Explainable AI for Equitable Nurse Scheduling: Pragmatic Pre-Post Implementation Study.

Background: Inequitable and time-consuming shift scheduling contributes to nurse burnout, dissatisfaction, and turnover. In Taiwan, annual nurse turnover reaches 11.6%, with rigid 3-shift systems and unfair workload distribution frequently cited as key drivers. Although artificial intelligence (AI) scheduling tools exist, most lack transparency and do not formally address algorithmic bias, limiting clinical adoption. Objective: This study aimed to design, deploy, and evaluate a transparent, fairness-audited, explainable AI-enabled nurse scheduling decision support system (XAI-NSDSS) to reduce administrative burden, eliminate experience-based algorithmic bias, and enhance staff acceptance in a real-world hospital setting. Methods: A pragmatic before-after implementation study was conducted at a 671-bed teaching hospital in Taiwan (January-December 2023), involving 8 departments and 156 nurses (42 novice, 78 midlevel, and 36 experienced). A 6-month manual scheduling baseline (January-June 2023) was compared with a 6-month AI-assisted period (July-December 2023). The XAI-NSDSS integrates a random forest workload prediction model (R²=0.887), Shapley Additive Explanations-based explainability, a hybrid integer programming and binary differential evolution (IP+ BDE) optimizer, and a multidimensional fairness monitoring dashboard. A formal weight sensitivity analysis (WSA) was conducted across 7 prespecified weight configurations using full-factorial repeated-measures ANOVA to assess outcome robustness. Primary outcomes were scheduling time, error rate, and user satisfaction. Statistical analyses used linear mixed effects models (LMMs) and generalized estimating equations (GEE) with department as a random effect. Results: Monthly scheduling time decreased by 81.2% (mean 32.0, SD 8.0-mean 6.0, SD 2.0) hours; P<.001; Cohen d=4.33) and error rate decreased by 73.8% (mean 18.3, SD 4.3%-mean 4.8, SD 1.2%; P<.001; Cohen d=4.12). Nurse satisfaction improved from a mean of 3.2 (SD 0.8) to a mean of 4.4 (SD 0.6; P<.001), with 148 out of 156 nurses (94.9%) adopting the system by Month 3. Preexisting experience-based bias was fully eliminated: workload coefficient of variation (CV) decreased 50% (0.18-0.09; P<.001), disparate impact ratios normalized from 1.35-1.56 to 1.01-1.04, and preference satisfaction equity was achieved across experience tiers (ANOVA P=.38). Among 156 nurses, 82 (52.6%) regularly engaged with Shapley Additive Explanations; this engagement was positively associated with satisfaction (Pearson r=0.456; P<.001). The WSA across 7 configurations confirmed that the consensus-derived default weights achieved the highest composite quality score (mean 82.1, SD 3.2) and that disparate impact ratios remained within the 0.80-1.25 fairness threshold across all configurations (P=.12), demonstrating structural robustness of the fairness-auditing module. Conclusions: This study presents the first longitudinally validated explainable AI implementation framework for nurse scheduling with formal algorithmic fairness auditing and WSA. The XAI-NSDSS framework is replicable, scalable, and provides a practical blueprint for responsible AI adoption in health care workforce governance, with fairness guarantees that are robust to institutional customization of optimization priorities.