BACKGROUND:
Non-routine discharge following single-level cervical disc arthroplasty (CDA) is associated with increased morbidity and healthcare burden. Identifying key predictors can improve perioperative planning and patient outcomes. The aim of this study is to predict non-routine discharge following single-level CDA and identify key discharge predictors.

METHODS:
A cohort of 14,315 single-level CDA patients was identified from the National Inpatient Sample (NIS) database from 2016 to 2020. Eight machine learning models predicted non-routine discharge based on characteristics such as age, length of stay (LOS), comorbidities, and surgical factors. Chi-square and t-tests compared categorical and continuous outcomes, with significance set at the 0.05 level.

RESULTS:
Significant predictors of non-routine discharge included psychoses (1.5 % vs. 0.4 %, P = 0.019), LOS (3.0 ± 0.2 days vs. 1.3 ± 0.02 days, P <  0.001), and age (53.0 ± 0.9 vs. 47.4 ± 0.2 years, P <  0.001). Race was also a factor, with higher non-routine discharge rates among Black (14.4 % vs. 5.8 %) and Hispanic (9.8 % vs. 7.2 %) patients compared to White patients (68.0 % vs. 81.5 %, P <  0.001). AdaBoost achieved the highest accuracy (93.7 %), sensitivity (0.17), specificity (0.99), and AUC (72.5 %). Gradient Boosting (93.0 % accuracy, specificity 0.99, sensitivity 0.08, AUC 72.6 %) also performed well. Other models, including Random Forest (91.1 % accuracy) and Naive Bayes (92.7 %), showed high accuracy but lower sensitivities (RF: 0.11, NB: 0.19).

CONCLUSION:
Machine learning models, especially AdaBoost and Gradient Boosting, effectively predicted non-routine discharge after single-level CDA. Key predictors included psychoses, extended hospital stays, and older age. Recognizing these factors may support targeted interventions to enhance patient outcomes.