Machine learning identifies differential predictors of early bone outcomes after TSFE in a retrospective cohort

Abstract Background Predicting which patients will achieve optimal bone augmentation after transalveolar sinus floor elevation (TSFE) remains difficult due to the multiplicity of interacting factors. To explore the differential clinical and imaging factors influencing bone augmentation outcomes after TSFE and to establish a machine learning‑based multidimensional analysis framework. Methods This retrospective study included 100 patients (100 procedures) who underwent TSFE with simultaneous implant placement. We collected demographic, radiographic, and surgical parameters. The primary early (6-month) outcomes were endosinus bone gain (ESBG 2 ), apical bone level (ABL 2 ), osteogenesis bone density (OBD), and marginal bone loss (MBL). After correlation analysis, we used a random forest algorithm to dichotomize each outcome at candidate thresholds, build predictive models, and assess variable importance via permutation and Gini importance. Results Model performance revealed a clear dichotomy. Bone quantity was reliably predicted: the ESBG 2 model achieved excellent discrimination (AUC = 0.88), strongly correlated with intraoperative elevation height ( r  = 0.87, P  < 0.001). The ABL 2 model also performed well (AUC = 0.76), with graft material use as a key predictor. Conversely, bone quality and stability proved difficult to predict. The OBD model showed only moderate performance (AUC = 0.63), and the MBL model performed poorly (AUC 0.44–0.53). Only diabetes showed a weak association with MBL ( r  = 0.22, P  = 0.03). Conclusion Our analysis highlights a potential dichotomy in sinus augmentation outcomes: while bone quantity is predictable from surgical parameters, bone quality and marginal stability remain elusive within the early retrospective framework. This underscores the need for distinct clinical assessment strategies and external validation.