Machine Learning in Assessing Canine Bone Fracture Risk: A Retrospective and Predictive Approach

Abstract

In the ever-evolving world of veterinary care, the occurrence of bone fractures in canines poses a common and complex problem, especially in extra-small breeds and dogs that are less than 1 year old. The objective of this research is to fill a gap in predicting the risk of canine bone fractures. A machine learning method using a random forest classifier was constructed. The algorithm was trained on a dataset consisting of 2261 cases that included several factors, such as canine age, gender, breed, and weight. The performance of the algorithm was assessed by examining its capacity to forecast the probability of fractures occurring. The findings of our study indicate that the tool has the capability to provide dependable predictions of fracture risk, consistent with our extensive dataset on fractures in canines. However, these results should be considered preliminary due to the limited sample size. This discovery is a crucial tool for veterinary practitioners, allowing them to take preventive measures to manage and prevent fractures. In conclusion, the implementation of this prediction tool has the potential to significantly transform the quality of care in the field of veterinary medicine by enabling the detection of patients at high risk, hence enabling the implementation of timely and customized preventive measures.

1. Introduction

The field of veterinary medicine places a significant emphasis on the well-being of canines, as they have important roles in the lives of people [1]. Based on this understanding, our study underscores the need to allocate the same priority to canine health research as is typically given in human medicine. In response to this need, we have undertaken an innovative initiative to create and assess an algorithm using machine learning that is specially tailored to forecast the probability of bone fractures in canines. Given the limited sample size and the exploratory nature of this study, the results should be considered preliminary, providing a foundation for future research with larger datasets. The primary objective of this new method is to provide veterinarians, breeders, and dog owners with a tool for comprehensive risk assessments for specific bones, thereby enabling prompt interventions to mitigate the occurrence of fractures. The described technical progress introduces a data-driven approach to the prevention of bone fractures in canines and highlights the need to use machine learning technology for various species. The creation of multiple subgroups based on size, age, and fracture locations and types is essential for capturing the heterogeneity within the canine population. This stratification ensures that the model can account for variations in fracture risks that are specific to different demographic and physiological groups, thereby enhancing the precision and applicability of the predictive tool across diverse canine profiles.

Compared to the progress made in veterinary medicine, machine learning technologies have shown significant achievements in the area of human health, including the prediction of the likelihood of bone fractures. A fracture prediction model was established in a study that focused on a community-based cohort, demonstrating the possibility of tailored medical treatments [2]. In addition, a separate research project led to the development of four machine learning models designed to evaluate the likelihood of fractures in individuals with osteoporosis. This demonstrates the wide range of uses for these technologies [3]. In addition to these advancements, well-established tools such as FRAX^®, Garvan, and QFracture^® are still crucial for evaluating the risk of human fractures. This underscores the need for a complete strategy to manage health risks [4]. The comparison between human and veterinary medical progress highlights the significance of using machine learning technology to improve healthcare and preventative approaches in all areas of medicine, thereby emphasizing the mutual dependence of human and animal well-being.

Bone fractures are a prevalent health issue seen in canines and result from a multitude of contributing factors. Dogs frequently experience car accidents, falls [5], and diverse injuries inside the home, especially during training sessions [6] or as a result of being assaulted by another dog [7,8]. Nevertheless, it is essential to consider not just the fracture as a consequence but also the potential etiology of the fracture; we should ask ourselves how we could have potentially prevented the animal from having experienced a fracture [9]. These scenarios have the potential to apply significant external loads, including bending, compression, shearing, and torsional forces [10], to bones, often surpassing the bone’s ultimate strength and leading to damage.

The most frequently reported form of fracture in canines is a long-bone fracture. The humerus, radius, ulna, femur, tibia, and fibula are the most often injured bones [11], and fractures in almost the same bones constituted the results of another group of researchers [5]; in addition, they diagnosed pelvic/sacrum fractures. Rao observed a higher frequency of fractures in the femur, tibia, and radius in his dissertation [12]. Furthermore, metacarpal bone fractures were mentioned in a study by Gomaa et al. [13].

A wide variety of canine breeds exist, with each breed possessing distinct genetic and physical characteristics that may contribute to the occurrence of bone fractures. Mongrel dogs have been reported to have a notable prevalence of long-bone fractures, specifically in the tibia and fibula, mostly arising from incidents involving falls from elevated positions [8]. After a scientific literature review, it was indicated that there are significant disparities between large and small canine breeds. The incidence of thoracic limb fractures, particularly those affecting the radius and ulna, is higher in small breeds such as Pinschers, Poodles, and Brazilian Terriers [14]. Additionally, these fractures are often seen in miniature breeds weighing less than 3 kg, such as Chihuahuas and Yorkshire Terriers [15].

Fractures in large breeds exhibit variations; greyhounds are prone to fractures affecting the accessory carpal bones [16], while elbow coronoid process fractures are frequently seen in Labrador Retrievers and Golden Retrievers [17]. The fractures in question have been discovered in many breeds, including Rottweilers, Bernese Mountain Dogs [18], German Shepherds, and Spaniels. The inclination observed in these breeds may be ascribed to several variables, including genetic predisposition, body size, and levels of physical activity [19].

The age at which dogs are most likely to have bone fractures varies across different studies and is influenced by characteristics such as the breed, size, and lifestyle. Nevertheless, several studies have emphasized certain age categories in which bone fractures are more common: Lee et al. [20] observed that dogs often suffer from long-bone fractures between the ages of 1 and 3 years. This suggests that younger dogs, which are more physically active, may be more susceptible to such fractures due to their higher physical activity and perhaps less-established bone strength. In a study by Smith et al. [21], the authors observed that the age at which dogs usually have bone fractures was 4 months, with a range spanning from 2 to 10 months. This finding underscores the vulnerability of young puppies to fractures. In contrast, an investigation revealed that canines aged 6 to 9 years constituted 23.38% of fracture events, suggesting that older canines are equally susceptible to fractures [22].

Gender-based studies on dogs concerning fractures are limited; nevertheless, it is important to acknowledge the variations between males and females. Abo-Soliman et al. [23] demonstrated a significant difference in the frequency of appendicular bone fractures between male and female dogs, where the majority of the male canines (over 60%) exhibited fractures. R. Jain et al. [24] confirmed these findings by demonstrating that male dogs have a higher incidence of bone fractures. These findings may be attributed to variations in behavior, as male canines may exhibit a greater propensity for participating in behaviors that heighten the likelihood of fractures, such as engaging in more aggressive play, wandering, and fighting, in comparison to their female counterparts. In our previously published research, it was discovered that female canines have a potential susceptibility to osteoporosis, which also needs to be assessed [25]. This study serves as an applicative case study focused on Jakov’s Veterinary Centre in Vilnius, showcasing the practical implementation of machine learning in predicting canine bone fractures. It also highlights the potential for developing a comprehensive IT tool that can be integrated into clinical practice for regular veterinarians. Our objective is to shift the focus from treating fractures as mere outcomes to proactively assessing and mitigating fracture risks, thereby enhancing preventive care to a level similar to that in human medicine.

2. Materials and Methods

In order to begin the process of developing an algorithm for machine learning, we collected data only on canines that had previously experienced a fracture. The veterinary center located in Vilnius city was the source of all of the obtained data records. Starting from 1 January 2000 and continuing until 1 March 2024, all of the information on dog fractures was gathered and documented using Microsoft Excel 365. A comprehensive dataset of 2261 clinical cases was collected, encompassing canines that had experienced a fracture. The following data, as documented on the day of the fracture, were gathered: the breed, age, and weight of the animal. Additionally, the gender, the type of fracture, and the specific bone that was broken were noted. Furthermore, the cause of the fracture was collected during the history-taking process. The software programs Matlab 2023 and Microsoft Excel 365 were used to categorize and analyze the data.

2.1. Processing of Collected Data

The acquired data were subjected to independent analysis and grouping. Based on the available data, the dogs were initially categorized based on their age groups [26]. Specifically, the dogs were classified into the following groups: puppies less than 1 year in age, young adults 1 to 1.5 years in age, mature adults 2 to 6.5 years in age, seniors 7 to 9.5 years in age, late seniors 10 to 12 years in age, and geriatric dogs aged 12 years and above.

We conducted individual evaluations of the dog’s breed and gender, as well as the particular type of bone fracture experienced.

The fracture causes are classified as unknown, domestic trauma—including all instances involving falls from heights, such as beds, furniture, or a person’s arms when being held, and injuries experienced by dogs during play—and separate groups for vehicle accidents and incidents or harm caused by an assault by another canine.

To ensure that the model accurately reflects the diverse nature of the canine population, the data were stratified into subgroups based on critical factors, such as age, breed size, and fracture characteristics. This stratification is intended to enhance the model’s robustness by allowing it to capture and leverage specific patterns and trends unique to each subgroup, thereby improving the overall predictive accuracy.

The animals were categorized according to their breed-specific weight groups, which were as follows: extra-small dogs; small; medium–small; medium–large; large; and giant.

2.2. Random Forest Classifier Method

The random forest classifier is an example of an ensemble learning technique used for classification and regression tasks. It improves the performance and accuracy of machine learning models by aggregating the outcomes of numerous decision trees that are built throughout the training process. The approach described in Breiman’s research [27] study produces the class that represents the mode of the classes in the separate trees, which are used specifically for classification tasks. The random forest classifier was used in the development of a tool for veterinarians to predict canine bone fractures.

2.2.1. Mathematical Formulation

Given a set of decision trees, $\{D_1,D_2,\dots ,D_n\}$, the random forest classifier predicts an input vector, X, as follows:

$$\widehat{y}=\mathrm{mode}\{D_1\left(X\right),D_2\left(X\right),\dots ,D_n\left(X\right)\}$$

(1)

This prediction mechanism leverages the diversity of the decision trees, which are trained on different subsets of data and features, leading to a reduction in the model’s variance without a significant increase in bias.

2.2.2. Entropy and Information Gain

Entropy, a measure of the purity of the subset, is given by

$$H\left(S\right)=-\sum _{i=1}^n{p}_i{log}_2\left(p_i\right)$$

(2)

where $H\left(S\right)$ is the entropy of set S, n is the number of classes, and $p_i$ is the proportion of class i in S. This measure is crucial in the construction of the decision trees within the random forest, guiding the selection of splits that most effectively purify the child nodes.

Information gain, the metric used to select the splitting criterion, is defined as

$$IG(A,S)=H\left(S\right)-\sum _{t\in T}\frac{|S_t|}{\left|S\right|}H\left(S_t\right)$$

(3)

where $IG(A,S)$ is the information gain obtained by splitting set S for attribute A, T is the set of subsets created by splitting S by A, and $|S_t|$ is the size of subset t. Information gain is used to measure the effectiveness of an attribute in classifying the training data to maximize the gain at each split in the trees [28].

2.3. Computational Methodology for Predicting Canine Bone Fracture Risk

A fracture risk assessment tool was developed using Python 3.12.2 [29], and the code was produced. In the Python environment, the libraries used were pandas 2.2.1 [30], scikit-learn 1.4.0 [31], and openpyxl 3.1.2 [32].

This subsection describes the computational approach used for the prediction of bone fracture risk in canines. The methodology involves the utilization of a dataset that encompasses a range of physiological and categorical characteristics. This study incorporates a sophisticated machine learning methodology (i.e., random forest classifier) into the veterinary context to improve the accuracy of the prediction of health outcomes by considering a diverse range of factors, such as age, gender, breed, and weight. The overall protocol of the computational technique described in this study is shown in Figure 1.

The process consists of many crucial phases, each of which enhances the model’s capacity to provide accurate and practical predictions. The method encompasses many key stages, including data loading and pre-processing, model training and assessment, probability calculations for particular attributes, and the interpretation of the prediction findings. Every element is essential to the effectiveness of the model and will be explained in detail.

2.3.1. Data Loading and Pre-Processing

The first stage of the process is retrieving veterinary data from an Excel file named data.xlsx. The dataset contains essential data related to canines, including variables such as age, gender, breed, and weight. It is worth mentioning that ages categorized as “less than 1 year” are modified to 0.5 years to maintain uniformity in their numerical examination.

The pre-processing and data-splitting function is used for this purpose. After loading the data, the dataset undergoes pre-processing to enhance the feasibility of machine learning analysis. StandardScaler is used to scale numerical variables, such as the dog’s age in years and its weight, whereas OneHotEncoder is used to encode categorical features such as gender and breed. Pre-processing is essential for successfully managing categorical data and normalizing numerical variables, which improves the model’s capacity to learn. Subsequently, the data are divided into training and testing sets, with 20% of the dataset designated for model assessment.

2.3.2. Model Training

The train_model function is used to create a machine learning pipeline that combines pre-processing stages with RandomForestClassifier. The use of a pipeline facilitates the optimization of the whole process, including data preparation and model training, hence promoting a unified workflow. RandomForestClassifier was selected due to its effectiveness in classification tasks, making it very suitable for the prediction of bone fracture probabilities.

2.3.3. Probability Calculations and Predictive Analysis

This part focuses on the calculation of trait probabilities and conditional probabilities using the calculate_trait_probabilities and calculate_conditional_probabilities functions. The purpose of these functions is to calculate the probability distributions of different qualities included in the dataset, as well as the conditional likelihood of having a fracture for each attribute. This probabilistic research elucidates the association between certain characteristics and the likelihood of fractures, providing a more profound understanding of the elements that impact bone health in canines.

2.3.4. Prediction and Results Presentation

make_predictions_and_save: In the last stage, the trained model is used to forecast the likelihood of fractures in new patients, taking into account their distinct characteristics. Additionally, the function aggregates characteristic and conditional probabilities to provide a full evaluation of the risk of fractures. Fracture risk.xlsx, which contains the predictions and associated statistical analysis, is created and assessed. This file provides veterinarians with a comprehensive and organized summary of the predicted results.

The validate_input function function is used to validate input characteristics against established criteria before producing predictions. This process ensures that the data being examined are both legitimate and relevant.

3. Results

The next section provides a comprehensive statistical analysis of the gathered clinical cases, with a specific emphasis on the various characteristics of canine fractures. The data obtained provide a detailed understanding of the distribution patterns according to gender, breed, age, breed size, and the exact characteristics of fractures. Furthermore, we provide a novel methodology for evaluating the canine risk of bone fractures, which offers valuable insights into prospective factors that might enhance the process of clinical decision-making in the field of veterinary medicine.

3.1. Statistical Analysis

The prevalence of fractures among canines was assessed using 2261 cases.

The information displayed in Figure 2 reveals that the gender distribution of the cohort was skewed in favor of males, comprising 54.58% (n = 1234) of the total, while females constituted 45.42% (n = 1027). The observed gender disparity in fracture incidence among the canines under study is indicative of the sample population as a whole and suggests that males account for the majority.

Within the study cohort, it is evident that the incidence of fractures in extra-small breeds is significantly higher than in other breeds, accounting for 42.24% (n = 955) of the observed cases shown in Figure 3.

This phenomenon may be ascribed to the underlying physical traits seen in extra-small breeds. Individuals with a smaller stature and a comparatively lighter bone composition may exhibit increased susceptibility to fractures, particularly when exposed to physical force or trauma. Moreover, their comparatively large population size and increased probability of seeking care at veterinary clinics may influence the percentage of fractures recorded.

The fracture statistics are influenced by medium–large and medium–small breeds, which represent 19.81% (n = 448) and 12.65% (n = 286), respectively. These breeds may be associated with aspects such as their activity levels and susceptibility to injury during exercise or play. The combined fracture cases of large, small, and giant breeds are somewhat lower, which may be attributed to their stronger bone structures or under-representation within the population sample.

The fracture distribution among mongrel (mixed-breed) and purebred dogs, categorized by gender and shown in Figure 4, offers valuable information on the fracture patterns associated with various breeds in our research.

Mongrels make up 25.36% (n = 313) of male dogs, while purebreds make up a significant majority at 74.64% (n = 921). Likewise, in the population of female dogs, mongrels account for 23.08% (n = 237), while purebreds make up 76.92% (n = 790).

This study included a wide range of distinct canine breeds, 139 in total, suggesting that the occurrence of fractures among purebred dogs covers a wide genetic range. The significant representation of purebred dogs in both the male and female cohorts, accounting for more than 75% of fracture cases, suggests that purebreds may more commonly seek veterinary treatment or exhibit increased susceptibility to fractures as a result of inherent genetic factors or breed-specific anatomical characteristics. The empirical evidence consistently indicates a greater prevalence of fractures in purebred canines of both genders, suggesting that the risk of fractures is not only determined by gender but also potentially impacted by the breed and its corresponding traits.

Figure 5 displays the varying frequencies of fractures across various breeds with a focus on gender differentiation, specifically on those that accounted for more than 1% of all clinical cases.

Among American Staffordshire Terriers, fractures were seen only in females, accounting for 1.85% (n = 19) of the population, whereas no cases were documented in males. Fractures in Chihuahuas were observed only in females, accounting for 1.36% (n = 14), whereas males did not exhibit any such events. On the other hand, the Bernese Mountain breed had a greater distribution in males, with an incidence of 1.94% (n = 24), compared to females at 1.17% (n = 12). A notable frequency of fractures was observed in Yorkshire Terriers, with females at 11.10% (n = 114) and males at 9.24% (n = 114). A more even distribution was observed in Labrador Retrievers, with males exhibiting a slightly greater proportion of 3.81% (n = 47) compared to females at 3.51% (n = 36). Fractures were seen in 1.85% (n = 19) of female Italian Greyhounds, with 2.27% (n = 28) attributed to males. The analysis of fractures in French Bulldogs revealed a prevalence rate of 1.46% (n = 15) in females and 1.62% (n = 20) in males. The occurrence of fractures in Russian Toy Terriers was 2.14% (n = 22) in females and 3.32% (n = 41) in males. In contrast, fractures in Siberian Huskies were only documented in females, with a value of 1.07% (n = 11). Fractures were observed only in the males of certain breeds, such as the Miniature Pinscher (1.05%, n = 13), Pekingese (1.22%, n = 15), Smooth-haired Dachshund (1.22%, n = 15), and Cane Corso (1.05%, n = 13). Fracturing was seen in 7.89% (n = 81) of female Toy Terriers and 5.67% (n = 70) of male Toy Terriers. In contrast, German Shepherds had fracture rates of 7.79% (n = 80) for females and 8.83% (n = 109) for males.

The data are shown in Figure 6, facilitating a visual comparison of fracture incidence across various age groups and genders.

Within the smallest age group, namely, puppies under 1 year old, there is a comparable occurrence of fractures between males and females. Females have a prevalence of 48.49% (n = 498), and males have a distribution of 47.73% (n = 589). The observed high proportion is likely attributable to the heightened levels of activity and susceptibility to injuries in puppies, which may be attributed to their exploratory tendencies and the ongoing development of their musculoskeletal systems.

The frequency rates for young adult canines—namely, those aged 1 to 1.5 years—are 12.07% (n = 124) for females and 14.10% (n = 174) for males. The group of mature adults aged 2 to 6.5 years has a similar distribution across genders, with females accounting for 26.58% (n = 273) and males representing 26.18% (n = 323). This pattern may be attributed to the prolonged engagement in physical activity and the associated risk of injuries during this particular phase of life.

A reduction in the occurrence of fractures is seen in older canines—namely, those aged 7 to 9.5 years—with an occurrence of 6.72% (n = 69) for females and 6.81% (n = 84) for males. This lower fracture incidence may be related to a decrease in physical activity levels or more careful movements as dogs progress in age. The aforementioned pattern persists across individuals in the late senior and geriatric age cohorts, wherein the smallest proportions are recorded in canines aged 12 years and above, specifically 2.43% (n = 25) for females and 2.35% (n = 29) for males. This phenomenon may suggest a decrease in mobility that is mostly seen in individuals in this age group, which may contribute to a decreased likelihood of circumstances that could potentially result in fractures.

Figure 7 illustrates the many forms of fractures seen in canines of both female and male genders.

Both genders exhibit high proportions of simple and comminuted fractures. Simple fractures account for 45.67% (n = 469) of female cases and 46.52% (n = 574) of male cases. On the other hand, comminuted fractures constitute 41.48% (n = 426) of female cases and 42.71% (n = 527) of male cases.

While less frequent, multiple fractures still account for a significant proportion, with 10.13% (n = 104) in females and 8.10% (n = 100) in males. Less frequent forms of fractures include oblique, transverse, avulsion, open, and spiral fractures, all of which have less than a 2% occurrence, suggesting their relative infrequency among the population under investigation.

The observed distributions indicating the specific types of injuries experienced by different genders might be driven by variations in bone density and behavior between male and female dogs.

Figure 8 presents the distinct fracture types in both female and male dogs.

The occurrence of fractures in the ulna and pelvic bones is noteworthy, especially among females, with 15.19% (n = 156) for ulna fractures, which include fractures of the entire ulna, processus coronoideus, and processus anconeus, and 12.46% (n = 128) for pelvic bone fractures. These findings indicate that these regions are often affected by injuries.

Tibia and femur fractures are equally notable, suggesting that these long bones are often impacted by trauma or external loads. Males have a greater incidence of tibia fractures (14.02%; n = 173) when compared to females (11.68%; n = 120). Similarly, males have a slightly higher frequency of femur fractures (9.24%; n = 114) than females (9.06%; n = 93).

The lower frequency of fractures in bones such as the metacarpal and metatarsal bones and the even lower frequency of fractures in the jaw, hock, and radius indicates a reduced vulnerability to or incidence of external loads on these regions.

The classification of fracture etiology, as shown in Figure 9, is significant in the context of identifying preventative interventions and comprehending the epidemiology of fractures in canines. It is noteworthy that a significant proportion of fracture causes, 58.13% in females and 56.89% in males, was recorded as unknown. This reflects a common challenge in busy clinical settings, where there may be limited time to obtain and record detailed information about the cause of fractures, despite recommendations to document this data.

The classification of fracture etiology, as shown in Figure 9, is significant in the context of identifying preventative interventions and comprehending the epidemiology of fractures in canines.

The analysis of fracture causes in canine patients indicates a notable proportion of cases of unidentified origin, including 58.13% (n = 597) in females and 56.89% (n = 702) in males. This implies that, in many cases, the triggering incident resulting in the fracture is either not seen or not communicated during veterinarian consultation.

Trauma is the second most frequent cause of fractures, accounting for 23.08% (n = 237) in females and 19.77% (n = 244) in males. This indicates the influence of several damage mechanisms that are not particular to trauma. Vehicle accidents have been identified as a particular cause, representing 15.00% (n = 154) of fracture cases in females and 18.56% (n = 229) in males. This finding underscores the significant risk associated with injuries resulting from traffic incidents. The least frequent documented cause of fractures is an assault by another dog, which accounts for 3.80% (n = 39) of cases in females and 4.78% (n = 59) in males. This suggests that aggressiveness or interactions with other dogs might increase the risk of fractures.

3.2. Algorithmic Assessment of Canine Fracture Risk Based on Clinical Data

Within our tool, the user inputs data on the canine’s age, weight, gender, and breed. The algorithm outputs an extra Excel file with two tabs. The first tab, labeled “Animals Details”, presents the patient information entered, while the second tab, labeled “Fracture risk”, provides the data shown in

Table 1. Probability of fractures in various bones.

Bone	Risk (%)
Carpal bones	5
Elbow bones	22
Femur	9
Fibula	0
Hock	0
Humerus	4
Hyoid bone	1
Jaw	0
Metacarpal bones	8
Metatarsal bones	10
Patella	1
Pelvic bones	13
Radius	0
Ribs	2
Scapula	1
Skull	0
Spine	11
Tail	0
Tibia	12
Ulna	1

.

The table shows the resulting fracture probabilities for a randomly entered patient—a male Yorkshire Terrier aged 3 years and weighing 3.4 kg—with the greatest likelihood of sustaining a fracture in the elbow bones at a rate of 22%, followed by the pelvis bones at 13%, the tibia at 12%, and the spine at 11%. The findings presented in this study suggest the presence of a biomechanical vulnerability that is unique to the breed and size of the individuals. For instance, the heightened vulnerability of elbow bones may indicate the breed’s inclination for domestic accidents or traumas. On the other hand, it is noteworthy that bones such as the tibia, fibula, jaw, solitary radius, skull, and tail have a risk of 0%, suggesting that the individuals possess inherent strength and resilience in these specific regions. The careful evaluation of such material is important.

The probabilities obtained were used for training and assessment using the fracture data file that was presented with the programming code. The greater the number of records supplied in a data file for a single patient, the greater the likelihood of encountering an identical bone fracture. In the absence of a match, the algorithm assesses and computes similarities by comparing the distributions of the remaining data.

4. Discussion

In this section, a thorough assessment of the model’s validity is conducted through a comparative examination of the algorithm’s predictions and the actual observed clinical cases. We conducted a comparative analysis of the obtained statistical data from studies conducted by other researchers who studied canine bone fractures. The subdivision of the data into specific age and size categories, as well as by fracture type and location, was a strategic decision to enrich the model’s analytical depth. This approach allows the algorithm to discern nuanced patterns that may be lost in a more generalized analysis, thus reinforcing the model’s robustness and its ability to provide detailed risk assessments tailored to individual canines. Moreover, our attention was directed to the comparison of our model and regression analysis to help provide useful recommendations to veterinarians for use in their everyday clinical work.

4.1. Limitations of the Study

This study presents significant initial findings on the application of machine learning for predicting canine bone fractures. However, there are notable limitations that must be acknowledged. The primary challenge is the sample size of 2261 cases, which, while substantial, is not sufficiently large to ensure broad external generalization across all canine populations. This limited sample size may affect the robustness and applicability of the prediction model to diverse and unobserved canine subpopulations.

Additionally, the stratification into multiple subgroups based on size, age, fracture locations, and types, though essential for capturing heterogeneity, further reduces the effective sample size within each subgroup. Consequently, the results, while promising, should be interpreted with caution and viewed as preliminary. Future studies should aim to include larger and more diverse datasets to validate and potentially refine the prediction model.

The stratification into subgroups further emphasizes the need for larger datasets to enhance the model’s predictive power across diverse canine populations. This limitation should be explicitly discussed in future research conclusions.

4.2. Similarities to and Differences from Other Canine Fracture Studies

Our findings were highly similar to those of other authors who have researched bone fractures in canines, noting differences that indicate that males are more prone to fractures, constituting 55% of cases; for comparison, the study by Abo-Soliman et al. [23] reported 60% for males in canine fractures.

A pattern was also observed among breeds, with our cases predominantly involving mixed breeds, which consisted of 24.33% (n = 550), and breeds such as the Yorkshire Terrier 10.08% (n = 228), Toy Terrier 6.68% (n = 151), and German Shepherd 8.36% (n = 189), having a higher predisposition to fractures. Based on the analysis conducted by Kong and Shin, mixed breeds consisted of almost 53%—including Labrador Retrievers (23.52%), German Shepherds (11.76%), Doberman Pinschers (2.94%), and small breeds such as Pomeranians (5.88%) and Spitzes (2.94%)—in their study [33]. To provide a comprehensive understanding of our findings, we compared the fracture prevalence in our study with the general canine population demographics. Our study sample indicated that males, which constituted 54.58%, had a higher prevalence of fractures. This was analyzed in the context of the general population, where males are also more numerous, thus clarifying whether the higher incidence is due to their majority or an inherent predisposition to fractures. Similarly, a comparative analysis was performed for other variables, such as breed size and age, ensuring a thorough understanding of the fracture risks in relation to the source population.

The ages of the animals also matched: in our study, the largest number of dogs were less than 1 year in age, which accounted for 48.07% (n = 2261), and the group of young animals from 1 to 1.5 years of age accounted for 13.18% (n = 298). Lee et al. [20] discovered that the average age was 20 months, with the majority (80%) ranging between 2 and 24 months. The analysis conducted by Smith et al. [21] determined that the median age of dogs at the time of fracture was 4 months. Almost the same results were obtained by a group of epidemiologists who analyzed fractures [24]. Breed size was also a significant factor in our results. From all the analyzed cases, the most common fractures were seen in the elbow 23.13% (n = 523), ulna 15.43% (n = 349), pelvis 10.52% (n = 238), and tibia 12.95% (n = 293) bones between both genders. The ulna fractures represent the combined total of fractures involving the entire ulna, processus coronoideus, and processus anconeus. This approach ensures a comprehensive assessment of all significant fracture sites within the ulna, providing a detailed understanding of fracture patterns in this bone. According to Fazili, the humerus, radius, ulna, femur, tibia, and fibula bones are most frequently affected [11]. Almost the same distribution of fractures was observed in an epidemiology-based study [24]. Additionally, the research conducted by Aithal and Singh et al. [5] observed and discovered that, in most cases, femur fractures constituted more than 40%. Furthermore, the study also identified common pelvic/sacrum fractures.

The causes of the fractures were also compared. For the largest proportion of our analyzed cases, the cause of the fracture was unknown, as most owners could not explain what happened to their pet; these cases constituted 57.45% (n = 1299). Domestic traumas accounted for 21.27% (n = 481), car accidents for 16.94% (n = 383), and assaults by other dogs for 4.33% (n = 98). The most common cause of fractures in the study conducted by Keosengthong et al. [34] was vehicle accidents, which were the cause for almost 80% of canines; Kitshoff et al. [35] demonstrated that the most frequent cause of fractures was fighting with other canines, accounting for 62%. We acknowledge that in busy veterinary clinics, there may be insufficient time to consistently obtain and record the exact causes of fractures. This limitation is reflected in our data, where a substantial percentage of fracture causes were documented as unknown. This underscores the need for improved data-recording practices, as accurate etiological information is crucial for enhancing our understanding and prevention of fractures.

4.3. Other Machine Learning Tools Used for Fracture Risk Assessment and Their Limitations

Machine learning technologies are extensively used to evaluate the risk of bone fractures and use advanced algorithms to improve the precision and dependability of predictions. Much research has shown the potential of random forest classifiers. However, it is worth noting that a wide range of other machine learning approaches are also being used in the field of bone health and are providing useful insights and capabilities. Unfortunately, most of these tools can only practically be used for human medicine, not for animals. The aforementioned tools encompass the following.

The utilization of Cox regression models and an artificial neural network (ANN)-DeepSurv model provides a diverse set of approaches to evaluating prospective fracture chances in individuals diagnosed with osteopenia and osteoporosis. This showcases the adaptability and comprehensiveness of machine learning applications within this particular domain [3].

An additional tool that employs machine learning models to diagnose and categorize osteoporosis by detecting fractures from images and predicting fracture risk via the use of various data sources provides a comprehensive approach to tackling osteoporosis and related bone health issues [36].

Support vector machines (SVMs), gradient-boosting machines (GBMs), neural networks (NNETs), and regularized discriminant analysis (RDA) are being used to predict new fractures following treatment for osteoporotic vertebral compression fractures. The analysis underscores the wide range of machine learning algorithms that can be applied [37].

The use of artificial intelligence in medical diagnostics has been applied to the prediction of fragility fractures using MRI data. This integration of artificial intelligence enhances diagnostic accuracy and risk assessment, highlighting the unique application of AI in medical diagnostics [38].

Some researchers have explored the possibility of automation and machine learning in facilitating speedy and accurate evaluations of fracture detection, bone mineral density prediction, and fracture risk evaluation through the use of plain radiographs [39].

The combination of advanced imaging methods and deep learning has been seen in the use of deep learning models and high-resolution peripheral quantitative computed tomography (HR-pQCT) for the full investigation of bone health, with a specific emphasis on analyzing bone micro-architecture and fracture risk [40].

Random forest (RF) classifiers were chosen due to their notable attributes, including high resilience, fast handling of intricate datasets, and remarkable levels of accuracy, specificity, and sensitivity. RF specializes in handling complex data and can be used for the evaluation of canine fracture risk.

It is important to note that our study, being retrospective, did not include data on the breed conformation, such as skeletal slenderness. This is a notable limitation, as breeds with different skeletal builds may have varying susceptibilities to fractures. For instance, slender breeds like Italian Greyhounds might differ significantly from more robust breeds like German Shepherds, even though they are within the same size category. Additionally, we recognize that lifespan variations among breeds could influence fracture risk profiles, with some breeds having shorter lifespans while others may live significantly longer. Future studies should aim to incorporate these factors to provide a more nuanced understanding of fracture risks across different canine breeds.

While many of our findings align with established knowledge in veterinary epidemiology and preventive medicine, this study underscores the practical application of machine learning in a clinical setting. By leveraging this approach, we demonstrate its potential as a foundational tool that could evolve into a more comprehensive IT solution. Such a tool could be embedded in clinical software to alert veterinarians to potential fracture risks based on patient data, thereby enhancing preventive care and clinical decision-making.

4.4. Recommendations for Veterinarians

• Early development: Based on our findings, which indicated a higher fracture incidence in younger canines, veterinarians should advise owners of young dogs (aged 1.5 years and below) to regulate their physical activity to prevent bone and joint injuries. Emphasis should also be placed on providing a nutrient-rich diet to support optimal skeletal development.
• Small-breed considerations: Given the significant prevalence of fractures in extra-small breeds observed in our study, it is advisable for owners of these breeds to implement safety precautions at home, such as using specific padding to mitigate risks from falls or jumps.
• Domestic safety: Our analysis showed a notable number of fractures due to domestic accidents. Therefore, it is important for owners to minimize potential domestic hazards, ensuring a secure environment for their dogs.
• Sporting dogs: There is a correlation between high activity levels and fracture risk. Veterinarians should recommend appropriate warm-up and cool-down routines tailored to each breed to prevent muscle and joint injuries.
• Elderly dog care: With a higher fracture incidence noted in older dogs, especially due to reduced sensory perception, modifying care routines to include safer walking paths and attentive supervision is essential.
• Understanding canine body language: Considering the data showing fractures from dog-to-dog interactions, educating owners to understand canine body language can help prevent such incidents.

5. Conclusions

The statistical analysis conducted in this research provides a comprehensive understanding of fracture patterns across various groups within the canine population. By comparing our findings with the general canine population, we observed that, while extra-small and purebred dogs showed a higher proportion of fractures, this may be linked to their prevalence and inherent physical traits. For instance, although males exhibited a higher fracture incidence in our study, this aligns with their majority in the population, suggesting behavioral rather than biological predispositions. Additionally, the breeds with the highest incidence of fractures, such as Yorkshire Terriers, German Shepherds, and Toy Terriers, were contextualized against breed-specific risks. Puppies under the age of 1 year showed the highest occurrence of fractures, which is consistent with their developmental stage and activity levels, indicating both an increased risk and a higher likelihood of receiving veterinary care.

The distribution of fracture types exhibited a prevalence of simple and comminuted fractures, perhaps indicating the shared nature of the damage processes that give rise to these specific forms. The most frequently injured anatomical areas were the elbow bones, ulna, and pelvic bones, perhaps due to the biomechanical pressures applied to these bones.

When investigating the causes of fractures, a considerable proportion of cases remained unidentified, although trauma and vehicular accidents emerged as the prevailing causes, indicating possible avenues for the implementation of preventive interventions to mitigate fracture susceptibility.

The proposed fracture risk assessment algorithm, which was built and trained on a varied dataset, has a strong capacity to forecast fracture risk. This algorithm effectively included several parameters, including age, gender, breed, and weight. The aforementioned prediction model serves as evidence of the capacity of machine learning approaches to improve diagnostic precision and preventive measures in the field of veterinary medicine.

The incorporation of this algorithm into clinical practice has the potential to facilitate the timely detection of canines with an elevated susceptibility to fractures, hence enabling the implementation of customized preventative measures and enhanced owner education on fracture risks and injury prevention.

This study not only demonstrates the feasibility of applying machine learning to fracture risk assessment in a specific clinical setting but also lays the groundwork for the future integration of such predictive tools into broader veterinary clinical software systems, enhancing preventative care and clinical decision-making. However, due to the limitations in sample size and the preliminary nature of these findings, further research is necessary to confirm these results with larger datasets and diverse canine populations. Such future work will be essential for enhancing the external generalization of the model and developing more comprehensive predictive tools for veterinary medicine.

Future research should focus on incorporating breed-specific skeletal conformation and lifespan data to enhance the predictive accuracy of fracture risk models. This would allow for a more detailed assessment of how physical characteristics and lifespan differences influence fracture susceptibility in various canine breeds.