyzing contributing factors in transportation safety.

2.2.5. Logistic Regression

One of the machine learning algorithms used for classification tasks is logistic regression. This algorithm predicts the probability of the target class based on the input features. The fundamental formula for binary classification using logistic regression is illustrated in Equation (1):
P ( Y = 1 | X ) = 1 1 + e − ( β 0 + β 1 X )
where:
  • P(Y= 1|X) is the probability of the outcome being 1 given the value of the independent variable X.
  • β0 is the intercept term.
  • β1 is the coefficient associated with the independent variable X.
In this study, the severity outcome of horse-and-buggy crashes, designated as the dependent variable (1 = minor/no injury; 2 = fatal/serious injury), was analyzed and classified using logistic regression. Given one or more predictor variables, logistic regression predicts the probability of a binary outcome. The predictor variables are the numerical features derived from the preprocessed text data (crash narratives). Text vectorization converted the preprocessed text data into these numerical features.
The “Count-Vectorizer” technique, a robust and reliable tool in NLP, transformed the preprocessed text extracted from crash narratives into a precise numerical representation. This representation, suitable for input into the logistic regression model, ensured accurate and reliable results. Subsequently, the logistic regression model was trained using the numerical features derived from the text data and their corresponding severity outcome labels.
Additionally, the logistic regression model’s performance evaluation encompassed a comprehensive set of measures, including accuracy, precision, recall, and F1-score. The model’s generalization performance was rigorously assessed using the cross-validation technique, employing 5-fold cross-validation. This thorough evaluation ensured the validity and reliability of our findings. A more precise comprehension of this methodology is illustrated in Figure 2.

3. Results and Discussion

3.1. AlexNet-CNN Model for Crash Diagrams Processing

From the four-year study period database, we meticulously extracted and analyzed only those crashes involving horse and buggy. These were then further scrutinized using the robust AlexNet CNN architecture, a method known for its accuracy and reliability, resulting in a comprehensive dataset of 125 crashes (21% of the rural pedestrian crashes). Of these 125 crashes, 71 were classified as “Minor/No Injury”, and 54 were classified as “Fatal/Serious Injury”. These crashes were merged with their respective narratives based on the unique crash ID. This research, with its focus on crash severity outcomes, demonstrates our commitment to rigorous and thorough research.
The “Minor/No Injury” category encompasses individuals who have endured minor injuries in a horse-and-buggy crash, each with their unique narrative. In stark contrast, the “Fatal/Serious Injury” category represents the heart-wrenching tales of riders who have suffered severe injuries or even lost their lives in these crashes. The independent variables in this study are the collections of word corpora from these crash narratives. The following section presents the AlexNet CNN classification results related to the horse-and-buggy crash database, shedding light on the human experiences behind these statistics.
An essential aim of this research is to evaluate the efficacy of the AlexNet CNN-trained model. The dataset consists of 609 diverse rural pedestrian crash diagrams. Horse-and-buggy crash type has been identified and classified as a binary variable (horse-and-buggy crash = 1, or 0 if not). The K-fold cross-validation method was applied, with five determined as the value of k. The investigation was conducted on a GPU server (Kaggle personal notebook) utilizing Python 3.10.0 and the FastAI platform to develop and evaluate our approach [38].
This research established a consistent training environment for the AlexNet CNN architecture. Two learning rates of 0.01 and 0.001 were compared as a controlled magnitude of each step during the optimization process [39,40]. Additionally, the model’s performance was evaluated using accuracy and F-score as the principal metrics, allowing us to determine the percentage of correctly classified outputs [39,40,41]. Accuracy denotes the proportion of horse-and-buggy crashes that are classified correctly. It includes true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN) [42]. Equation (1) denotes accuracy. The F-score measures the accuracy of positive predictions and the ability to detect all positive cases. It combines recall (the capability to identify all positive cases) and precision (the accuracy of positive predictions) into a single value. By effectively balancing these two metrics, the F-score provides an integrated performance metric [42]. A higher value on a scale of 0 to 1 signifies superior performance, making it a valuable tool for assessing the efficacy of classification models [41]. Equation (3) denotes the F-score.
A c c u r a c y = T P + T N T P + T N + F P + F N
F - S c o r e = 2 × R e c a l l × P r e c i s i o n R e c a l l + P r e c i s i o n
A 5-fold cross-validation methodology for the AlexNet CNN model was employed to ensure an accurate and consistent assessment. Using this approach, the dataset was partitioned into five discrete subsets. Each subset was used as a validation set in turn while the model was trained on the remaining four subsets. After five iterations, each fold was used as the validation set once. This comprehensive evaluation of the model’s effectiveness across different subgroups was achieved by averaging the outcomes of these five iterations. The systematic application of this methodology effectively reinforced the reliability of our model in detecting horse-and-buggy crashes and allowed us to draw strong inferences about the efficiency of the AlexNet CNN architecture, particularly for transportation safety research. Table 2a presents the training accuracy and F-score results of the AlexNet CNN model, while Table 2b presents the validation accuracy and F-score results.
When training deep learning models like AlexNet convolutional neural networks (CNNs), the choice of learning rate is of utmost importance. The learning rate determines how much the model’s weights change for the loss gradient throughout each training iteration. A well-chosen learning rate can lead to efficient convergence and high performance, while a poorly chosen one can result in slow convergence or divergence, significantly impacting the model’s performance.
In this study, two different learning rates (0.01 and 0.001) were applied to train the AlexNet CNN model, and their impacts on model performance were analyzed as presented in Table 2a,b. The results showed that the learning rate of 0.001 led to better performance than the learning rate of 0.01. This finding suggests that a lower learning rate can lead to more accurate and efficient model training, which is crucial for applications where model performance is a priority.

3.2. NLP Techniques for Crash Narratives Analysis

The word cloud for the minor/no injury outcomes reveals several vital terms frequently appearing in crash narratives associated with less severe incidents. As shown in Figure 3a, common unigrams in this category include words such as “buggy”, “straight”, “front”, “corner”, “driver”, “passenger”, etc. These terms suggest that the crashes described in this category often involve minor impacts and lower levels of harm. In contrast, the word cloud for severe/fatal injury outcomes, depicted in Figure 3b, displays unigrams indicative of more severe incidents. Common terms in this category include “right”, “driver”, “front”, “center”, “straight”, “corner”, “unit”, “impact”, etc.
While the unigram analysis provides preliminary indicators, it alone cannot comprehensively determine the risk factors associated with each severity level of horse-and-buggy crashes. Therefore, we extended our analysis to include trigram analysis to gain more detailed insights. The trigram count plot, as depicted in Figure 4a,b, illustrates the frequency of trigrams (sequences of three words) in the text corpus. This depiction plays a crucial role in identifying the most prevalent trigrams, offering further insights into common phrases or patterns within the text. By examining these trigrams, we can better understand the contextual nuances and specific circumstances that may contribute to the severity of crashes, thereby facilitating a more thorough risk factor assessment.
The trigram analysis for fatal/serious injury outcomes, depicted in Figure 4a, reveals phrases indicating more severe incidents. The top common meaningful trigrams in this category include “straight horse drown”, “front right corner”, “straight Amish buggy”, “road traffic night”, “fail stop acd”, “inside vehicle eat”, and “unit front center”. These trigrams provide critical insights into the nature and context of severe crashes. Conversely, the trigram analysis for minor/no injury outcomes reveals several vital phrases frequently appearing in the crash narratives. As illustrated in Figure 4b, the top common trigrams in this category include sequences such as “driver go straight”, “front right corner”, “go straight horse”, “fail yield stop”, “corner vehicle driver”, and “pass Amish buggy”. These trigrams suggest that the incidents described in this category often involve minimal impacts and negligible harm to the individuals involved. The term “acd” is an abbreviation frequently used by police officers in crash narratives to refer to “accident”, and it has been retained here to ensure consistency with the original data.
By comparing the trigrams associated with minor/no injury outcomes to those linked with fatal/severe injury outcomes, we can identify several key distinctions. It becomes evident that nighttime crashes, due to the lower visibility of horse-and-buggy and the distracted driving of vehicle drivers, are a significant contributor to fatal or severe crashes. Equally important, passing behaviors to overtake horse-and-buggy often lead to minor injury crashes. These findings underscore the urgent need to address these issues to improve rural traffic safety.
The chi-square score analysis of trigrams, a crucial statistical tool, was conducted based on the target variable indicating either fatal/serious or minor/no injury outcomes. This analysis aimed to identify trigrams that exhibit significant associations with both severity levels, thereby providing valuable insights into the field of transportation safety and injury analysis.
The trigram chi-square scores, a precise measure of statistical significance, illustrate the association between trigrams and the severity outcome (dependent variable) in the text data. The chi-square test quantifies the extent to which the observed frequency of a trigram differs from the expected frequency in the absence of an association between the trigram and the outcome, ensuring the accuracy of our findings.
Table 3 presents the top 10 significant trigrams, ordered ascendingly by their p-values. The results demonstrate trigrams with high chi-square scores, indicating their strong association with crash injury outcomes. These trigrams include distracted and careless activities inside the vehicle by drivers, failure to yield actions, especially at stop-controlled intersections, passing behaviors, turning left maneuvers, issues with road lighting during the night, and lower response of elderly drivers. All these factors are critical indicators of significant severe factors at a 95% confidence level related to horse-and-buggy crashes, providing a comprehensive understanding of the injury outcomes.
Notably, the logistic regression analysis, a robust statistical technique, was conducted to predict severity levels based on the significant trigrams we identified. These trigrams, crucial in our model, are associated with crash severity. Logistic regression models the probability of a binary outcome, such as fatal/serious or minor/no injury outcomes in this study, using multiple predictor variables.
The logistic regression model, trained with the identified trigrams as predictor variables and the severity outcome as the binary target variable (minor/no injury = 1 or fatal/serious = 2), underwent a comprehensive performance assessment. Various metrics, including accuracy, precision, recall, and F1-score, were used. This thorough evaluation provides a clear understanding of the model’s ability to classify severity levels and its overall predictive efficacy. Table 4 presents the logistic regression model’s training results with a maximum of 1000 features. To ensure the accuracy of these results, the dataset was split into 80% for training and 20% for testing. This approach is a typical practice in machine learning studies, balancing the need for sufficient data for model training with maintaining enough data for reliable performance evaluation. Figure 5 shows the confusion matrix for the testing results. The correctly classified data are represented along the diagonal, while values outside the diagonal denote classification errors made by the logistic regression model. The accuracy of the model’s validation data is 80%, indicating a notable level of predictive accuracy. Additionally, the confusion matrix in Figure 5 shows that the classification errors are marginally lower for horse-and-buggy crashes involving fatal or serious injuries compared to those involving minor or no injuries.
The logistic regression model’s coefficients offer valuable insights into the relationship between the identified trigrams and crash severity. Each coefficient quantifies the change in the log-odds of the outcome for a one-unit change in the corresponding predictor variable, with all other variables held constant. By analyzing the magnitude and direction of the coefficients, we can not only understand the degree to which each trigram contributes to predicting fatal/serious or minor/no injury outcomes but also provide practical insights for improving transportation safety. Trigrams with higher coefficients demonstrate a stronger association with the respective severity level, indicating their more significant influence on the likelihood of a severe crash occurrence. Table 5 presents the top 10 coefficient values for the logistic regression model features, clearly depicting the most influential trigrams in predicting crash severity.
The sign of the coefficient (positive or negative) is a crucial indicator of the relationship between the predictor variable and the outcome. A positive coefficient points to a direct relationship, meaning that the presence of the text trigram is linked to a higher likelihood of a severe or fatal injury crash. On the other hand, a negative coefficient suggests an inverse relationship, indicating that the presence of the text trigram is associated with a lower likelihood of the outcome being a severe or fatal injury crash.
Moreover, the statistical significance of the coefficients, determined through hypothesis testing, plays a pivotal role in identifying trigrams that are consistently linked to crash severity. Significant coefficients (p < 0.05) are a strong indication that the corresponding trigrams have a reliable statistical impact on predicting severity outcomes.
For instance, let us consider the coefficient for the text trigram “driver careless drive”. This coefficient implies that for every one-unit increase in the occurrence or presence of this trigram, the log-odds of the outcome being a fatal injury crash significantly increase by 0.2478 units. In simpler terms, this means that the odds of the outcome (a fatal injury crash) increase by approximately 28.1%.
The interaction between modern motor vehicles and horse-drawn buggies on rural roadways presents unique and urgent risks and hazards. Gorucu et al. [43] have underscored the potential dangers inherent in this interaction, emphasizing issues such as differences in speed leading to rear-end collisions and the disparity in vehicle size resulting in severe injuries to horse-drawn buggy occupants. These concerns highlight the pressing need for a comprehensive understanding of horse-and-buggy crashes’ specific characteristics and dynamics. Our findings shed new light on several critical aspects of these crashes, corroborating existing research findings and offering fresh insights. We discovered that left- or right-turning and passing or overtaking maneuvers significantly increase the risk of injuries to buggy drivers compared to motor vehicle drivers. Additionally, our findings align with the observations made by Anderson regarding the limited visibility of horse-drawn buggies, particularly at night, which further amplifies these risks [3].
Furthermore, our analysis highlights several factors contributing to the severity of horse-and-buggy crashes. Table 5 indicates that trigrams such as “driver careless drive” and “driver night road” are significantly associated with an increased likelihood of severe or fatal crashes, with positive coefficients suggesting a direct relationship. Failure-to-yield actions involving elderly drivers are also notable, as evidenced by trigrams like “older driver fail”. While this trigram is associated with less severe crashes in some contexts, the nuanced role of failure-to-yield actions highlights the complex dynamics at play in crash severity. These findings emphasize the multifaceted risks inherent in horse-and-buggy interactions on rural roadways, particularly during nighttime or due to driver carelessness.
Given the complex and multifaceted nature of the risks associated with horse-and-buggy interactions on rural roadways, our research highlights the urgent need for targeted safety interventions. By prioritizing the safety of horse-and-buggy drivers and other motorists and implementing comprehensive safety measures, we can effectively reduce the risks associated with these interactions and significantly enhance overall road safety in rural areas.

4. Conclusions

This study delved into the severity levels of horse-and-buggy crashes in rural Michigan, introducing a novel approach by leveraging the power of CNN and NLP algorithms. The CNN algorithm was used to analyze crash diagrams, while the NLP algorithm was applied to process text data from crash narratives. The analysis aimed to identify key risk factors and enhance crash severity prediction.
The findings of this study underscore the critical risks associated with horse-and-buggy crashes, particularly those linked to left- or right-turning maneuvers, which result in significantly more severe injuries to buggy drivers than motor vehicle drivers. The study also identifies key factors contributing to crash severity, including limited visibility on rural roads during nighttime due to the absence of streetlights or reflective materials on buggies and distracted or careless actions by vehicle motorists. Failure-to-yield actions, particularly involving elderly drivers, also play a notable role, as the logistic regression analysis indicates. These insights emphasize the need for targeted safety interventions, such as improved road lighting, increased use of reflective materials on buggies, and public education campaigns to mitigate distracted driving and enhance yielding behaviors.
The logistic regression analysis results provide valuable insights into the relationship between the identified trigrams and crash severity. By examining the coefficients associated with each trigram, we can ascertain their contributions to predicting fatal/serious or minor/no injury outcomes. Trigrams with higher coefficients exhibit stronger associations with their severity levels, while those with lower coefficients have less predictive power.
Addressing these issues requires the development and implementation of comprehensive safety measures. These measures should include improved signage and road markings to alert drivers to horse-and-buggy traffic, educational campaigns targeting the general public and the Amish community, and technology integration to enhance the visibility of horse-drawn buggies. However, the key to success lies in community involvement. Collaboration between local authorities, transportation agencies, and community leaders is not just important; it is the driving force behind creating and enforcing policies that ensure the safety of all road users.
Road markings and signage, in particular, are crucial components of road infrastructure, serving as a primary means of communication between road authorities and users [44]. Studies have shown that properly designed and high-quality road markings and signs positively impact road user behavior and safety outcomes, making them cost-effective safety measures. These elements guide drivers and enhance the visual information available to human drivers and perceptual vehicle technologies, ultimately reducing crash risks [44].
Research has also shown that enhanced lighting conditions and visibility significantly reduce the severity of pedestrian crashes, particularly during nighttime or low-light conditions [45,46]. Evidence from similar contexts, such as agricultural vehicle crashes, highlights the importance of educational campaigns in mitigating rural roadway risks [47]. In these contexts, targeted driver education has been identified as a critical prevention strategy, effectively addressing crash risk factors such as visibility and driving behavior [47].
Significantly, this study deepens our understanding of horse-and-buggy crash dynamics and underscores the potential of advanced image processing and AI technologies in traffic safety research. These innovative methodologies, with their ability to uncover hidden patterns and insights overlooked by traditional analysis, hold promise for generating more accurate and actionable findings. This, in turn, instills a sense of optimism and hope for safer road environments for all users.
In conclusion, this study demonstrates the potential of CNN and NLP algorithms in analyzing the severity of horse-and-buggy crashes in rural Michigan. Leveraging logistic regression with text trigrams yielded valuable insights into the factors influencing crash outcomes. Future research endeavors hold promise in exploring additional machine learning models and larger datasets, which can further enhance predictive accuracy and deepen our understanding of horse-and-buggy crashes. By harnessing advanced analytical techniques, we can develop more effective strategies to prevent severe crashes and improve road safety in rural communities.
This study employed logistic regression due to its interpretability and robustness, making it well suited for identifying actionable factors influencing crash severity. While this approach provided valuable insights, future research could explore advanced machine learning algorithms, such as random forests or gradient boosting, to assess whether they significantly improve predictive performance.
The methodologies and findings of this study not only address the specific context of horse-and-buggy crashes in rural Michigan and open pathways for broader applications. By leveraging advanced machine learning techniques, such as CNN and NLP, we demonstrate the potential of integrating image processing with text mining for analyzing crash data in unstructured formats. This approach promises to expand research into other regions where horse-drawn carriages or similar modes of transport are still standard, such as parts of Europe, Asia, and the developing world. The ability to systematically extract and analyze data from crash narratives and diagrams can support safety interventions tailored to the unique dynamics of these transport modes.
Furthermore, our methodologies can be adapted to investigate safety challenges involving other vulnerable road users (VRUs), cyclists, pedestrians, and micro-mobility users who share roadways with motor vehicles [48,49]. Emerging technologies, such as automated data extraction and predictive modeling, can be employed to understand crash patterns and risk factors across various contexts, offering insights into effective interventions.
Future research could benefit from incorporating innovative methodologies that leverage text mining and machine learning for safety analysis. Studies have demonstrated the utility of NLP and topic modeling in understanding perceived safety concerns, as seen in cyclist safety research utilizing crash narratives and self-reported data [50]. Additionally, advancements in automated monitoring systems, such as stereoscopic techniques for infrastructure evaluation, highlight the potential for integrating diverse data sources to enhance transportation safety [51]. Moreover, using social media platforms for crowdsourced subjective safety data, analyzed through machine learning and sentiment analysis, presents a novel avenue for understanding road users’ affective states and their impact on mobility [52]. These approaches align with our efforts and suggest potential expansions to include broader data sources and emerging techniques for a more comprehensive understanding of transportation safety challenges.
Building on these advancements, future work could also explore integrating real-time data sources, such as connected vehicle systems and smart city infrastructure, to provide dynamic insights into crash risks. These systems, equipped with sensors and communication capabilities, can capture data on vehicle movements, environmental conditions, and near-miss events. Researchers could develop predictive frameworks to proactively identify and mitigate potential safety hazards by combining this information with machine learning models [53,54,55]. This direction aligns with the growing emphasis on proactive safety measures, contributing to the vision of zero road traffic fatalities.

Author Contributions

Conceptualization, B.Q.; methodology, B.Q.; software, B.Q.; validation, B.Q.; formal analysis, B.Q.; investigation, B.Q.; resources, B.Q.; data curation, B.Q.; writing—original draft preparation, B.Q.; writing—review and editing, B.Q., J.-S.O. and V.K.; visualization, B.Q.; supervision, J.-S.O. and V.K.; project administration, J.-S.O. and V.K.; funding acquisition, B.Q. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The original data presented in the study are openly available in [Michigan Traffic Crash Facts (MTCF)] at [https://www.michigantrafficcrashfacts.org/data/querytool/#q1;0;2023,2022,2021,2020;; (accessed on 1 June 2024)].

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Becklinger, N. An assessment of horse-drawn vehicle incidents from US news media reports within AgInjuryNews. Safety 2023, 9, 21. [Google Scholar] [CrossRef]
  2. Stein, R.E.; Dewalt, M.W. Prevalence of Accidents in Smaller Amish Settlements: 2015–2022. J. Plain Anabapt. Communities 2024, 4, 1–22. [Google Scholar] [CrossRef]
  3. Anderson, C. Horse and buggy crash study I: Common crash scenarios between a motor vehicle and the Amish/Old Order Mennonite horse and buggy. J. Amish Plain Anabapt. Stud. 2014, 2, 79–99. [Google Scholar] [CrossRef]
  4. Pérez-Zuriaga, A.M.; Dols, J.; Nespereira, M.; Garcia, A.; Sajurjo-de-No, A. Analysis of the consequences of car to micromobility user side impact crashes. J. Safety Res. 2023, 87, 168–175. [Google Scholar] [CrossRef]
  5. Yang, H.; Ma, Q.; Wang, Z.; Cai, Q.; Xie, K.; Yang, D. Safety of micro-mobility: Analysis of E-Scooter crashes by mining news reports. Accid. Anal. Prev. 2020, 143, 105608. [Google Scholar] [CrossRef]
  6. Kwayu, K.M.; Kwigizile, V.; Lee, K.; Oh, J.-S. Discovering latent themes in traffic fatal crash narratives using text mining analytics and network topology. Accid. Anal. Prev. 2021, 150, 105899. [Google Scholar] [CrossRef]
  7. Zhang, Y.; Lu, H.; Qu, W. Geographical detection of traffic accidents spatial stratified heterogeneity and influence factors. Int. J. Environ. Res. Public Health 2020, 17, 572. [Google Scholar] [CrossRef]
  8. Santos, K.; Dias, J.P.; Amado, C. A literature review of machine learning algorithms for crash injury severity prediction. J. Saf. Res. 2022, 80, 254–269. [Google Scholar] [CrossRef]
  9. Azhar, A.; Ariff, N.M.; Bakar, M.A.A.; Roslan, A. Classification of driver injury severity for accidents involving heavy vehicles with decision tree and random forest. Sustainability 2022, 14, 4101. [Google Scholar] [CrossRef]
  10. Muhammad, I.; Liu, L.; Muhammad, Z.; Arshad, J. A comparative study of machine learning classifiers for injury severity prediction of crashes involving three-wheeled motorized rickshaw. Accid. Anal. Prev. 2021, 154, 106094. [Google Scholar] [CrossRef]
  11. Nayak, R.; Piyatrapoomi, N.; Weligamage, J. Application of text mining in analysing road crashes for road asset management. In Engineering Asset Lifecycle Management, Proceedings of the 4th World Congress on Engineering Asset Management (WCEAM 2009), Athens, Greece, 28–30 September 2009; Springer: London, UK, 2010; pp. 49–58. [Google Scholar]
  12. Rahman, M.; Kockelman, K.M.; Perrine, K.A. Investigating risk factors associated with pedestrian crash occurrence and injury severity in Texas. Traffic Inj. Prev. 2022, 23, 283–289. [Google Scholar] [CrossRef] [PubMed]
  13. Kwayu, K.M.; Kwigizile, V.; Zhang, J.; Oh, J.-S. Semantic N-gram feature analysis and machine learning–based classification of drivers’ hazardous actions at signal-controlled intersections. J. Comput. Civ. Eng. 2020, 34, 4020015. [Google Scholar] [CrossRef]
  14. Athuraliya, C.D.; Gunasekara, M.K.H.; Perera, S.; Suhothayan, S. Real-time natural language processing for crowdsourced road traffic alerts. In Proceedings of the 2015 Fifteenth International Conference on Advances in ICT for Emerging Regions (ICTer), Colombo, Sri Lanka, 24–26 August 2015; pp. 58–62. [Google Scholar]
  15. Gao, L.; Wu, H. Verb-based text mining of road crash report. In Proceedings of the 92nd Annual Meeting of the Transportation Research Board, Washington, DC, USA, 13–17 January 2013. [Google Scholar]
  16. Hou, L.; Chen, H.; Zhang, G.; Wang, X. Deep learning-based applications for safety management in the AEC industry: A review. Appl. Sci. 2021, 11, 821. [Google Scholar] [CrossRef]
  17. Nixon, M.; Aguado, A. Feature Extraction and Image Processing for Computer Vision; Academic Press: Cambridge, MA, USA, 2019. [Google Scholar]
  18. O’Mahony, N.; Campbell, S.; Carvalho, A.; Harapanahalli, S.; Hernandez, G.V.; Krpalkova, L.; Riordan, D.; Walsh, J. Deep learning vs. traditional computer vision. In Advances in Computer Vision, Proceedings of the 2019 Computer Vision Conference (CVC), Las Vegas, NV, USA, 25–26 April 2019; Springer: Cham, Switzerland, 2020; Volume 1, pp. 128–144. [Google Scholar]
  19. Alzubaidi, L.; Zhang, J.; Humaidi, A.J.; Al-Dujaili, A.; Duan, Y.; Al-Shamma, O.; Santamaría, J.; Fadhel, M.A.; Al-Amidie, M.; Farhan, L. Review of deep learning: Concepts, CNN architectures, challenges, applications, future directions. J. Big Data 2021, 8, 1–74. [Google Scholar] [CrossRef]
  20. Yuan, Z.-W.; Zhang, J. Feature extraction and image retrieval based on AlexNet. In Proceedings of the Eighth International Conference on Digital Image Processing (ICDIP 2016), Chengu, China, 20–22 May 2016; Volume 10033, pp. 65–69. [Google Scholar]
  21. MTCF. Michigan Traffic Crash Facts (MTCF). 2024. Available online: https://www.michigantrafficcrashfacts.org/ (accessed on 1 June 2024).
  22. Calhoun, B.C.; Uselman, H.; Olle, E.W. Development of Artificial Intelligence Image Classification Models for Determination of Umbilical Cord Vascular Anomalies. J. Ultrasound Med. 2024, 43, 881–897. [Google Scholar] [CrossRef]
  23. Abbas, R.F. Review on some methods used in image restoration. Int. Multidiscip. Res. J. 2020, 10, 13–16. [Google Scholar] [CrossRef]
  24. Samir, S.; Emary, E.; El-Sayed, K.; Onsi, H. Optimization of a pre-trained AlexNet model for detecting and localizing image forgeries. Information 2020, 11, 275. [Google Scholar] [CrossRef]
  25. Chen, H.-C.; Widodo, A.M.; Wisnujati, A.; Rahaman, M.; Lin, J.C.-W.; Chen, L.; Weng, C.-E. AlexNet convolutional neural network for disease detection and classification of tomato leaf. Electronics 2022, 11, 951. [Google Scholar] [CrossRef]
  26. Al Tawil, A.; Shaban, A.; Almazaydeh, L. A comparative analysis of convolutional neural networks for breast cancer prediction. Int. J. Electr. Comput. Eng. 2024, 14, 3406–3414. [Google Scholar] [CrossRef]
  27. Fang, A.; Kornblith, S.; Schmidt, L. Does progress on ImageNet transfer to real-world datasets? Adv. Neural Inf. Process. Syst. 2024, 36, 25050–25080. [Google Scholar]
  28. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. ImageNet classification with deep convolutional neural networks. Commun. ACM 2017, 60, 84–90. [Google Scholar] [CrossRef]
  29. Arteaga, C.; Paz, A.; Park, J. Injury severity on traffic crashes: A text mining with an interpretable machine-learning approach. Saf. Sci. 2020, 132, 104988. [Google Scholar] [CrossRef]
  30. Banks, G.C.; Woznyj, H.M.; Wesslen, R.S.; Ross, R.L. A review of best practice recommendations for text analysis in R (and a user-friendly app). J. Bus. Psychol. 2018, 33, 445–459. [Google Scholar] [CrossRef]
  31. Manning, C.; Schutze, H. Foundations of Statistical Natural Language Processing; MIT Press: Cambridge, MA, USA, 1999. [Google Scholar]
  32. Manning, C.D.; Surdeanu, M.; Bauer, J.; Finkel, J.R.; Bethard, S.; McClosky, D. The Stanford CoreNLP natural language processing toolkit. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics: System Demonstrations, Baltimore, MD, USA, 23–24 June 2014; pp. 55–60. [Google Scholar]
  33. Feldman, R.; Sanger, J. The Text Mining Handbook: Advanced Approaches in Analyzing Unstructured Data; Cambridge University Press: Cambridge, MA, USA, 2007. [Google Scholar]
  34. NLTK. Natural Language Toolkit. 2024. Available online: https://www.nltk.org/ (accessed on 1 June 2024).
  35. Hadi, Z.; Sunyoto, A. Detecting Fake Reviews Using N-gram Model and Chi-Square. In Proceedings of the 2023 6th International Conference on Information and Communications Technology (ICOIACT), Yogyakarta, Indonesia, 10 November 2023; pp. 454–458. [Google Scholar]
  36. Fachrurrozi, S.; Shidik, G.F.; Fanani, A.Z.; Al Zami, F. Increasing Accuracy of Support Vector Machine (SVM) By Applying N-Gram and Chi-Square Feature Selection for Text Classification. In Proceedings of the 2021 International Seminar on Application for Technology of Information and Communication (iSemantic), Virtual, 18–19 September 2021; pp. 42–47. [Google Scholar]
  37. Cavnar, W.B.; Trenkle, J.M. N-gram-based text categorization. In Proceedings of the SDAIR-94, 3rd Annual Symposium on Document Analysis and Information Retrieval, Las Vegas, NV, USA, 11–13 April 1994; Volume 161175, p. 14. [Google Scholar]
  38. Kaggle. 2024. Available online: https://www.kaggle.com (accessed on 15 June 2024).
  39. Wang, S.-H.; Xie, S.; Chen, X.; Guttery, D.S.; Tang, C.; Sun, J.; Zhang, Y.-D. Alcoholism identification based on an AlexNet transfer learning model. Front. Psychiatry 2019, 10, 454348. [Google Scholar] [CrossRef]
  40. Kalaiarasi, P.; Rani, P.E. A comparative analysis of AlexNet and GoogLeNet with a simple DCNN for face recognition. In Advances in Smart System Technologies: Select Proceedings of ICFSST 2019; Springer: Singapore, 2021; pp. 655–668. [Google Scholar]
  41. Singh, I.; Goyal, G.; Chandel, A. AlexNet architecture based convolutional neural network for toxic comments classification. J. King Saud Univ. Inf. Sci. 2022, 34, 7547–7558. [Google Scholar] [CrossRef]
  42. Schonlau, M.; Zou, R.Y. The random forest algorithm for statistical learning. Stata J. 2020, 20, 3–29. [Google Scholar] [CrossRef]
  43. Gorucu, S.; Murphy, D.J.; Kassab, C. Injury risks for on-road farm equipment and horse and buggy crashes in Pennsylvania: 2010–2013. Traffic Inj. Prev. 2017, 18, 286–292. [Google Scholar] [CrossRef]
  44. Babić, D.; Babić, D.; Fiolic, M.; Ferko, M. Road markings and signs in road safety. Encyclopedia 2022, 2, 1738–1752. [Google Scholar] [CrossRef]
  45. Batouli, G.; Guo, M.; Janson, B.; Marshall, W. Analysis of pedestrian-vehicle crash injury severity factors in Colorado 2006–2016. Accid. Anal. Prev. 2020, 148, 105782. [Google Scholar] [CrossRef]
  46. Houten, R.V.; Kwigizile, V.; Oh, J.S.; Mwende, S.; Qawasmeh, B. Effective Pedestrian/Non-Motorized Crossing Enhancements Along Higher Speed Corridors. No. SPR-1734; Michigan Department of Transportation, Research Administration: Lansing, MI, USA, 2023. [Google Scholar]
  47. Franklin, R.C.; King, J.C.; Riggs, M. A systematic review of large agriculture vehicles use and crash incidents on public roads. J. Agromed. 2020, 25, 14–27. [Google Scholar] [CrossRef]
  48. Qawasmeh, B.; Oh, J.S.; Kwigizile, V. Micro-Mobility Safety Assessment: Analyzing Factors Influencing the Micro-Mobility Injuries in Michigan by Mining Crash Reports. Future Transp. 2024, 4, 1580–1601. [Google Scholar] [CrossRef]
  49. Qawasmeh, B.S. Safety Assessment for Vulnerable Road Users Using Automated Data Extraction with Machine-Learning Techniques. Ph.D. Thesis, Western Michigan University, Kalamazoo, MI, USA, 2024. [Google Scholar]
  50. Janstrup, K.H.; Kostic, B.; Møller, M.; Rodrigues, F.; Borysov, S.; Pereira, F.C. Predicting injury-severity for cyclist crashes using natural language processing and neural network modelling. Saf. Sci. 2023, 164, 106153. [Google Scholar] [CrossRef]
  51. Wang, J.; Zhang, S.; Guo, H.; Tian, Y.; Liu, S.; Du, C.; Wu, J. Stereoscopic monitoring of transportation infrastructure. Autom. Constr. 2024, 164, 105472. [Google Scholar] [CrossRef]
  52. Abedi, M.M.; Sacchi, E. A machine learning tool for collecting and analyzing subjective road safety data from Twitter. Expert Syst. Appl. 2024, 240, 122582. [Google Scholar] [CrossRef]
  53. Qawasmeh, B. Enhancing Work Zone Safety: Evaluating Static Merge Strategies Through Microscopic Traffic Simulation. Open Transp. J. 2024, 18, e18744478330254. [Google Scholar] [CrossRef]
  54. Qawasmeh, B.; Oh, J.-S.; Kwigizile, V.; Qawasmeh, D.; Al Tawil, A.; Aldalqamouni, A. Analyzing Daytime/Nighttime Pedestrian Crash Patterns in Michigan Using Unsupervised Machine Learning Techniques and their Potential as a Decision-Making Tool. Open Transpl. J. 2024, 18, e26671212352718. [Google Scholar] [CrossRef]
  55. Al Tawil, A.; Almazaydeh, L.; Qawasmeh, D.; Qawasmeh, B.; Alshinwan, M.; Elleithy, K. Comparative Analysis of Machine Learning Algorithms for Email Phishing Detection Using TF-IDF, Word2Vec, and BERT. Comput. Mater. Contin 2024, 81, 3395–3412. [Google Scholar] [CrossRef]
Safety 11 00001 g001
Figure 1. Data extraction framework.
Figure 1. Data extraction framework.
Safety 11 00001 g001
Safety 11 00001 g002
Figure 2. Crash narratives analysis flow.
Figure 2. Crash narratives analysis flow.
Safety 11 00001 g002
Safety 11 00001 g003
Figure 3. Unigram displayed as word clouds. (a) for minor/no injury; (b) for fatal/serious injury.
Figure 3. Unigram displayed as word clouds. (a) for minor/no injury; (b) for fatal/serious injury.
Safety 11 00001 g003
Safety 11 00001 g004
Figure 4. (a) Trigram word frequency for fatal/serious injury outcome for horse-and-buggy crashes. (b) Trigram word frequency for minor/no injury outcome for horse-and-buggy crashes.
Figure 4. (a) Trigram word frequency for fatal/serious injury outcome for horse-and-buggy crashes. (b) Trigram word frequency for minor/no injury outcome for horse-and-buggy crashes.
Safety 11 00001 g004
Safety 11 00001 g005
Figure 5. The confusion matrix of the logistic regression model.
Figure 5. The confusion matrix of the logistic regression model.
Safety 11 00001 g005
Table 1. AlexNet convolutional neural network architecture [28].
Table 1. AlexNet convolutional neural network architecture [28].
Layer TypeOutput ShapeNumber of FiltersKernel SizeStride
Input227 × 227 × 3---
Convolutional 155 × 55 × 969611 × 114
Max Pooling 127 × 27 × 96-3 × 32
Convolutional 227 × 27 × 2562565 × 51
Max Pooling 213 × 13 × 256-3 × 32
Convolutional 313 × 13 × 3843843 × 31
Convolutional 413 × 13 × 3843843 × 31
Convolutional 513 × 13 × 2562563 × 31
Max Pooling 36 × 6 × 256-3 × 32
Fully Connected 14096---
Fully Connected 24096---
Fully Connected 31000---
Table 2. (a) The AlexNet CNN model’s training accuracy and F-score values for 10-epochs. (b) The AlexNet CNN model’s validation accuracy and F-score values for 10-epochs.
Table 2. (a) The AlexNet CNN model’s training accuracy and F-score values for 10-epochs. (b) The AlexNet CNN model’s validation accuracy and F-score values for 10-epochs.
(a)
#EpochLearning RateFoldAccuracyPrecisionRecallF-Score
100.0110.78320.83620.76220.7975
20.79240.84280.76720.8032
30.80150.87320.73650.7990
40.80350.82630.79930.8126
50.81260.82230.81650.8194
Mean0.800.840.780.81
100.00110.84630.88460.82560.8541
20.85540.90030.80490.8499
30.83280.89540.77260.8295
40.89900.84600.92240.8825
50.86280.82430.89630.8588
Mean0.860.870.840.85
(b)
#EpochLearning RateFoldAccuracyPrecisionRecallF-Score
100.0110.79460.84730.75920.8008
20.80030.88460.74730.8102
30.79380.85320.75870.8032
40.81990.83240.80910.8206
50.82350.86380.78370.8218
Mean0.810.860.770.81
100.00110.83750.91050.77280.8360
20.84750.90350.78310.8390
30.85830.88540.81660.8496
40.90230.89940.89820.8988
50.92060.84891.00820.9217
Mean0.860.870.870.89
Table 3. Trigrams chi-square scores.
Table 3. Trigrams chi-square scores.
# FeatureTop Features (Trigram)Chi-Square Scorep-Value
1Activity inside vehicle5.5240.0187
2Distracted activity inside5.5240.0187
3Fail yield stop5.0690.0244
4Driver overtaking passing4.4140.0356
5Stop sign older4.3440.0371
6Rear left corner4.3440.0371
7Inside vehicle eating4.1430.0418
8Driver night road4.1430.0418
9Driver careless drive4.1430.0418
10Older driver fail4.1430.0418
Table 4. Logistic regression model training results.
Table 4. Logistic regression model training results.
Severity OutcomePrecisionRecallF1-ScoreModel Accuracy
Minor/No Injury = 10.701.000.820.83
Fatal/Serious Injury = 21.000.820.90
Table 5. Top 10 features coefficients of the logistic regression model.
Table 5. Top 10 features coefficients of the logistic regression model.
Features
(from the Trigrams)		Coefficientp-ValueSignificant at 95% CL
Stop sign older−0.3210240.01245Significant
Fail stop acd−0.3183330.00863Significant
Driver overtaking passing−0.3092560.02653Significant
Older driver fail−0.2267450.02435Significant
Driver night road0.2454400.00258Significant
Driver careless drive0.2475150.00023Significant
Driver going straight0.2941940.36525Not Significant
Activity inside vehicle0.3008430.00385Significant
Distracted activity inside0.3008430.00385Significant
Going straight horse0.3460140.25314Not Significant
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Qawasmeh, B.; Oh, J.-S.; Kwigizile, V. Investigating Injury Outcomes of Horse-and-Buggy Crashes in Rural Michigan by Mining Crash Reports Using NLP and CNN Algorithms. Safety 2025, 11, 1. https://doi.org/10.3390/safety11010001

AMA Style

Qawasmeh B, Oh J-S, Kwigizile V. Investigating Injury Outcomes of Horse-and-Buggy Crashes in Rural Michigan by Mining Crash Reports Using NLP and CNN Algorithms. Safety. 2025; 11(1):1. https://doi.org/10.3390/safety11010001

Chicago/Turabian Style

Qawasmeh, Baraah, Jun-Seok Oh, and Valerian Kwigizile. 2025. "Investigating Injury Outcomes of Horse-and-Buggy Crashes in Rural Michigan by Mining Crash Reports Using NLP and CNN Algorithms" Safety 11, no. 1: 1. https://doi.org/10.3390/safety11010001

APA Style

Qawasmeh, B., Oh, J. -S., & Kwigizile, V. (2025). Investigating Injury Outcomes of Horse-and-Buggy Crashes in Rural Michigan by Mining Crash Reports Using NLP and CNN Algorithms. Safety, 11(1), 1. https://doi.org/10.3390/safety11010001

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop