<h3>Purpose/Objective(s)</h3> Significant controversy remains over sufficient dose to minimize local failure (LF) for SBRT treatment of NSCLC treated to 50 Gy in 5 fractions (BED = 100). We hypothesized that dosimetric predictors can be determined by developing survival machine learning (SML) models, identifying highest impact features and determining cutoffs for guiding treatment planning. <h3>Materials/Methods</h3> Five hundred fifty-two NSCLC SBRT patients treated from 2009-2017 with prescription of 50 Gy in 5 fractions were retrospectively identified. Median follow-up was 23 months. Five clinical features (age, gender, tumor location, clinical max diameter, and sphericity) and ten dosimetric features (min, max, mean, standard deviation, D5, D20, D40, D60, D80, and D95) for both the IGTV and ‘PTV-IGTV' were extracted. Five SML models, including elastic net, survival random forest, and gradient boost, were investigated. Additionally, an Auto-Encoder Survival Deep Network (AE-SDN) was developed combining the reconstruction loss of AE with Cox regression module for modeling time to event, LF in this case. The quality of predictions was assessed using Harrell's concordance index (C-index). For explainability, SHAP analysis was used to study features' influence on the observed LF events. Cutoff points were derived using SHAP analyses, and log-rank test and Cox regression were used to assess group differences based on these cutoffs. <h3>Results</h3> Five hundred patients were included in this analysis. Table 1 summarizes this cohort. Thirty-one patients experienced LF at a median time of 21 months. AE-SDN had the best performance (C-index, 0.79; 95% CI, 0.78-0.81). Stratifying the test sets using AE-SDN's output, the low-risk patients had significantly lower LF (HR, 0.23; 95% CI, 0.1-0.57; p<0.005). Using SHAP analysis and log-rank testing on the AE-SDN's outputs, significant identifiers of reduced LF risk were female gender (HR, 0.43; 95% CI, 0.2-0.89; p=0.019), clinical max diameter < 2.3 cm (HR, 0.3; 95% CI, 0.15-0.6; p<0.005), sphericity > 0.5 (HR, 0.23; 95% CI, 0.11-0.48; p<0.005), IGTV min dose >50.92 (HR, 0.27; 95% CI, 0.13-0.56; p<0.005), and PTV-IGTV mean dose >52.32 (HR, 0.38; 95% CI, 0.19-0.77; p=0.005). <h3>Conclusion</h3> AE-SDN outperformed other SML models in LF prediction. SHAP analysis on AE-SDN indicated that, in our cohort, patients receiving IGTV min and PTV-IGTV mean higher than 50.92 and 52.32 Gy, respectively, had significantly lower LF risks. These dosimetric factors require validation. The prognostic power of gender, clinical max diameter, and sphericity were also presented.