Bone loss in young adults with HIV following antiretroviral therapy containing tenofovir disoproxil fumarate regimen using machine learning.

OBJECTIVE

Bone mineral density (BMD) monitoring, primarily relying on dual-energy X-ray absorptiometry (DEXA), remains inaccessible in resource-limited regions, making it difficult to promptly address bone loss in people with HIV (PWH) on long-term ART-containing TDF regimens and assess the prevalence of bone loss. Our objective is to identify the frequency of PWH experiencing bone loss after long-term ART with a TDF regimen and to develop a predictive model of HIV-infected high-risk populations containing TDF long-time ART, for providing more appropriate ART regimens for PWH in clinical practice, particularly in resource-limited settings.

METHODS

Our study retrospectively screened PWH under long-term follow-up at Peking Union Medical College Hospital (PUMCH) from January 2000 to August 2024. These individuals were either treatment-naive or treatment-experienced and had been on containing TDF ART regimen for over 5 years. BMD was assessed using DEXA every 1-2 years in this center. We selected predictive factors utilizing machine learning methods, including Random Forest, XGBoost, LASSO regression, and logistic regression. The results were visualized using a nomogram.

RESULTS

Our study enrolled a total of 232 PWH who have contained TDF ART regimens for more than 5 years. Twenty-five percent (58/232) of the patients experienced bone loss, primarily including osteopenia and osteoporosis. Further results showed that the LASSO regression model was the most suitable for the current dataset, based on a comparison of LASSO regression, Random Forest, XGBoost, and logistic regression models including age, gender, LPV/r, baseline CD4+ T count, baseline VL, baseline body weight, treatment-naïve TDF, ART duration, percentage of CD38+CD8+T, percentage of HLA-DR+CD8 T, and CD4+/CD8+ ratio, with AUC values of 0.615, 0.507, 0.593, and 0.588, respectively. We identified age, gender, and LPV/r as the most relevant predictive factors associated with bone loss based on LASSO regression. Then the results were visualized and plotted in a nomogram.

CONCLUSION

Our study quantified the frequency and established a nomogram based on the LASSO regression model to predict bone loss in PWH on long-term containing TDF ART. The nomogram guides identifying individuals at high risk of bone loss due to prolonged TDF exposure. Clinicians can leverage the predicted risk to design personalized ART regimens at the initiation of therapy or to switch from TDF-containing to TDF-free regimens during treatment. This approach aims to reduce the incidence of bone loss, particularly in resource-limited settings.