Wearable sensors offer a promising platform for non-invasive glucose monitoring by indirectly predicting glucose levels from physiological signals. However, machine learning models trained on such data often suffer degraded performance when applied to new individuals due to distribution shifts in physiological patterns. This study investigates how the inter-subject distribution shift impacts the performance of glucose prediction models trained on wearable data. We utilize the BIGIDEAs dataset, which includes simultaneous recordings of glucose levels and multimodal physiological signals. Personalized XGBoost regression models were trained on data from 10 subjects and evaluated on 5 held-out subjects to assess cross-subject generalization. Distribution shifts in glucose profiles between training and test subjects were quantified using the Anderson-Darling (AD) statistic. Results show that models trained on one individual performed poorly when tested on others. Repeated measures correlation analysis revealed significant positive correlations between the AD statistic and model performance metrics, including RMSE, NRMSE, and MARD. Our findings highlight the challenge of inter-individual generalization and the need for distribution-aware models. We propose personalized calibration and subject phenotyping as future directions to enhance model generalizability.