Assessment of fracture risk in diabetic patients

Abstract

Patients with diabetes often experience reduced bone strength, resulting in a higher fracture risk. This decline and increased susceptibility stem from intricate interactions within the bone microstructure. However, current gold standard methods for assessing bone strength, such as bone mineral density, and widely-used fracture risk assessment tools do not accurately predict fracture risk in diabetic patients. Therefore, it is crucial to incorporate additional indicators that evaluate bone quality and specific markers relevant to diabetes to enhance the accuracy of predictive models. Moreover, the selection of appropriate algorithms for model construction is essential. This review aims to introduce indicators from both imaging examinations and laboratory indicators that hold significant value for inclusion in fracture risk prediction models for diabetic patients. Additionally, this study provides an overview of the research progress in fracture risk prediction models for diabetic patients, serving as a valuable reference for clinical practice.

Introduction

Diabetes is a group of chronic metabolic diseases characterized by long-term hyperglycemia, which can be caused by the disturbance of insulin secretion, insulin resistance, or both [1]. The global trend of population aging has contributed to a continuous increase in the incidence of diabetes [2]. The global prevalence of diabetes among adults between the ages of 20 and 79 years was estimated to be 8.8% in 2015 and is expected to increase to 10.4% by 2040 [3]. Studies have shown an approximately 11% prevalence rate of diabetes in adults in China [4]. Long-term exposure to chronic hyperglycemia in patients with diabetes mellitus (DM) can lead to multiple organ dysfunction, which is harmful to the eyes, kidneys, nerves, heart, blood vessels and musculoskeletal system [5]. The risk of fractures in patients with DM is significantly higher compared to that in healthy individuals. For elderly individuals, fractures not only impact quality of life but also potentially endanger life, particularly in cases of hip and spinal fractures.

The increased risk of fractures in patients with DM is attributable to various factors. Direct causes include reduced bone mineral density (BMD) and structural bone damage associated with diabetes, while indirect factors elevate the risk of falls, which further increases fracture risk (Fig. 1). Research by Westgard demonstrated that both type 1 diabetes mellitus (T1DM) and type 2 diabetes mellitus (T2DM) are associated with an elevated risk of fractures [6]. BMD is considered the gold standard for diagnosing osteoporosis and plays a crucial role in assessing fracture risk. Due to the decrease of BMD in patients with T1DM, the risk of fracture increases [7]. However, the BMD of patients with T2DM is generally normal or increased, and fractures are as common as in patients with T1DM, which may be due to a decrease in bone strength [8]. Furthermore, several major hypoglycemic drugs such as insulin, sulfonylureas, and sodium-glucose cotransporter 2 inhibitors are associated with an increased risk of hypoglycemia, which can contribute to falls, particularly among elderly patients. The heightened risk of fractures has emerged as a significant health concern in patients with DM, underscoring the challenge of relying solely on BMD for a comprehensive assessment of bone health.

Fig 1.

Three main reasons for the increased risk of fracture in patients with DM

The emergence of machine learning has introduced new avenues for predicting fracture risk among diabetic patients. With rapid advancements in machine learning, widely utilized algorithms such as support vector machine (SVM), logistic regression (LR), and random forest (RF) have found extensive applications in the medical domain. These applications encompass early disease prediction, precise diagnostics, medical image analysis, and the formulation of personalized treatment plans, all of which have yielded significant outcomes. For instance, Wu et al leveraged machine learning to develop a tool for predicting osteoporosis risk in patients with T2DM [9]. We will summarize the role and performance of imaging, laboratory examination and machine learning in fracture prediction in subjects with diabetes.

Imaging examination

Dual X-ray absorptiometry

Dual X-ray absorptiometry (DXA) is a scanning method used to assess bone strength. It can measure several bone indicators, including BMD, trabecular bone score (TBS), vertebral fracture assessment (VFA), and body composition analysis. BMD is considered the gold standard for diagnosing osteoporosis and is commonly used in predicting fracture risk.

Nevertheless, BMD does not accurately reflect fracture susceptibility in individuals, particularly those with T2DM. Addressing this limitation requires a comprehensive approach that integrates multiple bone parameters for a more detailed assessment of fracture risk. The necessity for developing and implementing novel bone measurement tools arises from the inherent constraints of DXA. While DXA offers simplicity, noninvasiveness, convenience, and cost-effectiveness, it lacks the ability to capture the complex three-dimensional characteristics of the skeleton.

Bone mineral density

BMD is effective in predicting fracture risk in patients with osteoporosis, but exhibits limited efficacy in predicting fracture risk in patients with DM and osteoporosis. Studies have indicated a substantial correlation between T1DM and a reduction in systemic BMD [10]. However, the risk of fragility fractures in T1DM patients surpasses predictions made solely by the DXA approach [11, 12]. In contrast, BMD values in T2DM patients are typically normal or elevated, rendering them inadequate for reflecting fracture risk in this population [13]. DXA measurements have revealed increased BMD in overweight and obese adults, suggesting that increased body weight may mitigate skeletal mechanical insufficiency. Furthermore, BMI exhibited a positive correlation with BMD and a negative correlation with fracture risk. Hence, a higher BMI may contribute to the relatively elevated BMD and diminished fracture risk observed in patients with T2DM [14].

Trabecular bone score

TBS is a texture parameter used to assess bone microstructure by analyzing variations in pixel grayscale within spine DXA images, providing valuable insights into bone quality [15]. Lower TBS values indicate compromised bone structure and reduced strength. TBS, assessed using the widely available DXA method, does not necessitate additional devices and, like BMD, independently predicts fracture risk [16]. Multiple studies underscore the potential of the TBS for predicting fracture risk among diabetic patients [16–20]. Interestingly, TBS effectively reveals differences in bone microarchitecture and quality, even when BMD levels are identical, thereby partially addressing the limitations of BMD in assessing bone quality. Research conducted by Koromani et al. revealed that patients with T2DM had lower TBS compared to controls, particularly showing a significant reduction in TBS among those in early stages of diabetes [21]. Dule et al [22] discovered that despite normal or increased BMD in women with T2DM (N=126 vs 117 in an age-matched control group), TBS assessments revealed damage to bone micro-structure. Similarly, Ubago-Guisado et al [23] found that patients with T2DM exhibited significantly higher values for trabecular and cortical parameters of aBMD and 3D-DXA compared to the control group. However, TBS values were significantly lower in the T2DM group compared to the control group. In conclusion, incorporating TBS improves the predictive accuracy of fracture risk assessment. Nevertheless, TBS has limitations, such as the lack of clinical studies directly comparing T1DM and T2DM regarding TBS disparities, incomplete insights into how diabetes treatments affect TBS and the constraint of evaluating bone microarchitecture solely at the lumbar spine.

Vertebral fracture assessment

VFA is a rapid and low-radiation method employed for the detection of vertebral fractures, utilizing DXA to visualize the thoracic and lumbar spine [24]. The VFA exhibits good sensitivity (87-93%) and specificity (93-95%) in detecting moderate and severe vertebral fractures. However, its performance diminishes for identifying mild fractures and in the presence of conditions such as scoliosis or osteoarthritis [25]. Studies have established that the presence of morphometric vertebral fractures independently increases the risk of subsequent spinal or hip fractures, irrespective of BMD [26, 27]. Consequently, VFA plays a crucial role in predicting vertebral fracture risk, especially in patients with normal BMD. Both T1DM and T2DM patients exhibit a greater risk of vertebral fractures than nondiabetic individuals [6, 21, 28, 29]. In patients with diabetes who have a normal BMD, the VFA is a valuable tool for predicting the risk of vertebral fractures.

Body composition

BMI is widely used to assess overweight and obesity. However, it may not accurately reflect variations in body composition proportions, since individuals with identical BMI values can differ significantly in their distribution of fat and lean mass [30, 31]. DXA is a well-validated technique for body composition analysis. In a single full-body DXA scan, the three major components of the human body—fat mass, lean body mass, and bone mineral mass—can be accurately measured with high precision and in a short scanning time [32]. Numerous studies have indicated that both lean mass and fat mass play pivotal positive roles in bone mineral mass among young and middle-aged adults [33, 34]. Nevertheless, it has also been demonstrated that fat mass independently exerts an adverse effect on bone mass [35, 36]. Although the findings are not entirely consistent, the majority of studies suggest that lean mass plays a dominant role in maintaining bone mass.

Quantitative computed tomography

Quantitative computed tomography (QCT) is a method employed to assess the microstructure of the distal radius and tibia bones [37]. In contrast to DXA, QCT is a three-dimensional imaging modality capable of distinguishing between trabecular and cortical bone compartments, while also quantifying cortical bone thickness and bone geometry [38]. In a study by Heckelman and colleagues utilizing CT, it was found that the minimum cross-sectional area of the femoral neck was significantly lower in postmenopausal women with T2DM than in nondiabetic controls [39]. Most current studies utilizing QCT to assess bone mass in T2DM patients are cross-sectional observational studies and lack prospective investigations. Finite element analyses of proximal femurs based on computed tomography scans (CTFEA) have been developed to predict femur stiffness and hip fracture risk. CTFEA has demonstrated superiority over DXA in various studies [40–45]. Ram Naresh Yadav et al. demonstrated that damage-based finite element analysis can aid in predicting fracture risk in T2DM patients by predicting changes in bone mechanical properties and pathological alterations in mechanical response [46].

High-resolution peripgeral quantitative computed tomography

High-resolution peripheral QCT (HR-pQCT), similar to QCT, is employed for three-dimensional imaging to assess volumetric BMD (vBMD), bone geometry, and microstructure specifically at the distal radius and distal tibia [38]. The high resolution of the 3D bone images generated by HR-pQCT mitigates challenges in measuring vBMD and other parameters. This imaging technique allows for the separate evaluation of cortical and cancellous bone, demonstrating a high sensitivity to subtle changes in bone volume [47, 48]. Studies have revealed that there are no significant differences in the vBMD or bone microarchitecture of the distal radius between patients with T1DM and controls. However, the trabecular density, trabecular thickness, trabecular volume fraction, and overall BMD were significantly lower, indicating trabecular damage in T1DM patients [49–53]. In a longitudinal HR-pQCT study of postmenopausal women with T2DM, patients in the diabetes group exhibited greater tibial cortical porosity, potentially adversely impacting bone strength [54]. Several small-scale studies have employed HR-pQCT to assess differences in bone microarchitecture between T2DM patients and nondiabetic subjects, but the findings are inconclusive, possibly because of the small sample sizes used [55–57].

Histomorphology

Morphological analysis of bone tissue enables direct assessment of bone remodeling rates at the tissue level, and micro computed tomography (micro-CT) is a valuable tool for evaluating bone microstructure [58]. Animal model studies have consistently demonstrated that rodent models of T1DM and T2DM exhibit low bone turnover, microstructural degradation, and reduced strength [59–62]. However, due to the invasive nature of this technique, there are limited studies involving human subjects with diabetes, and the results vary across different studies. A study using micro-CT to analyze the tibia of patients with T2DM showed that patients with T2DM had high tibial cortical porosity and impaired bone quality [63]. Another study using micro-CT to measure the proximal femur in patients with T2DM also showed that patients with T2DM had greater cortical porosity and an increased tendency to fracture. It should be noted that the patient samples used in both studies were taken from cadaveric specimens [64].

Laboratory examination

Micro indentation test

The micro indentation test assesses bone condition by applying pressure with a stainless-steel probe featuring a spherical tip and observing the depth of indentation. A deeper indentation indicates a decrease in the resistance of the bone to mechanical stress. This technique comprises cyclic micro indentations and impact micro indentations. Impact microindentation, while not noninvasive, aids in detecting skeletal conditions in diabetic patients and assists in assessing skeletal changes. This method generates a bone material strength index (BMSi), where a lower value indicates reduced fracture strength [65]. Nilsson et al reported a lower BMSi in T2DM patients [66]. Farr JN et al. observed significantly lower BMSi in T2DM patients with similar BMD to the control group. They also identified a negative correlation between HbA1c levels and BMSi over the past decade [67]. Furthermore, Furst JR noted a negative correlation between BMSi and the accumulation of advanced glycation end products [68].

Laboratory indicators

Bone turnover markers

Bone turnover markers can be divided into the bone formation index and the bone resorption index. Bone formation indicators reflect osteoblast activity and bone formation status and include bone-specific alkaline phosphatase (b-ALP), type 1 collagen N-terminal propeptide (P1NP), type 1 collagen C-terminal propeptide (P1CP), and osteocalcin (OC).

The bone resorption indices represent osteoclast activity and bone resorption levels and include hydroxyproline (HOP), pyridinamine (Pyr), deoxypyridinoline (DPD), carboxyl end cross-linked peptide (CTX) of type 1 collagen, amino end cross-linked peptide (NTX) of type 1 collagen, and tartrate-resistant phosphatase 5b (TRAP5b).

Patients with diabetes exhibit alterations in various bone turnover markers. In a study, both T1DM and T2DM patients showed a significant decrease in osteocalcin and C-terminal telopeptide of type 1 collagen (CTX-1) compared to those in the control group. However, there were no notable differences in alkaline phosphatase (ALP), N-telopeptide of type 1 collagen or hydroxyproline [69]. Furthermore, the parathyroid hormone (PTH) levels in T2DM patients tended to be lower than those in the control group [70].

Interestingly, individuals with T1DM demonstrated lower levels of OC, P1NP, and S-nuclear factor receptor activator kappa β ligand (RANKL) than did those with T2DM [15]. Additional research is required to elucidate the impact of changes in these bone turnover markers on fracture risk in patients.

Biochemical measurements

Compared with those of normal people, a number of physiological indicators of diabetic patients change, some of which can be used as indicators to evaluate the bone status of diabetic patients. First, glycosylated hemoglobin (HbA1c) reflects the average blood glucose level of diabetic patients in the past few months. Long-term hyperglycemia can affect bone health. A cross-sectional study of men with T2DM showed that HbA1c levels ≥ 7% were associated with low bone turnover [71].

Studies indicate that diabetes patients have lower levels of 25-hydroxyvitamin D (25-(OH)-VD) and estimated glomerular filtration rate (eGFR) than those without diabetes, which further negatively impacts skeletal health [9]. In addition, some studies used T2DM patients without osteoporosis or fracture as the control group, and T2DM patients with major site fracture or osteoporosis as the experimental group. More than ten routine laboratory indices were detected and compared, and several indicators with significant differences were obtained. For example, eGFR [9, 72, 73], 25-(OH)-VD [9], HbA1c, hemoglobin, prealbumin, albumin [74], high-density lipoprotein cholesterol (HDL-C), apolipoprotein A, total cholesterol (TC) [75], systolic blood pressure, the urinary albumin-to-creatinine ratio (uACR) [72] and serum albumin. These indicators are suggested to be included in the final prediction model as predictive variables, but due to the complex mechanism and extensive effects of diabetes, more studies are needed to clarify the specific mechanism of bone damage in patients with diabetes. Moreover, more relevant indicators were screened to predict the risk of fracture in patients.

Machine learning

The commonly utilized clinical fracture risk prediction model FRAX has proven effective in forecasting fracture risk over the next 10 years. However, its efficacy diminishes when applied to patients with DM, particularly in underestimating fracture risk in those with T2DM [76]. The Garvan Fracture Risk Calculator (FRC) model, similar to the FRAX model, exhibits limitations in predicting fracture risk among individuals with diabetes. The Q fracture model is unique because it includes diabetes as a predictive variable for fracture risk, yet its applicability in the Asian population is suboptimal [77]. In attempts to predict fracture risk in diabetic patients, Sijia Chu et al explored various models, such as decision tree (DT), gradient boosting decision tree (GBDT), LR, RF, SVM, extreme gradient boosting (XGBoost), and probabilistic classification vector machine (PCVM) models. Their findings indicate that PCVM delivers the most effective results [78]. However, it is essential to note that the relevant features selected in this study were obtained using a form of hypothesis testing akin to linear regression. Alternative methods such as the weighted Gini index may yield different correlation features and impact the final predictive effectiveness of the model.

Classical fracture risk prediction model

Fracture risk assessment tool

In 2008, the World Health Organization (WHO) collaborating Centre in Sheffield, UK, released FRAX a computer-based algorithm available at http://www.shef.ac.uk/FRAX, to estimate an individual's 10 years risk of hip and major osteoporotic fractures, including clinical spine, distal forearm, and proximal humerus fractures. The FRAX incorporates seven binary clinical risk factors, which include prior fragile fractures, parental history of hip fractures, smoking (Tobacco smoking is entered into FRAX as yes or no depending on whether the patient currently smokes tobacco), systemic glucocorticoid use, high alcohol intake (The input to FRAX asks for a positive entry if the patient takes three or more units of alcohol daily), rheumatoid arthritis, and other secondary causes of osteoporosis [79]. Although FRAX offers a straightforward approach to estimate fracture risk, it has certain limitations. Notably, parameters associated with diabetes are not included in the FRAX, potentially leading to an underestimation of fracture risk in individuals with T2DM [80]. Presently, several FRAX-like fracture prediction tools have been proposed, demonstrating enhanced predictive performance compared to FRAX (Table 1).

Table 1.

Comparison of FRAX-like Fracture Risk Assessment Tools^{[a] [b] [c] [d]}

Reference	Tool	Research population	Area under the curve
Bonaccorsi, Gloria et al	Derived FRAX (DeFRA)	postmenopausal women with T2DM, n=119; consecutive healthy postmenopausal women, n=118	0.89
	Nonadjusted FRAX		0.73
	Adjusted FRAX		0.80
Kong, Xiaoke et al	Chinese diabetes fracture risk (CDFR)	T2DM (age, 55.1±11.9 yr.) n=1730, 66% were male	0.80
	Nonadjusted FRAX		0.75
	RA-adjusted FRAX		0.75
	Age-adjusted FRAX		0.76
Wendy A Davis et al	FDS1 hip fracture risk calculator	T2DM (age, 65.0±10.0 yr.) n=1251, 48.8% were male	0.84
	Q Fracture		0.82
	Nonadjusted FRAX		0.80
Chuan Fengning et al	Diabetic Fracture Risk Assessment (DFRA)	T2DM (age, 64 yr.) n=1855, 45.5% were female	0.80
	RA-adjusted FRAX		0.73
	Nonadjusted FRAX		0.74
	T-adjusted FRAX		0.71
	Age-adjusted FRAX		0.74

[a] BONACCORSI G, MESSINA C, CERVELLATI C, et al. Fracture risk assessment in postmenopausal women with diabetes: comparison between DeFRA and FRAX tools [J]. Gynecol Endocrinol, 2018, 34(5): 404-8.

[b] KONG X K, ZHAO Z Y, ZHANG D, et al. Major osteoporosis fracture prediction in type 2 diabetes: a derivation and comparison study [J]. Osteoporos Int, 2022, 33(9): 1957-67.

[c] DAVIS W A, HAMILTON E J, BRUCE D G, et al. Development and Validation of a Simple Hip Fracture Risk Prediction Tool for Type 2 Diabetes: The Fremantle Diabetes Study Phase I [J]. Diabetes Care, 2019, 42(1): 102-9.

[d] CHUAN F, GAO Y, LIAO K, et al. A simple fragility fracture risk score for type 2 diabetes patients: a derivation, validation, comparison, and risk stratification study [J]. Eur J Endocrinol, 2023, 189(5): 508-16.

Q Fracture model and Garvan FRC model

FRAX has been validated in 26 studies across 9 countries, Garvan FRC in 6 studies across three countries, and Q Fracture in 3 studies conducted in the United Kingdom and the Republic of Ireland [81]. Notably, the Q Fracture model stands out as the sole risk prediction model incorporating diabetes as a predictor variable for fracture risk. However, its practical application is hindered by the model's restricted use due to the incomplete consideration of risk factors [77]. The Garvan FRC, developed using data from the Dubbo Osteoporosis Epidemiology Study (DOES) involving both women and men, lacks the incorporation of diabetes as a predictor variable [82, 83]. Additionally, no study has assessed the performance of the Garvan FRC model in patients with T1DM or T2DM.

Common algorithms for modeling

DT, GBDT, XGBoost and RF

A DT is akin to a tree-like structure, resembling a flow chart where each internal node denotes a judgment condition, typically a test on a specific attribute. Branches emanating from these nodes represent the outcomes of the tests, while each leaf node signifies a distinct category. The crux of the algorithm lies in selecting the attribute with the highest information entropy, deeming it the optimal attribute for decision-making [84]. DTs stand out as widely embraced predictive modeling tools due to their commendable interpretability, accuracy and efficient prediction capabilities and often require minimal tuning efforts [85]. In a retrospective cross-sectional exploration of factors influencing vertebral fractures, researchers employing decision trees discerned a notable revelation: the pivotal determinant of vertebral fractures leaned toward TBS rather than toward BMD [86]. However, it is crucial to acknowledge that DT possess inherent limitations, notably their constrained ability to model intricate data and complex feature relationships. In scenarios demanding a nuanced understanding of such complexities, opting for alternative and more advanced machine learning techniques may prove to be more apt.

GBDT is a machine learning algorithm rooted in the classical DT approach. Consequently, GBDT has the advantages of high prediction efficiency, accuracy, and interpretability [87]. However, the advent and progression of big data present novel challenges for GBDT, particularly in striking a balance between accuracy and computational efficiency. Traditionally, GBDT processes all data instances for each input feature to estimate the information gain across all potential split points. This approach, while effective, encounters significant computational complexity proportional to both the number of features and the number of samples. The consequence is prolonged processing times, particularly when dealing with substantial volumes of big data [88]. As a result, addressing the computational demands of big data poses a key challenge for GBDT.

XGBoost, a variant of the GBDT algorithm and part of the ensemble learning paradigm, outperforms its predecessor by virtue of heightened efficiency and a plethora of algorithmic and engineering enhancements [89]. Notably, XGBoost is a more streamlined and optimized iteration of GBDT. In the pursuit of further advancements, Guolin Ke and colleagues introduced LightGBM, a novel GBDT algorithm distinguished by two groundbreaking techniques: gradient-based one-sided sampling (GOSS) and exclusive feature bundling (EFB). Notably, research has demonstrated that LightGBM surpasses both XGBoost and stochastic gradient boosting (SGB) in terms of computing speed and memory consumption [88]. Despite these notable strides, there is a dearth of research on the application of LightGBM in predicting fracture risk. The current landscape presents an opportunity for exploration, and delving into the potential of LightGBM in the context of fracture risk prediction could yield valuable insights into its adaptability and effectiveness within this specific domain.

The RF algorithm represents an extension of the bagging integration of DT. In this approach, tests inserted into each node are chosen from a randomly selected subset of potential tests, rather than considering all of them. This infusion of randomness typically contributes to an enhancement in the overall performance of the integration [85]. Lisa Langsetmo and her colleagues discovered that RF harness the inherent flexibility of tree-based models, enabling them to model nonlinear relationships, as well as pairwise and higher-order interactions. However, despite this flexibility, their study revealed that employing RF models to predict hip fractures did not outperform simpler models lacking the sophisticated flexibility intrinsic to RF [90]. This nuanced insight underscores the importance of considering the specific characteristics of the problem at hand when selecting a predictive modeling approach, as the enhanced flexibility of certain models may not always translate to superior predictive performance in every scenario.

LR

LR involves training a model with a given set of n data (training set) and subsequently classifying one or more given sets of data (test set). Each set of data is represented as a P-dimensional vector, and the model trained by LR operates within this P-dimensional space [91]. Following classification, a category or a similar classification emerges in one of the dimensions. Reference [92] delves into the detailed definition of machine learning and explores the distinctions from statistical modeling, although the differences between the two may not be readily apparent [93]. Through a comparison of clinical prediction models, Evangelia Christodoulou and her colleagues discovered that when there is a low risk of bias, the area under the curve (AUC) of LR models and machine learning models in predicting clinical risk tend to be similar. However, machine learning outperforms LR when there is a high risk of bias [91]. In a cross-sectional study involving 808 patients with T2DM, researchers employed LR analysis to assign weights to individual risk factors. This approach was used to develop a scoring tool for predicting vertebral fractures in patients with T2DM. The study identified significant risk factors, including the duration of diabetes and BMI, which were translated into risk scores. This score proved effective in predicting vertebral fractures in patients with T2DM and was noteworthy for its validity, even in the absence of BMD data provided by DXA [75].

SVM and PCVM

The SVM training algorithm constructs a model that aims to maximize the gap between two categories by mapping training examples to points in space. Subsequently, new examples are also mapped into this space, and their categorization is determined based on the side of the gap they fall on [94]. SVM stands out for its robustness in recognizing subtle patterns within complex datasets compared to other machine learning methods. However, certain drawbacks, such as nonprobabilistic outputs and a linear correlation between the number of support vectors and the size of the training set, have led to dissatisfaction, particularly when dealing with large datasets [95, 96]. Yosibash Z et al developed an SVM algorithm based on finite element analysis for assessing hip fracture risk in both T2DM and non-T2DM patients. The AUC for predicting hip fracture risk in non-T2DM patients was 0.84, that for predicting hip fracture risk in T2DM patients was 0.92, and the combined AUC was 0.88 [97]. Wang et al demonstrated that osteoporosis diagnosis in T2DM patients can be achieved by analyzing serological test results using an SVM model based on bone metabolism biomarkers. The SVM model, which incorporates features such as sex, age, BMI, total-type collagen amino-terminal extension peptide, and OC, yielded the best results with an accuracy of 88% [98].

In the realm of advanced SVM variants, the PCVM synthesizes the strengths of both the SVM and the relevant vector machine, providing a sparse Bayesian solution to classification problems. Sijia et al. study confirmed the superiority of the PCVM over the SVM in predicting fracture risk in diabetic patients. This highlights the ongoing refinement and development of machine learning algorithms to address specific challenges and optimize performance in healthcare applications.

Common problems with model overfitting

The challenge of overfitting is prevalent in predictive models and arises when a model becomes overly complex, the sample size is too small, or the training data are overemphasized. Overfitting manifests as a situation where the model excels on the training data but falters when faced with new, unseen data. However, various techniques are employed to address and mitigate overfitting issues. For instance, DTs counteract overfitting by constraining the depth of the tree, setting minimum sample requirements for nodes, and employing pruning techniques to eliminate unnecessary branches; GBDTs enhance model performance by amalgamating multiple weak learners, often in the form of shallow decision trees. RF alleviates overfitting by constructing numerous decision trees and aggregating prediction results through voting mechanisms.; LR, SVM, XGBoost, and PCVM manage model complexity and mitigate overfitting by introducing regularization parameters [78]. These strategies reflect a nuanced and diverse set of approaches aimed at achieving a balance between model complexity and generalization, ultimately improving a model's performance on new, unseen data.

Conclusion:

The ability of various factors, such as BMI, smoking status, alcohol intake, the use of diabetes medications, and the presence of diabetes-related complications, to predict fracture risk in diabetic patients is complex. This complexity underscores the need for advanced and comprehensive assessment models. An excellent fracture risk prediction model not only needs to adopt appropriate algorithms but also needs to incorporate as many relevant predictive variables as possible while remaining simple. In addition to incorporating gold standard indicators such as BMD for predicting fracture risk, other indicators should be considered. For example, the TBS, VFA, QCT, body composition analysis and several other laboratory indicators have been used.

Integrating these factors into the predictive model can better reflect patients’ bone health status, thereby more accurately assessing their risk of fractures.

Additionally, with the advancement of telemedicine and mobile health technologies, we can anticipate the emergence of more convenient and real-time systems for monitoring and intervening in fracture risk. Through smartphone apps, wearable devices, and remote monitoring technology, patients can easily track their bone health indicators and receive personalized health advice and intervention measures to prevent fracture events in a timely manner. We can better prevent and manage the risk of fractures in diabetic patients, ultimately improving their quality of life and health outcomes.

Funding

Project funded by the Sichuan Provincial Department of Science and Technology (Grant Number: 2024NSFSC0590).

Data availability

Not applicable.

Declarations

Conflict of interest

No conflict of interest exists in the submission of this manuscript, and the manuscript is approved by all authors for publication. I would like to declare on behalf of my co-authors that the work described was original research that has not been published previously, and not under consideration for publication elsewhere, in whole or in part. All the authors listed have approved the manuscript that is enclosed.