There is a great demand for the development of approaches to improve assessment of bone strength and associated hip fracture risk. Here, we present a multivariate analysis that highlights the potential contribution of combining bone tissue water parameters to predict proximal femur stiffness. The importance of bone composition and geometry to bone mechanics is well established [9,10,11,12]. This study is novel because of its multivariate approach, which reveals the combined role of different bone properties related to the ultimate mechanical function of the proximal femur, including different types of tissue water content in both cortical and trabecular bones. In this study, we showed that multivariate regression models to predict bone stiffness based on the combination of multiple tissue properties were markedly stronger than models using individual properties as single explanatory variables. This outcome sheds new insights to advance our understanding of the multifactorial role of tissue properties in proximal femur stiffness, with emphasis on the role of cortical and trabecular bone water, beyond the role of any individual properties alone.

We used PLS leave-one-out cross validation to create and validate the multivariate models. This machine learning technique is reliable and widely used to evaluate model fitting and predictive ability by performing a k-fold cross-validation in which the number of k subsets is equal to the number of observations in the dataset [27,28,29]. In other words, it creates k–1 iterative models (training sets) in which one data point (testing sets) is excluded at a time and used as input to validate its independent prediction by the model created without it. In comparison with MLR, PLS offers advantages such as suitability for models with small sample sizes and numerous explanatory variables, and its ability to handle multicollinearity among co-variates without compromising model performance [27,28,29]. In addition, whereas MLR relies on adjusted R2 as a metric to evaluate the goodness-of-fit of models while accounting for the number of predictor variables, PLS typically utilizes the cross-validated R2 (also expressed as Q2) [27,28,29]. Cross-validated R2 is preferred in PLS because it not only reflects the fit of the model to the training data but also provides a measure of predictive accuracy and robustness when applied to unseen data, thus mitigating the risk of overfitting that adjusted R2 cannot detect [35]. Outcomes from this approach showed that the combination of cortical and trabecular total and tightly bound water was a significant contributor to markedly improved models, with water content being inversely correlated with bone stiffness. This suggests that tissue water content may be an important potential biomarker of bone quality.

This result is in line with several previous studies that have shown an association between tissue water content and bone mechanics in both clinical [36,37,38] and pre-clinical studies [19, 39]. Here, our findings expand the current knowledge by assessing water content in both cortical and trabecular tissues of the femoral neck, a region prone to fragility fractures and thus of great clinical relevance [2]. To help interpreting the findings of this study, we need to delve into how water can be found in bone tissue [33, 40, 41]. Water comprises around 20% of bone tissue in volume, being distributed in different compartments throughout the tissue. Free or pore water is located within bone porosity at the microscale, such as in the Haversian canals and the lacuno-canalicular system; loosely bound water is associate with the surfaces and interfaces between collagen fibrils and mineral crystals at the submicron and nanoscales; and tightly bound water is found as part of the collagen triple helical structure or the mineral lattice at the molecular and atomic scales. Loss of bone water by dehydration has been shown in many studies to lead to an increase in stiffness [41], which is in accordance with our overall results.

Interestingly, it is possible to delve deeper into this inverse relationship between bone water and stiffness considering water content in different tissue compartments. Free water content is a direct surrogate of cortical bone porosity [20, 42], such that a greater free water content directly reflects a greater bone porosity, which has been linked to a decrease in bone stiffness [40]. Loosely bound water plays key roles in transferring loads by promoting sliding between collagen and mineral; a reduction in this interface water content impairs this process, resulting in stiffer bones [40, 43]. Tightly bound water is thought to provide structure to the length of collagen fibrils [38, 44] and order and mechanical stability to mineral–mineral interfaces via their surface hydrated layers [45, 46]; a loss of tightly bound water may lead to a disruption in the mineralized matrix at the molecular and crystal level, leading to stiffer and more brittle bones [47].

Here, using NIR spectroscopy, we describe water content in two different bone tissue compartments: total water (which includes free and loosely bound water) and tightly bound water associated with collagen and mineral [33], assessed both in cortical and trabecular tissues. Our results not only corroborate previous findings, but also suggest that assessment of water content in different tissue compartments may provide a valuable approach to improve prediction of bone mechanical function. This is illustrated, for example, when a multivariate PLS model using as input the combination of total water content of both cortical and trabecular bone and tightly bound water content of trabecular bone more thoroughly explains whole-bone stiffness (i.e., leads to a higher R2 value) than models based on analysis of water on an individual tissue compartment. Overall, this shows that to elucidate how bone water assessment may inform on bone mechanical function, multiple water pools within bone tissues should be taken into consideration.

In standard clinical practice, BMD scores are often combined with other clinical risk factors (such as female sex, older age, low body mass index) via the predictive algorithm FRAX Fracture Risk Assessment Tool to predict 10-year fracture risk in individuals aged 50 years and older [8, 48]. However, the sensitivity of FRAX to predict fracture risk remains low at ~ 50%, missing nearly half of the women at high risk of hip fractures [7, 8]. Here, our study is motivated by the need to improve fracture risk prediction; however, current applications of infrared spectroscopy for a comprehensive assessment of bone composition are not suitable for in vivo evaluations. Thus, further consideration is necessary to discuss a translational pathway in which tissue compositional properties, especially bone water content measurements, may be implemented in clinical systems to predict fracture risk. Bone water is particularly interesting because it can be assessed in clinical settings by ultra-short echo time (UTE) magnetic resonance imaging (MRI) [20, 37, 42], with MRI outcomes correlated to those obtained by NIR spectroscopy [20, 42]. Specifically, UTE-MRI can be used clinically to assess free water, loosely bound water, and total water in bone, while further MRI-based approaches are being explored to assess tightly bound or structural water [40]. Moreover, with advances of new spectroscopic methods using fiber optic probes, including Raman and NIR spectroscopy, clinical assessment of bone tissue composition and quality may be on the horizon [39, 40, 49, 50]. Ultimately, the translational potential of this study is on laying a foundation for future studies aiming to combine and incorporate different types of bone water content metrics to improve multivariate algorithms in clinical models for predicting hip fracture risk.

It is also important to discuss other limitations of this study. Although the small sample size available (N = 12) is not optimal for robust statistical modeling, our results reveal a promising and novel direction towards the better understanding and assessment of bone mechanical function at the proximal femur. This may lay a foundation for future studies with larger samples sizes and different cohorts of donors, focusing for example on postmenopausal women at higher risk of typical osteoporotic femoral neck fractures [51]. Another limitation of this study is that we used stiffness instead of ultimate load as the parameter for correlation of bone biomechanics. Stiffness was chosen because it can be obtained without loading the bone to failure, allowing to preserve the intact femoral neck for structural and compositional analyses after mechanical testing. Stiffness is often used as a surrogate for bone strength due to the significant correlation between these parameters [52,53,54,55,56]; however, in order to directly investigate the relationship between tissue properties and bone strength, future studies may consider assessing the ultimate load instead of stiffness.

This study provides new insights into the complex relationship between bone tissue water content and bone stiffness. By investigating total and tightly bound water across cortical and trabecular tissues from femoral neck samples, we found that higher bone water content was consistently linked to lower proximal femur stiffness. Importantly, our multivariate models, which incorporated multiple tissue properties as explanatory variables, revealed a synergistic effect: the combination of water content across different tissue compartments resulted in better predictions of stiffness than any individual metrics alone. These findings suggest that bone water may serve as a valuable biomarker for assessing bone quality, offering new opportunities to better understand bone fragility and fracture risk. Additionally, this research sets the stage for future clinical and pre-clinical studies, highlighting the potential for incorporating bone water measurements into advanced diagnostic tools and in the design of tissue-engineered materials optimized for bone mechanical performance.