With the advancement of antiretroviral therapy (ART), the life expectancy of people with HIV (PWH) has significantly increased (Negredo et al., 2018). However, non-AIDS-defining events have become increasingly prominent, now representing major contributors to morbidity and mortality among PWH receiving suppressive ART (Maartens et al., 2014). These conditions primarily encompass cardiovascular diseases, non-AIDS-defining cancers, metabolic syndrome, liver and kidney diseases, as well as bone-related complications, such as bone loss and osteonecrosis (Dominguez-Molina et al., 2016).

Bone health in PWH is significantly influenced by a complex interplay of traditional risk factors, such as aging, comorbidities, low body mass index (BMI), smoking, alcohol abuse, and vitamin D insufficiency. Additionally, HIV-related factors, including the infection itself, low CD4 count, and antiretroviral therapy (ART) (Rey et al., 2015; Biver, 2022; Compston, 2016), further complicate this issue. Notably, certain ART medications, particularly tenofovir disoproxil fumarate and protease inhibitors (PIs), have been shown to exacerbate the loss of bone mineral density (BMD) associated with HIV infection. Research indicates that PWH can experience an additional 2%–6% reduction in BMD within the first 1–2 years of initiating ART, a rate of bone loss comparable to that observed in postmenopausal osteoporosis (McComsey et al., 2010; Han et al., 2020; Compston, 2014).

Integrase strand transfer inhibitors (INSTIs) have emerged as a new frontline option for PWH (Scarsi et al., 2020). However, tenofovir disoproxil fumarate (TDF) and protease inhibitors (PIs) continue to be the first-line ART drugs in resource-limited settings (Takou et al., 2019; Tang et al., 2023). Bone involvement is characterized by the presence of osteopenia or osteoporosis as identified through dual-energy X-ray absorptiometry (DEXA) scanning (Atencio et al., 2021; El Maghraoui and Roux, 2008). Current international practice guidelines recommend that DEXA scans be conducted for PWH aged 50 years and older (Negredo et al., 2018; Lustig et al., 2016). Despite these recommendations, access to DEXA scanning remains limited in many low-middle-income countries. While some studies have explored fracture risk prediction using mathematical models or machine learning techniques, such as FRAX and Garvan (Kanis et al., 2009; Nguyen et al., 2008; Galassi et al., 2020), these approaches may not be generalizable to PWH due to variations in population characteristics and sample size limitations. Consequently, no fracture risk prediction model specifically tailored for PWH that is available for clinical use.

This study aims to develop a predictive model for bone loss in PWH using deep learning techniques, with a particular focus on those who are either undergoing long-term tenofovir disoproxil fumarate (TDF)-based ART regimens or are about to initiate such treatment. By enabling early identification of bone loss risk in PWH, the model seeks to inform optimal ART regimen selection, especially in resource-limited settings, ultimately improving the quality of life for PWH.

Over the past 2 decades, the development of more effective and better-tolerated anti-HIV therapies has significantly extended the long-term survival of most patients receiving treatment. This now closely aligns with that of the general population and may be comparable in a significant subset (Samji et al., 2013; Lewden et al., 2007). Osteoporosis is the most prevalent bone disease in the general population, with skeletal fragility fractures being its primary complication, leading to substantial medical, functional, and economic burdens (Yin and Falutz, 2016). Although fractures can occur at any age, individuals with HIV, both male and female, who are receiving treatment face a higher-than-expected fracture risk compared to the general population of the same age (Dong et al., 2014). The reported prevalence of low BMD in PWH varies widely in the literature, ranging from 13.9% to 88.3% (Paccou et al., 2018; Cazanave et al., 2008; Cervero et al., 2018; Shao et al., 2024; Meng et al., 2022). Specifically, ART-related bone loss is more pronounced in regimens that include TDF. A meta-analysis demonstrated that PWH on stable ART experienced a more significant decline in BMD when treated with TDF, with an annual decrease of 0.67% in the lumbar spine and 0.35% at the total hip after the first year (Starup-Linde et al., 2020). Our previous study also indicated that PWH exposed to TDF for 2 years experienced a 4.37% reduction in hip BMD compared to baseline (Guo et al., 2021). However, few studies on changes in BMD among long-term PWH undergoing TDF-containing ART regimens. A 5-year prospective study by Han et al. showed that after 60 months of TDF-containing ART, BMD decreased by 5.4% in the lumbar spine and 4% in the total hip, respectively (Han et al., 2020). However, few reports on the frequency of low BMD or fractures associated with TDF-containing ART. In our study, the frequency of low BMD among PWH on long-term TDF-based ART was approximately 25% (Table 1), which was relatively high. Moreover, with the extension of ART duration, the incidence of cumulative bone loss showed an upward trend (Supplementary Figure S1). This further suggests that in resource-limited settings, particularly where BMD measurement is not feasible, long-term TDF-containing ART regimens should be chosen with caution. In cases where suitable alternatives are unavailable, expanding the indications for free treatment with new INSTIs may be a consideration for future clinical experts.

Furthermore, we compared the clinical characteristics of PWH who experienced bone loss with those in a control group without bone-related adverse events. We found that the bone loss group had a lower pre-treatment body weight and a higher proportion of regimens including LPV/r, which were consistent with previous studies. A meta-analysis indicated that low body weight may largely account for the high prevalence of low BMD reported in PWH (Bolland et al., 2007). Katherine et al. also suggested that low body weight was more strongly negatively associated with BMD in HIV-positive persons (Kooij et al., 2015). A large number of studies have consistently concluded that TDF in combination with LPV/r results in a greater decrease in BMD and a higher risk of fracture than PWH with the unenhanced TDF group (Guo et al., 2021; Hill et al., 2018). Recent studies in China had also found that ART regimens containing LPV/r can lead to a decrease in BMD in the short term (Shao et al., 2024; Guan et al., 2021). This may be related to the fact that LPV/r increases the AUC of TDF, thereby potentially exacerbating the bone adverse events of TDF (Vitoria et al., 2016). Meanwhile, bone adverse events are also associated with gender and age, with women more likely to experience osteopenia or osteoporosis than men (Ahmed et al., 2023). Older age was also associated with a greater risk of bone loss, especially in postmenopausal women (Jaqua et al., 2022; Sharma et al., 2022). Our study also reached a consistent conclusion that although the median age of PLWH was less than 50 years (Table 2). Thus, it also potentially suggests that monitoring of BMD is equally important in PWH younger than 50 years of age, informing future adjustments to actual guidelines (Atencio et al., 2021).

Our study did not observe any differences in immune markers between individuals with and without bone loss including CD4⁺ T cell count, HIV VL and advanced disease status, although some studies had reported that the immune status of PWH may also be a potential factor influencing BMD decline after ART. A randomized trial from the START bone mineral density substudy showed that immediate ART (CD4 >500 cells/μL) resulted in greater BMD declines than deferred ART (CD4 <350 cells/μL) at the hip (−2.5% versus −1.0%, p < 0.001) and spine (−1.9% versus −0.4%, p < 0.001), further finding revealed lower CD4 count was a significant predictor of greater BMD loss at both the spine and the hip in the deferred ART group (Hoy et al., 2017). A lower CD4 count and higher HIV viral load have been associated with lower bone mass in cross-sectional studies, suggesting a role for HIV infection or the immunological response to HIV in bone loss (Kooij et al., 2015; Brown et al., 2009). However, the results of the SMART study were different. These data showed that hip BMD declined by 0.8% per year for up to 4 years of follow-up in participants who were ART-experienced at study entry, and continued using ART, but TNF-α and IL-6 increased in the intermittent ART group, and decreased in the continuous ART group. There was no evidence for an association of TNF-α, IL-6, or CD4+T cell count with any of the other BMD outcomes (Hoy et al., 2013). The variables included in these different studies vary, and as a result, the conclusions drawn may not be consistent. The more influencing factors are included, the more the conclusion will reflect the factors most closely related to BMD, just like the SMART study and this study. Immunological factors may indirectly affect bone density by influencing bone resorption and bone formation (Hoy et al., 2013).

Intriguingly, in our study, there was also a significant difference in the duration of ART with TDF between the group that experienced bone loss and the control group, where the duration was rather longer (Table 2). This seemed to contradict the finding in our study that the cumulative incidence of bone adverse events was higher as the duration was longer. One possible explanation is that the occurrence of bony adverse events with long-term use of TDF regimens occurs only in specific PWH, although most studies consider BMI or body weight, age, gender, etc., as risk factors for low BMD. In actual clinical practice, some postmenopausal HIV-infected women had not experienced bone loss with TDF-containing ART regimens. It was especially important to early identification of people who may experience bone loss. Our study compared the ROC of different machine models and found that the AUC curve was still the largest for LASSO regression. Further, three variables, age, sex and LPV/r, were selected based on LASSO regression brushing to be most associated with clinical resolution of bone loss. Finally, our results were visualized thought a nomogram. Lasso regression was superior to univariate analysis and can resolve multicollinearity between variables. In this study, we developed a nomogram based on a Lasso-logistic regression model, which can help clinical practitioners to predict PWH who develop low BMD with TDF, especially in low-middle-income countries where BMD monitoring was not available, to further optimize the treatment regimen, such as switch TDF to tenofovir alafenamide (TAF). Many clinical trials reported that switching from a TDF-containing regimen to one containing TAF, has been associated with BMD increases as measured by DEXA (Maggiolo et al., 2019; Orkin et al., 2017) and to inform the future expansion of the indications for new first-line INSTIs.

Our study also had several limitations that should not be ignored. Firstly, it was a retrospective investigation with inevitable retrospective bias. Meanwhile, this was a single-centre study with a relatively small sample size, limiting the generalizability of the findings. Furthermore, older PWH TDF-containing ART regimes were not included in our model, meaning that changes in bone mineral density (BMD) could not be predicted for PWH aged 60 years or older. However, in older patients, experienced clinicians considering bone adverse events would also try to avoid TDF-containing ART regimens. Besides, PWH (People with HIV) tend to be younger in this study, with 25% of them under the age of 50 showing bone loss. This may be related to unhealthy lifestyle habits such as smoking, drinking, lack of exercise, and high sugar and salt intake. However, we lack this data and cannot further include these factors for a comprehensive evaluation. An even more flawed point was that this study model did not include PWH who were not containing TDF ART as a control or healthy control, and it did not identify specific subgroups of PWH, which further limits the generalizability of the findings. And the AUC value is relatively low, it may reflect the complexity and heterogeneity of the data itself. To strengthen our conclusions, We will continue to optimize the model and explore additional factors that may affect its performance in future studies.

Our study found that long-term ART (more than 5 years) with TDF-containing regimens is associated with a relatively high frequency of bone loss, including osteopenia and osteoporosis. We also developed a predictive model for bone loss using a LASSO-penalized logistic regression approach, which we visualized with a nomogram. This model can assist clinicians in optimizing ART regimens, such as discontinuing TDF and LPV/r in favor of TDF/LPV/r-free alternatives, or using TAF or EFV 400 mg, especially in resource-limited areas where BMD monitoring may not be available.