Risk prediction models for postmenopausal osteoporosis: a systematic review and meta-analysis study.

backgroundPostmenopausal osteoporosis usually happens 5 ~ 10 years after menopause. Low awareness, low detection rates and high morbidity have prevented the possibility of early or preventive interventions, thus increasing the social and economic burden on families and societies. A reliable prediction model for postmenopausal osteoporosis has the potential to guide the prevention, but regarding the early prediction of postmenopausal osteoporosis without fracture, this field has not been sufficiently studied. Although many scholars have developed several prediction models to estimate the risk of postmenopausal osteoporosis without fractures, the evidence about the model quality and clinical applicability is scarce.

methodNine databases (Medline, Embase, Web of science, CINAHL, The Cochrane Library, CNKI, SinoMed, Wanfang, VIP data) were systematically searched from 1 January 2014 to 1 May 2024. Two researchers independently extracted data using the CHARMS checklist and assessed bias using the PROBAST tool. The primary outcomes of interest were related to the model’s discriminative ability (assessed by pooled AUC values) and calibration performance (evaluated using calibration curves or the calibration intercept and slope). We performed meta-regression and sensitivity analyses to explore the influence of important factors, such as data sources, machine learning methods, and types of predictor variables, on the aforementioned. results. Additionally, subgroup analyses were conducted based on data sources, machine learning. methods, and types of predictor variables. The study was registered in the PROSPERO database (registration number CRD42024542498).

resultsA total of 8,549 records were initially identified, and 7 studies (comprising 19 models) were ultimately included. All models were developed based on Asian population data. The risk of bias assessment showed: 1 study had a low risk, 1 study had an unclear risk, and 5 studies had a high risk. The sample sizes ranged from 319 to 4,417 participants. The reported AUC of the models ranged from 0.639 to 0.921; however, the vast majority of studies lacked reports on calibration performance. The pooled C-statistic (AUC) was 0.78 (95%CI: 0.73–0.83). Sensitivity analysis yielded robust results (AUC=0.77). Subgroup analysis indicated that models combining demographic and laboratory data demonstrated the best performance (AUC=0.92). Significant publication bias and substantial heterogeneity (I² = 98%) were observed among the studies.

conclusionCurrent machine learning-based prediction models for postmenopausal osteoporosis without fractures, as presented in the included studies, demonstrate good discriminative ability but are generally characterized by a high risk of bias, a notable lack of calibration performance evaluation, and insufficient validation of clinical utility. Furthermore, existing models are developed entirely on Asian population data, which limits their generalizability to other populations. Future research should focus on strictly adhering to prediction model research guidelines (such as PROBAST), enhancing the reporting of model calibration and clinical utility, and assessing model generalizability through external validation in multi-center studies encompassing diverse ethnicities and regions.