Stacked Machine Learning Models Exploring Resin Matrix Ceramics Aesthetic Alteration

To investigate the long-term optical properties of CAD/CAM resin matrix ceramics (RMCs) by a thermocycler and further predict aesthetic alteration using machine learning (ML). Three RMCs as experimental groups: Hyramic (Hy), Lava Ultimate (LU), and Vita Enamic (VE), and one lithium disilicate ceramic as the control group: UP.CAD (UP) in two opacities (high and low translucency, HT and LT), two thicknesses (0.5 and 1.0mm), and cementation or not were evaluated (n=5). L*, a*, and b* were measured at baseline and every 10000 thermocycles up to 40000. L*, whiteness index (WI), and translucency parameter (TP) were calculated and analyzed with t-tests and ANOVA. The dual-step stacked ML models were developed to predict aesthetic alteration. Opacities, thicknesses, material types, and thermocycling affected optical properties significantly (P<0.05). The highest L* (66.77±0.18), WI (31.35±0.39), and TP (43.68±1.57) were found in 1.0mm HT VE, 1.0mm LT LU, and 0.5mm HT VE, respectively. ML predicted aesthetic deterioration timings were 18584, 7828, 36577, and 27828 thermocycles for UP, Hy, LU, and VE, respectively, with R2/MSE (L*: 0.9633/0.2375, WI: 0.9724/0.4017, and TP: 0.9801/1.2205). Compared to UP, RMCs showed greater alterations in L* and WI, lower TP, and ML predicted aesthetic deterioration between 7828 and 36577 thermocycles. Notably, LU exhibited superior WI and extended aesthetic longevity.