This study examined the implementation outcomes of personalized exercise programs in university physical education and developed an evaluation model for identifying effective individual responses. A 16-week quasi-experimental design was used with 586 valid undergraduate participants from compulsory physical education classes. The intervention group received individualized exercise programs generated from baseline fitness tests, exercise preference, self-efficacy, injury-risk screening and weekly workload records, while the control group followed conventional class-based instruction. The evaluation framework combined standardized fitness improvement, moderate-to-vigorous physical activity, adherence, safety and preference matching. LightGBM was used to predict whether students reached the predefined response threshold, and SHAP was applied to explain variable contribution. After 16 weeks, the intervention group increased its physical fitness composite score by 8.49 points, compared with 3.57 points in the control group. Weekly MVPA increased by 57.4 min in the intervention group and by 24.6 min in the control group. The intervention group also showed higher adherence, lower overload rate and fewer discomfort reports. The response prediction model achieved an AUC of 0.842, an F1 score of 0.786 and an RMSE of 3.18 for score-gain prediction. Ablation analysis indicated that removing workload records caused the largest performance decline. The results suggest that personalized exercise programs can improve university physical education when baseline diagnosis, process workload control and interpretable evaluation are integrated within the same teaching cycle.
