Explainable deep learning for healthcare workforce attrition: a methodological study on the Watson healthcare synthetic benchmark.

Background: Retaining qualified nurses and allied-health staff has become a central occupational-health concern for modern hospital systems, yet existing attrition-prediction models face a persistent trade-off between predictive accuracy and auditability. The quantitative contribution of modifiable occupational factors-working hours, work environment, compensation, and commuting burden-to individual departure decisions remains insufficiently characterized. Methods: We developed an end-to-end interpretable deep-learning framework and evaluated it on the publicly available Results: The proposed network achieved an area under the receiver-operating-characteristic curve of 0.934 ± 0.026 and an area under the precision-recall curve of 0.761 ± 0.091, matching or exceeding all three baselines while preserving minority-class sensitivity. Dual-level SHAP analysis consistently identified overtime status, environment satisfaction, marital status, job involvement, tenure in current role, age, and commuting distance as the dominant drivers of the predicted attrition label on this benchmark. It revealed interpretable non-linear thresholds together with auditable, case-by-case rationales. Conclusion: On this synthetic benchmark, a compact interpretable deep network matches strong tabular baselines while providing axiomatic, dual-level SHAP explanations that translate into three candidate retention levers-overtime reduction, working-environment improvement, and commuting support. Because the records are synthetic, these levers should be regarded as data-driven hypotheses rather than validated interventions, and prospective evaluation on real, de-identified hospital workforce data remains a necessary next step.