INTRODUCTION AND BACKGROUND
BACKGROUND
Osteoporosis is a skeletal disorder that increases the risk of fractures due to compromised bone strength, which encompasses both bone quality and bone density Bone quality is evaluated through factors such as micro-architecture, turnover, and mineralization, while bone density, or bone mineral density (BMD), is measured as the mineral content in a specific area of bone Currently, BMD is assessed using bone densitometry, particularly through dual-energy X-ray absorptiometry (DXA), which is considered the gold standard for measuring bone density This quick, painless, and non-invasive technique accurately quantifies the mineral absorption in bones, primarily at sites like the lumbar spine, femoral neck, or hip, making it essential for diagnosing osteoporosis.
Osteoporosis is defined by the World Health Organization based on Bone Mineral Density (BMD), specifically using the BMD T-score This score is calculated by taking the difference between an individual's BMD and the peak BMD of a reference population of young, healthy adults, then dividing by the standard deviation For femoral neck measurements, the Third National Health and Nutrition Examination Survey (NHANES III) database for White women aged 20–29 years is recommended as the reference range.
In 2008, DXA instruments were designed to calculate bone mineral density (BMD) values using established references According to the World Health Organization (WHO), osteoporosis is characterized by a BMD at the femoral neck that is 2.5 standard deviations or more below the average for young adult females, indicated by a T-score of ≤ -2.5 A T-score of femoral neck BMD ≤ -2.5 signifies severe osteoporosis when accompanied by additional clinical factors.
A study conducted in Australia revealed that around 6% of men and 23% of women over the age of 50 are affected by osteoporosis, with these figures rising to 13% for men and 43% for women over 70.
Osteoporosis ultimately leads to osteoporotic or fragility fractures, which are defined as any loss of bone continuity at various sites These fractures predominantly occur in individuals aged 50 and older, often linked to low bone mineral density (BMD) and typically resulting from low-trauma incidents, such as falls from standing height or less, without underlying diseases like cancer However, this definition may overlook the true prevalence of osteoporotic fractures, as high-trauma incidents can also be classified as osteoporotic The most frequently affected areas include the hip, spine, and distal forearm.
Advancing age is a significant risk factor for fractures, particularly in women over 60, who face a fracture risk that is 10.2 times greater than pre-menopausal women (Jiang et al., 2013) In Caucasian women, the risk of fracture doubles with every five-year increase in age (Nguyen et al., 2007; Taylor et al., 2007).
Advancing age significantly increases the risk of hip fractures in Caucasian men, comparable to the risk observed in women, with studies indicating a 1.5-fold higher risk (Cummings et al., 1995; Kanis et al., 2005) Additionally, for every five years of aging, the risk of humeral fractures rises approximately 2.8-fold, while forearm fractures increase by about 1.6-fold in men (Nguyen et al., 2001) Although the exact mechanisms linking aging to fracture risk remain unclear, they may be related to the cumulative decline in physical health that accompanies aging (Kelsey & Samelson, 2009).
Low Bone Mineral Density (BMD) significantly increases fracture risk, with each standard deviation decrease in BMD correlating to a 1.5 to 3.0-fold rise in fracture likelihood BMD is particularly predictive of hip fractures, which have a risk gradient of 2.6 Interestingly, while individuals with osteoporosis face a higher relative fracture risk, those with osteopenia experience the highest absolute number of fractures—approximately five times more than those with osteoporosis This disparity has prompted researchers to explore additional risk factors for fractures beyond BMD measurements.
Gender, history of falls, and prior fractures are significant risk factors for fractures Women face a lifetime fracture risk of 40-50%, compared to 13-22% for men Falls are strongly linked to fractures due to weak muscle strength and increased body sway Nursing home residents experience two to six falls annually, with up to 5% resulting in fractures Notably, having a prior fracture doubles the risk of an osteoporotic fracture, while an asymptomatic vertebral fracture can increase the risk of subsequent vertebral fractures by 7 to 10 times.
In addition to previously mentioned factors, several risk elements contribute to fracture susceptibility, such as a family history of fractures, lifestyle choices, nutritional habits, and genetic predispositions Key risk factors include excessive alcohol consumption, smoking, long-term glucocorticoid steroid use, low body weight, inadequate calcium intake, early menopause, and specific medications that disrupt bone metabolism (McCloskey et al., 2012).
Numerous studies have uncovered various genetic variants linked to bone density and the risk of osteoporotic fractures In 2008, researchers identified 77 single nucleotide polymorphisms (SNPs) associated with bone mineral density (BMD), with 16 of these SNPs specifically correlating to fracture risk Over the past decade, additional genetic variants relevant to fracture risk have been discovered However, it is important to note that the individual effect sizes of these identified genetic variants on BMD and fracture risk are generally modest.
Osteoporosis research prioritizes identifying high-risk individuals for timely intervention, as the lifetime risk of hip and vertebral fractures in women over 50 is approximately 32%, comparable to or exceeding the risk of invasive breast cancer Alarmingly, one in five women who experience a hip fracture may not survive the first year, with increased mortality persisting for up to ten years In men, the lifetime risks for these fractures are about 11%, similar to the likelihood of being diagnosed with prostate cancer Without intervention, the number of osteoporotic fractures is projected to reach 3 million by 2021 due to population aging These fractures impose a significant financial burden, encompassing direct costs such as hospital care and rehabilitation, as well as indirect costs like lost workdays In Australia alone, direct costs related to osteoporotic fractures were estimated at $1.9 million annually, with predictions of 183,105 cases and a total cost of $33.6 billion from 2012 to 2022 Therefore, identifying high-risk individuals is crucial for effective disease management and early intervention to mitigate the impact of fractures and their associated consequences.
Several predictive models have been created to assess fracture risk, with the Garvan Fracture Risk Calculator and the FRAX tool being the most prominent The Garvan Fracture Risk Calculator, established through a 20-year study of over 3,500 elderly individuals in Dubbo, evaluates fracture risk over 5 and 10 years by considering factors such as age, sex, femoral neck bone mineral density (BMD), body weight, previous fractures, and falls in the last year Meanwhile, FRAX®, introduced in 2007, estimates the 10-year risk of major osteoporotic fractures—including clinical spine, hip, forearm, and proximal humerus fractures—using 13 different risk factors.
In 2009, the QfractureScores model was introduced to estimate a 10-year fracture risk based on 20 non-BMD risk factors, as highlighted in the English cohort study by Hippisley-Cox and Coupland While other fracture risk assessment models, such as Fracture INDEX (Black et al., 2001; Robbins et al., 2007), FRISK (Henry et al., 2006), and a simple four-item risk score (Pluijm et al., 2009), have been developed, they are not commonly utilized in clinical practice.
MOTIVATION OF THE THESIS
Current fracture risk models, such as the Garvan fracture risk calculator and the FRAX tool, exhibit varying levels of accuracy, with AUC values ranging from 0.61 to 0.85 (Bolland et al., 2011; Langsetmo et al., 2011; Sambrook et al., 2011; Sandhu et al., 2010) A study in a Canadian population demonstrated a notable concordance between the Garvan model's predictions and actual fracture occurrences, achieving scores of 0.69 for women and 0.70 for men (Langsetmo et al., 2011) This concordance was even higher for hip fractures, with scores of 0.80 for women and 0.85 for men Additionally, a performance comparison of QfractureScores and FRAX was performed using data from English postmenopausal women.
Both models demonstrated a competitive discrimination with an AUC of 0.67 However, for hip fracture discrimination, QfractureScores achieved a higher accuracy with an AUC of 0.71, compared to FRAX's AUC of 0.64 This indicates that there is still potential for improvement in these models.
The low accuracy of current fracture predictive models may stem from their failure to consider the interactions among various risk factors Utilizing an Artificial Neural Network (ANN) approach could effectively address these interactions, enhancing predictive accuracy.
Artificial Neural Networks (ANN) effectively mimic human brain functions to model complex real-world relationships, including the interactions among various variables Recent research has demonstrated the application of ANN in detecting vertebral fractures in postmenopausal women with osteoporosis and predicting mortality after hip fractures, showing greater accuracy compared to traditional statistical methods like logistic regression Despite these advancements, the use of ANN in osteoporosis research remains limited.
Incorporating genetic factors alongside clinical factors into fracture risk predictive models can enhance their accuracy and uniqueness Although individual BMD-associated genetic variants have modest effect sizes, a simulation study by Tran et al (2011) demonstrated that a genetic profile comprising up to 50 variants could improve fracture prediction accuracy by 10% in terms of AUC This finding aligns with other research on genetics and fracture risk (Estrada et al., 2012; Huang et al., 2003; Makovey et al., 2007) Despite this potential, no existing tools currently integrate these genetic variants as risk factors in their models.
Developing multivariable risk prediction algorithms that integrate genetic and non-genetic factors can enhance the identification of individuals at high risk for fractures By utilizing advanced modeling techniques, we can improve genetic profiling, leading to better fracture risk assessment and personalized prevention strategies In the context of translational genetics in osteoporosis, addressing five key questions is essential to bridge existing knowledge gaps.
(1) How can we derive a genetic profiling of the genetic data to predict individual bone mineral density and risk of fracture;
(2) Does genetic profiling improve the prediction accuracy of fracture beyond that obtained with conventional clinical risk factors (Nguyen & Eisman, 2013);
(3) Does genetic profiling contribute to the relative rate of BMD change;
(4) Does the incorporation of genetic profiling improve the clinical utility of existing predictive model for fracture risk, and
(5) How to develop a predictive model for fracture risk, in which the interaction between risk factors are taken into account for a better performance
The studies reported in this thesis have been based on the following specific hypotheses:
Hypothesis 1: Fracture risk and BMD are partly determined by an osteogenomic profile derived from BMD-related genetic variants, each with modest effect
Hypothesis 2: Osteogenomic profile can help to predict fracture risk over and above clinical risk factors
Hypothesis 3: Osteogenomic profile contributes to the variation in the rate of
Hypothesis 4: Osteogenomic profiling can help improve the net benefit of predictive model for fracture risk over and above that of the clinical risk factors
Hypothesis 5: Artificial Neural Network can improve the accuracy of hip fracture prognosis.
AIMS OF THESIS
In order to address the questions, I have conducted a series of studies with the following specific aims:
Aim 1: To construct an osteogenomic profile based on a number of genetic variants associated with low BMD (reported from the genome-wide association studies)
In the Dubbo Osteoporosis Epidemiology study, participants were genotyped for a list of SNPs linked to low bone mineral density (BMD) A weighted genetic risk score (GRS) was then calculated for each individual by multiplying the number of risk alleles by sex-specific regression coefficients derived from genome-wide association studies (GWAS) related to BMD.
Aim 2: (1) To assess the association of the osteogenomic profile and BMD; and
This study aims to evaluate the role of osteogenomic profiling in personalizing fracture risk assessment and to create a clinico-genetic model for predicting fracture risk The first objective utilizes multiple linear regression, with bone mineral density (BMD) as the outcome variable and genetic profiling as the primary risk factor, while adjusting for various covariates For the second objective, a Cox proportional hazards model is employed in a longitudinal study to analyze the relationship between genetic risk scores (GRS) and the likelihood of experiencing different types of fractures, including hip, vertebral, and wrist fractures The effectiveness of predictive models, both incorporating and excluding the osteogenomic profile, is then compared using various metrics, such as the area under the curve (AUC).
10 the receiver-operating characteristic (ROC) curve (AUC), net reclassification index, and calibration
Aim 3: (1) To derive the long-term rate of BMD change for each individual from their serial BMD measurements; and (2) to determine the contribution of the osteogenomic profile to the variability of bone loss The data for this aim is from longitudinal study, in which each individual was followed in a long time with multiple femoral neck BMD measurements ΔBMD, expressed as annual percent change-in-BMD, is determined by linear regression analysis for each individual Multiple linear regression model is used to assess the association between the genetic profiling and BMD changes, adjusted for other covariates Mix effects model is also used as another method to compare the analysis results
Aim 4: To assess the clinical utility of osteogenomic profile in fracture risk perdition based on the net clinical benefit metric In this study, a novel approach called decision curve analysis is used to assess four Cox’s proportional hazard models with and without GRS to compare their clinical net benefit Net clinical benefit is defined as the true positive (benefits) minus the false positive (potential harms), weighed by the cost:benefit ratio Four predictive models are: Base model includes clinical risk factors only (i.e age, prior fracture, history of fall); Base + GRS model adds GRS into the base model; Garvan model includes FNBMD into the Base model; and Full model includes Base + GRS + BMD
Aim 5: To improve the prediction of hip fracture by using Artificial Neural
This study explores a network approach by developing three 3-layer feed-forward artificial neural networks (ANN), each utilizing distinct sets of risk factors The research aims to compare the accuracy of the most effective neural networks against other predictive models that employ the same risk factors.
This study explores 11 distinct methods of development, including logistic regression, Cox’s proportional hazard model, K-nearest neighbor, and support vector machine, to predict hip fractures The model is initially trained on one cohort and subsequently validated for its predictive performance in a separate cohort.
CONTRIBUTIONS
This thesis explores the under-researched area of genetic profiling and machine learning in fracture risk assessment By constructing a genetic profile, it demonstrates that this profile significantly improves the accuracy of existing fracture predictive models, as evidenced by enhanced metrics such as area under the ROC curve, calibration, net reclassification, and net clinical benefit Furthermore, the study reveals that advanced machine learning techniques, particularly neural networks, can further refine fracture risk predictions This research contributes valuable insights to the field of osteoporosis, advancing our understanding of fracture risk assessment.
An osteogenomic profile, derived from 68 genetic variants linked to bone mineral density (BMD) in prior research, is unique to each individual and plays a crucial role in assessing osteoporosis risk.
This thesis reveals a significant link between the osteogenomic profile and bone mineral density (BMD), as well as a notable correlation between the osteogenomic profile and fracture susceptibility By integrating this profile into current predictive models that utilize clinical risk factors, we can enhance the accuracy of fracture prediction These results indicate that genetic profiling of BMD-related genetic variants could improve prognostic performance in assessing fracture risk.
12 improve the accuracy of fracture prediction over and above that of clinical risk factors alone, and help stratify individuals by fracture status
This thesis research highlights the role of the osteogenomic profile in bone mineral density (BMD) changes, revealing that individuals with a "high risk" osteogenomic profile tend to experience greater loss in femoral neck BMD Despite this association, the osteogenomic profile accounts for less than 2% of the total variance in BMD.
A new metric for model assessment, known as net benefit, has been introduced to evaluate the clinical utility of genetic profiling in fracture risk prediction models The findings indicate that for risk thresholds above 0.15, genetic profiling significantly enhances the net clinical benefit for both men and women when combined with existing clinical risk factors However, when bone mineral density (BMD) is included in the model, adding genetic profiling does not improve the net benefit.
A series of three-layer feed-forward artificial neural networks (ANNs) have been developed for predicting hip fracture risk The findings indicate that incorporating both bone mineral density (BMD) and non-BMD risk factors significantly enhances prediction accuracy This suggests that ANNs are a promising alternative to traditional methods, offering valuable support in stratifying individuals for effective clinical management.
STRUCTURE OF THESIS
This thesis comprises seven chapters, along with an appendix and references Each of the main chapters—IV, V, VI, and VII—includes a discussion and conclusion The structure of the thesis, excluding the first chapter, is organized as detailed in the subsequent chapters.
Chapter II examines the literature on osteoporosis and osteoporotic fractures, defining key terms and exploring the consequences of these fractures It identifies potential risk factors for fractures and offers an overview of current predictive models for assessing fracture risk The chapter also highlights gaps in the existing literature regarding the prediction of osteoporotic fractures and proposes a strategy for improving fracture risk assessment.
Chapter III outlines various methodologies employed in this research, detailing the study design, settings, and the collection and management of bone health data, which encompasses anthropometric, clinical, lifestyle, and genetic information Additionally, this chapter presents the statistical methods, evaluation metrics, and machine learning techniques utilized throughout the thesis.
Chapter IV outlines the creation of an osteogenomic profile based on various genetic variants associated with bone mineral density (BMD) It examines the relationship between the osteogenomic profile, BMD, and the risk of fractures Additionally, the chapter introduces an innovative predictive model that integrates the osteogenomic profile into the current clinical framework for enhanced prediction accuracy This model was developed using the Cox proportional hazards approach applied to longitudinal study data, demonstrating superior performance.
The new model demonstrates superior performance compared to previous models lacking an osteogenomic profile, as evidenced by various evaluation metrics such as the area under the receiver-operating characteristic curve (AUC), net reclassification, calibration, and decision curve analysis.
Chapter V analyzes the rate of bone mineral density (BMD) change for each individual based on a series of measurements taken during the follow-up period It also examines the relationship between genetic profiling and the rate of bone loss The study presents the impact of genetic profiling on the variation in bone loss through multiple linear regression analysis.
Chapter VI presents a novel metric termed "net clinical benefit" to evaluate the added value of genetic profiling in enhancing existing predictive models for fracture risk This chapter discusses the comparison of models with and without genetic profiling, utilizing decision curve analysis to illustrate the differences in their effectiveness.
• Chapter VII reports an application of feed-forward Artificial Neural Network classifier for predicting hip fracture in post-menopausal women in
Over a decade, a comparative analysis was conducted among three neural networks utilizing BMB, non-BMD, and a combination of both to identify the most effective model The significance of various risk factors was evaluated, and the performance of this methodology was benchmarked against the results of alternative approaches.
Chapter VIII summarizes the research findings, highlighting the impact of osteogenomic profiles on bone mineral density (BMD), variations in BMD, and fracture susceptibility It also evaluates the effectiveness of the various methods employed throughout the study.
This thesis presents 15 significant findings that contribute to the current understanding of fracture risk Furthermore, it explores future research directions focusing on the genetic factors influencing fracture risk and the application of machine learning techniques for improved fracture prediction.
LITERATURE REVIEW
OSTEOPOROSIS AND OSTEOPOROTIC FRACTURE
Bone is an inelastic organ that supports and protects other organs in the body Bone tissue is composed of organic matrix and inorganic salts (Christoffersen & Landis,
The organic matrix of bone is primarily composed of over 90% type I collagen, with the remaining 5% consisting of proteoglycans The inorganic components mainly include phosphate and calcium, which, along with non-collagenous matrix proteins, create a scaffold that provides bone tissue with its characteristic stiffness and resistance.
Bone health relies on a delicate balance between bone formation and bone resorption, processes that continuously remodel bone tissue An imbalance in these processes can alter bone structure and microarchitecture, resulting in increased fragility Osteoporosis occurs when bone resorption surpasses bone formation, leading to a significant loss of bone strength and integrity, making bones more prone to fractures.
Osteoporosis, a term first introduced in the early 20th century, was officially defined by the World Health Organization (WHO) in the 1990s According to a 1994 WHO study group, osteoporosis is described as a skeletal disorder marked by low bone mass and a compromised micro-architecture of bone tissue, which increases the risk of fractures.
Focusing solely on bone mass overlooks critical aspects of bone health, such as the deterioration of trabecular connectivity and thickness, along with the buildup of micro-damage within the bone structure.
Osteoporosis, defined in 2000 as “a skeletal disorder characterized by compromised bone strength predisposing a person to an increased risk of fracture” (NIH, 2001), involves both bone quality and bone density Bone quality is assessed through micro-architecture, bone turnover, and mineralization, while bone density, or bone mineral density (BMD), measures the mineral content per defined area of bone Current technology primarily evaluates BMD at specific sites such as the lumbar spine, femoral neck, or hip Dual-energy X-ray absorptiometry (DXA) is the gold standard for measuring bone density, utilizing X-ray absorption to provide a quick, painless, and accurate diagnosis of osteoporosis (Watts et al., 2004).
Figure 2-1: Normal bone (left) and osteoporotic bone (right)
Bone consists of a sturdy outer layer known as cortical bone and a lighter, spongy inner structure called trabecular bone, which together provide strength, flexibility, and reduced weight In cases of osteoporosis, the healthy trabecular bone gradually deteriorates, leading to weakened bone structure.
Osteoporosis is defined by the correlation between low bone mineral density (BMD) and an increased risk of fractures BMD typically rises with age until it peaks around 20-30 years, after which it begins to decline, especially in menopausal women At any given age, BMD follows a normal distribution with a consistent standard deviation (SD) The standardized BMD measurement, known as the BMD T-score, is calculated by subtracting the normal BMD (ρBMD) of a young, healthy same-sex adult reference population from an individual's BMD and then dividing by the standard deviation (SD).
The Third National Health and Nutrition Examination Survey (NHANES III) established a reference range for femoral neck measurements in Caucasian women aged 20–29 years, as recommended by Kanis et al (2008) Dual-energy X-ray absorptiometry (DXA) instruments are designed to calculate these values based on pre-defined reference standards.
According to the World Health Organization (WHO), Table 2-1 categorizes bone health based on T-score, utilizing dual-energy X-ray absorptiometry (DXA) for bone mineral density (BMD) measurements, with the total hip as the preferred reference site (Kanis et al., 2000) Specifically, osteoporosis is defined by the WHO as a BMD value of 2.5 or lower at the femoral neck.
A T-score of -2.5 or lower in the femoral neck bone mineral density (BMD) indicates severe osteoporosis, particularly when accompanied by one or more fragility fractures (WHO, 1994) According to an Australian study by Henry et al (2011), approximately 6% of men and 23% of women over the age of 50 are affected by osteoporosis, with these figures rising to 13% for men and 43% for women over 70 years old.
Table 2-1: Classification of bone density based on T-score
T-score < 2.5 SD and a prevalent fracture
Normal Osteopenia Osteoporosis Severe osteoporosis
Another way to standardise BMD is to compare the bone density of an individual with average BMD (aBMD) of healthy one at the same age:
SD aBMD refers to the standard deviation of average bone mineral density (BMD) at a specific age, while aBMD represents the average BMD for that age group A Z-score of -0.5 indicates that an individual's bone density is 0.5 standard deviations below the average for their age cohort.
As mentioned above, loss of bone density is common in men and women aged
Bone loss in individuals aged 50 and older is predominantly observed in trabecular bone, which is more metabolically active due to its larger surface area compared to cortical bone In menopausal women, a decline in bone mineral density is linked to reduced production of reproductive hormones, but factors beyond these hormones may also contribute to bone loss prior to menopause Research suggests that increased sensitivity to endogenous glucocorticoids and changes in body composition could play a role However, the rapid acceleration of bone loss, particularly after the age of 70, remains poorly understood.
Bone loss is influenced by several risk factors, notably advancing age and low body mass index (BMI), which are significant predictors of individual bone loss Other contributing factors include low bone mineral density (BMD), physical inactivity, early menopause, and genetic predispositions Research, such as the Dubbo Osteoporosis Study, highlights the importance of these elements in understanding and predicting bone health outcomes.
A study by the Epidemiology Study (DOES) found that femur bone loss in elderly women is influenced by factors such as advancing age, low body weight, significant weight loss, physical inactivity, and low baseline bone mineral density (BMD), which together account for 13% of the total variance in bone loss (Nguyen et al., 1998) While genetic factors play a significant role in bone loss variation, many associated genes and genetic variants remain unidentified (Wu et al., 2013).
Osteoporosis ultimately leads to osteoporotic fractures, defined as breaks in the continuity of bone at any skeletal site These fractures predominantly affect individuals aged 50 and older, often linked to low bone mineral density (BMD) and resulting from low-trauma incidents, such as falls from a standing height However, the definition may underestimate the true prevalence of osteoporotic fractures, as those resulting from high-trauma events can also be classified as osteoporotic The most commonly affected areas include the hip, spine, and distal forearm.
RISK FACTORS FOR OSTEOPOROSIS FRACTURES
Recent advancements in understanding the risk factors for osteoporotic fractures have revealed that low bone mineral density (BMD) is not the sole predictor of fracture risk Research indicates that approximately 50% of individuals who experience fractures do not meet the criteria for osteoporosis (Nguyen et al., 2007) Consequently, the evaluation of fracture risk has evolved to encompass a broader range of factors beyond just BMD.
23 using BMD cut-off thresholds toward the use of unique profile of risk factors for fracture prediction (Kanis et al., 2005; Nguyen et al., 2007)
Fractures are influenced by various risk factors, which can be categorized into non-modifiable and modifiable groups Non-modifiable risk factors include advancing age, being female, Caucasian ethnicity, a high-risk genetic profile, a family history of osteoporotic fractures, and previous fracture history In contrast, modifiable risk factors encompass elements that can be changed, such as low bone mineral density, a history of falls, long-term steroid use, inadequate nutritional intake of protein, vitamin D, or calcium, lifestyle choices like smoking, alcohol abuse, and physical inactivity, as well as estrogen deficiency.
Identifying individuals at high risk of fractures is crucial for early prevention, and this can be achieved through the assessment of both non-modifiable and modifiable risk factors Numerous studies and tools have been developed to differentiate those at high risk of osteoporotic fractures from the general population (Kanis et al., 2007; Nguyen et al., 2007; Sambrook et al., 2011; Trémollieres et al., 2010).
Table 2-2: Classification of risk factors for fracture Modifiable risk factors Non-modifiable risk factors
- Having a high-risk genetic profile
Advancing age is a critical risk factor for fractures, with the incidence of osteoporotic fractures rising exponentially across all demographics (Cummings et al., 1995; Franic & Verdenik, 2018; Chang et al., 2004) In Caucasian women, each five-year increase in age correlates with a twofold increase in fracture risk and a 1.5-fold higher risk of hip fractures (Cummings et al., 1995; Nguyen et al., 2005; Taylor et al., 2004) Similarly, Caucasian men experience a threefold increase in hip fracture risk with every ten-year age increment, a statistic comparable to that of women (Kanis et al., 2005) Additionally, for men, the risk of humerus fractures increases by 2.8 times and forearm fractures by 1.6 times for every five-year age increase (Nguyen et al., 2001).
The relationship between aging and fracture risk is particularly pronounced for hip, proximal humerus, and vertebral fractures, while no significant difference is observed for distal forearm and foot fractures This discrepancy may stem from variations in daily activities between older and younger individuals Although the exact mechanisms linking aging to fractures remain unclear, it is likely connected to the cumulative decline in physical health that accompanies aging.
Women are more likely to have fracture than men (Alswat, 2017; Cawthon,
The lifetime fracture risk for a 60-year-old woman is significantly higher, ranging from 40-50%, compared to 13-25% for men of the same age This disparity in fracture risk between genders can vary based on geographic regions and ethnic backgrounds.
Women face a twofold higher risk of non-vertebral fractures compared to men when adjusted for age and body size The incidence of vertebral fractures in women significantly increases at age 55, while men see a rise at age 65 Similarly, the risk of hip fractures escalates for women at age 65, compared to age 75 for men This heightened fracture risk in women is attributed to lower volumetric bone density, smaller bone size with age, reduced muscle mass and strength, and accelerated bone loss associated with menopause.
Apart from advancing age, low BMD is the strongest predictor for fracture (Burger et al., 1999; Cranney et al., 2007; Cummings et al., 1993; Cummings et al.,
Numerous studies have consistently highlighted the significant relationship between low bone mineral density (BMD) and an increased risk of fractures, despite variations across different populations (Aspray, 2015) Specifically, a decrease of one standard deviation in BMD is linked to a 1.5 to 3.0-fold rise in fracture risk, influenced by factors such as region, ethnicity, and clinical context (Cummings et al., 1993; Cummings et al., 1995; Kanis et al., 1997; T Nguyen et al., 1993b) In particular, for hip fractures, a similar decrease in femoral neck BMD correlates with heightened fracture risk.
BMD is associated with 2.6 times increase risk, after adjusting for age (Cummings et al.,
Bone mineral density (BMD) measured at a specific site is the most accurate predictor of fractures at that same location, with hip BMD being the best indicator for hip fractures compared to measurements at the spine, wrist, or radius In both research and clinical settings, the lumbar spine and femoral neck are the primary sites for BMD measurement However, BMD at the lumbar spine can be influenced by age-related degenerative changes, leading to artificially elevated readings Consequently, femoral neck BMD is considered the optimal measure for assessing the risk of major osteoporotic fractures and is frequently utilized in fracture risk prediction tools.
BMD measurements at the lumbar spine and femoral neck are probably discordant (Faulkner et al., 1999; Mounach et al., 2009; Pickhardt et al., 2015; Woodson,
2000), leading to uncertainty in fracture risk assessment (Alarkawi et al., 2016)
A decrease of 1 standard deviation in the lumbar spine (LS) T-score, when compared to the femoral neck (FN) T-score, is linked to a 30% rise in fracture risk among low-LS women (Alarkawi et al., 2016) Furthermore, non-osteoporotic women with an osteoporotic LS T-score face a 10% to 13% higher absolute risk of fracture over five years compared to those without osteoporotic readings.
The LS T-score is a valuable indicator for predicting fractures, even if there is discordance in bone mineral density (BMD) measurements across different sites Research shows a strong correlation between BMD measurements at various body locations, supporting the use of BMD from one site to assess fracture risk at another (Alarkawi et al., 2016; Abrahamsen et al., 1997; Imamoto et al., 1998; Namwongprom et al., 2011; Pickhardt et al., 2015).
Osteoporosis significantly increases fracture risk, affecting one-third of individuals aged 70 and over half of those aged 80, yet BMD criteria only account for less than half of fracture cases in older women Interestingly, the highest number of fractures occurs in individuals with osteopenia, which is about five times more than those with osteoporosis This discrepancy has prompted researchers to explore additional risk factors for fractures beyond just BMD measurements.
According to the Kellogg International Work Group, a fall is defined as an event where an individual unintentionally comes to rest on the ground or a lower level, excluding incidents caused by violent impacts, loss of consciousness, sudden paralysis, or seizures Each year, approximately one-third of individuals aged 65 and older experience at least one fall, with those aged 80 and above experiencing multiple falls annually Various risk factors contribute to the occurrence of falls among the elderly.
28 advancing age, sex, races, BMI, alcohol consumption, and sleep problems, physical inactivity, dizziness, and other health problems (Grundstrom et al., 2012)
A history of falls significantly increases the risk of fractures, with around 10% of falls resulting in a fracture Notably, approximately 90% of hip fractures are caused by falls Furthermore, each fall experienced in the past 12 months is linked to a 23% increase in the risk of osteoporotic fractures.
The correlation between falls and reduced muscle strength significantly contributes to increased body sway Individuals who experience falls are 1.5 times more likely to suffer hip fractures than those who do not fall, regardless of factors such as femoral neck bone mineral density (FNBMD).
CURRENT RESEARCH OF FRACTURE RISK ASSESSMENT
Early prediction of fragility fractures is crucial for osteoporosis prevention Over the last two decades, several studies have introduced predictive tools for assessing fracture risk, including FRAX®, the Garvan Fracture Risk Calculator, and QfractureScores, which evaluate fracture risk over 5 to 10 years These tools consider multiple risk factors, such as female gender, advancing age, low bone density, history of falls and fractures, and family history, enhancing their accuracy in predicting fractures.
The Garvan Fracture Risk Calculator, developed by Nguyen et al (2007; 2008), utilizes data from the Dubbo Osteoporosis Epidemiology Study (DOES), which began in 1989 with 3,500 participants in Dubbo The study provided data on 426 fractures in women, including 96 hip fractures, and 149 fractures in men, with 31 being hip fractures This information was used to create nomograms for assessing fracture risk over 5- and 10-year periods.
Figure 2-2: The online tool - Garvan Fracture Risk Calculator
(www.garvan.org.au/bone-fracture-risk, accessed in September 2018)
The models incorporate clinical factors such as age, sex, body weight, and femoral neck bone mineral density (BMD), which is optional They also consider the history of prior fractures after age 50, categorized by the number of fractures (none, 1, 2, or 3 or more), as well as the history of falls in the past year (none, 1, 2, or 3 or more) In cases where femoral neck BMD is unavailable, weighting will be applied Calibration of these models has been conducted exclusively on the Australian population, and there is no adjustment for competing mortality risks.
An independent study evaluated Garvan’s algorithm on a Canadian population, involving 4,152 women and 1,606 men aged 55 to 95 years at baseline The study had a follow-up period of 8.6 years, focusing on 699 individuals with low outcomes.
A recent evaluation of 37 trauma fractures, including 97 hip fractures, demonstrated favorable outcomes in both men and women regarding discrimination and calibration For low-trauma fractures, Harrell’s C indicated a concordance of 0.69 for women and 0.70 for men, while hip fractures showed even higher concordance rates of 0.80 for women and 0.85 for men Over a 10-year prediction period, the observed rates of low-trauma fractures aligned with predicted risks across all subgroups, except for the highest quintile in both genders Similarly, for hip fractures, the observed risk matched the predicted risk in all subgroups except for the highest quintile in women.
FRAX® is a tool developed by the Centre for Metabolic Bone Diseases in Sheffield, UK, to estimate the 10-year probability of osteoporotic fractures, including clinical spine, hip, forearm, and proximal humerus fractures, tailored to specific regions Key risk factors for fractures include prior fragility fractures, age, sex, body mass index, prolonged glucocorticoid use, secondary osteoporosis, rheumatoid arthritis, a family history of hip fractures, current smoking, and alcohol consumption of three or more units per day (Kanis et al., 2009).
Figure 2-3: The online tool (FRAX®) for calculation of 10-year absolute risk of fracture (assessed in September 2018)
A series of meta-analyses identified clinical risk factors (CRFs) from nine global prospective cohorts (Kanis JA, 2007), highlighting that each CRF varies in its relative importance for different types of fracture risk, with bone mineral density (BMD) being the most significant factor influencing management decisions In certain populations, such as institutionalized elderly individuals, fracture risk assessments can be conducted without BMD measurements, utilizing cost-effective options like calcium and vitamin D Notably, FRAX distinguishes itself from other fracture risk assessments by incorporating mortality into its analysis through a competing mortality adjustment approach.
39 may have the same risk of fracture, but different risk of death, thus they will have different the 10-year fracture probability
FRAX assessment results vary significantly across different countries, with the latest version encompassing 47 nations and calibrated to reflect the specific fracture and mortality epidemiology of each This variability raises critical considerations when establishing thresholds for high- and low-risk probabilities Optimal values are recommended based on country-specific health economics studies (Leslie & Lix, 2014).
The QfractureScores tool, available at www.qfracture.org, employs two fracture risk algorithms to assess the individual risk of osteoporotic and hip fractures over a decade Developed using data from the QResearch database, this tool analyzes osteoporotic fracture information collected over 15 years, from January 1, 1993, to June 30, 2008, involving over one million women and men aged 30 to 85 from England and Wales for its validation and development.
The study examined osteoporotic fractures, including hip, wrist, and spine fractures, over a follow-up period of 1 to 10 years, focusing on hip fractures in particular The QfractureScores evaluated various clinical risk factors (CRFs) but excluded bone mineral density (BMD) from their analysis Notably, the significant risk factors differed between sexes and fracture types, with more CRFs identified for women compared to men While certain variables were linked to the risk of osteoporotic fractures, they did not correlate with hip fracture risk The updated 2012 ROC statistics indicated a prediction accuracy of 0.79 for women and 0.71 for men regarding osteoporotic fractures, contrasting with higher accuracies of 0.89 for women and 0.88 for men specifically for hip fractures.
40 well calibrated with predicted risks closely matching observed risks FRAX and QfractureScores’s comparison were conducted in 246 postmenopausal women aged 50–
A study conducted over 85 years across six centers in Ireland and the UK found that the area under the curve (AUC) for fracture discrimination is approximately 0.67 Notably, QfractureScores demonstrated greater accuracy for hip fracture discrimination, achieving an AUC of 0.71 compared to 0.64 for FRAX Further evaluations of QfractureScores in older men and women are necessary The key advantage of QfractureScores is their ability to predict fracture risk without requiring laboratory tests, making them suitable for both clinical applications and self-assessment.
2.3.4 Other predictive models for fracture risk
Many fracture risk estimators and assessment tools exist from various studies, but most have not undergone independent validation Among these is the Fracture INDEX, developed from the Study of Osteoporotic Fractures.
In recent studies, various models have been developed to assess the 5-year risk of hip fractures, including an 11-factor model from the Women's Health Initiative (Aragaki et al., 2007), the FRISK model from the Geelong Osteoporosis Study in Australia (Henry et al., 2006), and a straightforward four-item risk score derived from the Rotterdam Study and the Longitudinal Aging Study Amsterdam (Pluijm et al., 2009).
THE GAP IN LITERATURE
Osteoporosis impacts approximately 20% of postmenopausal women and 10% of elderly men, leading to fragility fractures as a significant consequence Research is currently focused on creating models to identify individuals at high risk for fractures Two primary predictive models used to assess an individual's absolute fracture risk are the Garvan’s fracture risk calculator and the tool developed by Nguyen et al in 2007.
The Fracture Risk Assessment Tool (FRAX) developed by Kanis et al (2007) has shown limitations in its predictive accuracy, with discrimination rates between fracture and non-fracture ranging from 0.61 to 0.85, as noted in validation studies by Sambrook et al (2011) and Sandhu et al (2010) These shortcomings may stem from the models' failure to incorporate genetic factors and the interactions among various risk factors, indicating a need for further enhancements in fracture prediction Additionally, there remain gaps in the existing literature that warrant further exploration.
Genetic factors play a crucial role in determining peak bone mass, bone structure, and the risk of bone deterioration and fragility fractures Research indicates that up to 80% of the variance in bone mineral density (BMD), a significant risk factor for fractures, is influenced by genetics Twin studies reveal that nearly 50% of the variability in fracture susceptibility can be attributed to genetic factors, while up to 40% of the variance in bone loss is also linked to genetics Despite this evidence, specific genes or genetic loci related to bone loss remain unidentified.
Bone mineral density (BMD) is a key indicator of fracture risk and bone loss; however, the role of genetic variants linked to BMD in enhancing fracture prediction accuracy has been uncertain Recent research has identified several genes and loci that are associated with low BMD and an increased risk of fractures, providing new insights into genetic influences on bone health.
42 thousands of single nucleotide polymorphisms (SNPs) via genome-wide association studies (GWAS) (Estrada et al., 2012),(Styrkarsdottir et al., 2008; van Meurs et al., 2006)
A recent meta-analysis of genome-wide association studies identified 56 loci linked to bone mineral density (BMD) (Estrada et al., 2012) Additionally, a study involving children demonstrated a connection between genetic profiling and variations in BMD (Mitchell et al., 2016).
Genetic variants associated with fracture risk are prevalent in the general population, yet their individual effect sizes are minimal, limiting their utility in predicting fractures While bone mineral density (BMD) is a key predictor of fracture risk, the role of BMD-related genetic variants in enhancing fracture prediction remains unclear The translation of these genetic variants into practical fracture risk assessments has not been systematically explored.
2.4.2 Predictive value of risk factors
One of the key challenges in developing prediction models is assessing their prognostic performance, particularly distinguishing between prediction and association, a distinction often overlooked in predictive medicine While traditional models aim to identify statistically significant factors that influence an outcome variable, the true objective of prediction is to find observable factors that can accurately predict new outcomes Consequently, statistically significant factors may not always serve as effective predictors or classifiers, highlighting the importance of focusing on predictive accuracy rather than mere statistical significance.
43 significant but can be a useful predictor (Guyon, 2003) In this thesis, I am more interested in prediction rather than association, and a good predictive model is not necessarily a good associative model
2.4.3 Performance evaluation of predictive models
Recent studies suggest that genetic factors may enhance fracture prediction beyond bone mineral density (BMD) and clinical risk factors However, it remains uncertain if models incorporating an osteogenomic signature outperform those without it The effectiveness of these models is typically evaluated through their discriminatory power, measured by the area under the ROC curve (AUC), which indicates the model's capability to differentiate between individuals at risk of fractures and those who are not (Nguyen et al., 2013; Pencina & D'Agostino, 2015) The AUC is particularly valuable during the initial phases of test assessment (Halligan et al., 2015).
Recent findings indicate that the Area Under the Curve (AUC) is not sensitive to model performance, even with strong predictors, and may not be the best method for evaluating the contribution of genetic markers (Cook, 2008) AUC fails to encompass predicted probability values and model goodness-of-fit, merely summarizing test performance within the ROC space, which is seldom utilized (Lobo et al., 2008) Additionally, AUC does not account for prevalence or the varying misclassification costs associated with false-negative and false-positive diagnoses Consequently, it is inadequate for assessing the clinical impact of genetic profiling on fracture prediction, as changes in ROC AUC may have minimal direct implications for clinicians (Halligan et al., 2015) This limitation necessitates the exploration of alternative metrics to evaluate the integration of genetic profiling into existing predictive models.
In recent years, decision curve analysis (DCA) has emerged as an alternative method for evaluating the prognostic performance of models by assessing net benefits (Vickers et al., 2008; Vickers & Elkin, 2006) This approach involves a trade-off at each risk level, weighing the benefits against potential harms or unnecessary additional testing By quantifying these factors through true positive and false positive rates, DCA allows for a comprehensive evaluation of competing risk predictive models without relying on external economic or cost-effectiveness data.
Utilizing a predictive model for treatment decisions offers both advantages and risks The primary benefit is accurate diagnoses, or true positives, achieved by effectively treating high-risk patients Conversely, potential harms arise when low-risk patients receive inappropriate treatment To assess the overall clinical benefit, one must calculate the net clinical benefit by subtracting potential harms from benefits, while factoring in a treatment threshold function.
The net clinical benefit, along with the cost of misclassification and fracture prevalence, provides a valuable interpretation of clinical impact by highlighting the effects of correct and incorrect diagnoses While the Area Under the Curve (AUC) is beneficial for evaluating model development, the net clinical benefit offers a more relevant assessment of the model's clinical implications for treatment decisions (Halligan et al., 2015) Although the Decision Curve Analysis (DCA) method is commonly applied in cancer research (Barbieri et al., 2011; Ishioka et al., 2012; Vickers et al., 2016), its application in osteoporosis research remains unexplored.
Development of predictive model for fracture is a challenging endeavour One reason is that many factors interactively contribute to an individual's hip fracture
Hip fractures are influenced by several risk factors, including low bone mineral density, a previous history of hip fractures and falls, female gender, older age, lower body weight, insufficient physical activity, reduced muscle strength, high alcohol intake, and smoking.
Statistical models, such as the World Health Organization's fracture risk assessment tool (FRAX®) and the Garvan Fracture Risk Calculator, have been developed to evaluate the risk of hip fractures based on various risk factors The effectiveness of these models is measured using the Area Under the Curve (AUC), which indicates how well the predicted fracture probabilities align with actual fracture occurrences Previous studies have compared these models to assess their reliability in predicting fracture risk.
Recent studies indicate that the Area Under the Curve (AUC) for hip fracture prediction models ranges from moderate (0.7) to good (0.85) However, a notable limitation of these models is their failure to account for potential interactions among risk factors, highlighting an opportunity to enhance the accuracy of hip fracture predictions.
PROPOSED STRATEGY OF OSTEOPOROSIS ASSESSMENT
A recent simulation study indicated that a genetic profile consisting of up to 50 variants, each with a modest effect size (odds ratio of 1.01-1.35), can enhance fracture prediction accuracy by 10 percentage points in the area under the curve (AUC) (Tran et al., 2011) Additionally, genome-wide association studies (GWAS) have identified several common genetic variants linked to bone mineral density (BMD) Notably, a GWAS involving Icelandic, Danish, and Australian populations discovered 74 SNPs associated with BMD (Styrkarsdottir et al., 2008) Furthermore, a meta-analysis of GWAS revealed that more than
70 SNPs were statistically associated with BMD in various ethnicities (Estrada et al.,
This thesis proposes that genetic profiling using polygenic scores enhances the prediction of fractures beyond traditional clinical risk factors in existing models To test this hypothesis, a strategic approach was developed to achieve specific research aims.
An osteoporosis signature is established using 68 genetic variants linked to low bone mineral density (BMD), as identified in a genome-wide association study (Styrkarsdottir et al., 2008) This signature utilizes weighted genetic risk scores (GRS) to assess the genetic predisposition to osteoporosis.
47 constructed for each individual by summing the products of the number of risk alleles and the sex-specific regression coefficients [associated with BMD reported from GWAS]
The study evaluates the contribution of Genetic Risk Scores (GRS) in predicting Bone Mineral Density (BMD) within the general population, analyzing its variance compared to other clinical factors It further investigates the relationship between GRS and fracture risk, focusing on its potential to redefine individual risk categories, regardless of existing risk factors To achieve this, Cox’s proportional hazard models are developed, both incorporating and excluding GRS.
This article compares predictive models for fracture risk, focusing on those with and without genetic profiling, by evaluating discrimination, calibration, and reclassification Discrimination is assessed using the area under the receiver operating characteristics curve (AUC) to determine how well each model differentiates between individuals who will sustain a fracture and those who will not Calibration measures the accuracy of fracture predictions by comparing predicted outcomes with actual observations Additionally, a net reclassification analysis is performed to evaluate the added prognostic value of genetic profiling This analysis estimates the predicted fracture risk for individuals across three risk categories, highlighting the number of individuals reclassified in both models If genetic profiling proves beneficial, the model incorporating it should show an increased fracture probability for those at risk and a decreased probability for those not at risk.
The rate of changes in bone mineral density (BMD) is assessed through individual follow-up measurements, allowing for an analysis of the relationship between genetic risk scores (GRS) and the rate of BMD changes The annual percent change in BMD, denoted as ΔBMD, is calculated to quantify these changes effectively.
This study employs linear regression analysis on individuals with a minimum of two femoral neck bone mineral density (BMD) measurements Multiple linear regression models are created incorporating genetic risk scores (GRS), age, and factors such as BMD, weight, and BMI to predict changes in BMD rates Additionally, a mixed effects model is utilized to assess the variability in BMD change rates associated with the GRS.
The clinical utility of genomic risk scores (GRS) is assessed through decision curve analysis (DCA), which evaluates four predictive models: Model I (Base model) incorporates only clinical risk factors (CRF); Model II (Garvan model) adds femoral neck bone mineral density (BMD) to CRF; Model III (Garvan+GRS) combines CRF, femoral neck BMD, and GRS; and Model IV (Base+GRS) includes CRF alongside GRS The analysis calculates and compares the clinical net benefits, focusing on true positives and false positives across these models.
Artificial Neural Networks (ANN) are utilized to create predictive models for hip fractures in post-menopausal women, aiming to enhance prediction accuracy by analyzing the interactions among various risk factors This study focuses on developing ANN models based on a cohort of hip fracture patients and validating them in a separate cohort Three distinct ANN models are compared: one using only Bone Mineral Density (BMD), another based solely on clinical risk factors, and a third that combines both BMD and clinical risk factors.
Overfitting is a significant concern in model building, which this study addresses by employing 5-fold, 5-times-repeated cross-validation on the training dataset This approach minimizes instability caused by varying local minima during training in each fold To effectively stratify individuals with or without hip fractures, the optimal risk threshold is determined using the Youden J-index, ensuring maximum sensitivity and specificity The consistency between training and test results further validates the model's reliability.
49 that there was no over-fitting in the models The best ANN model is also compared with other traditional models including Garvan model, logistic regression model, and Cox’s model