The global nursing workforce crisis is one of the most pressing challenges in modern health care. The World Health Organization (WHO) projects a shortage of 5.9 million nurses by 2030, with high-income countries experiencing turnover exceeding 15% annually [1]. In Taiwan, the annual nursing turnover rate reached a 10-year high of 11.6%, with scheduling-related dissatisfaction identified as a leading contributor to burnout and early career exit [2,3]. Inefficient scheduling creates a vicious cycle; unfair shift distribution increases fatigue, which accelerates attrition, which further strains remaining staff.

Nurse scheduling is a complex combinatorial optimization problem requiring simultaneous consideration of regulatory constraints (labor laws and union agreements), operational requirements (shift coverage and skill mix), and individual staff preferences (time-off requests and work-life balance) [4]. Traditional manual scheduling is time-intensive, error-prone, and chronically fails to achieve fair workload distribution, particularly disadvantaging novice nurses who lack institutional leverage to negotiate favorable assignments [5].

Artificial intelligence (AI) and operations research approaches have shown promise in nurse scheduling. Recent implementations include mixed integer programming heuristics [6], genetic algorithms [7], and machine learning–based prediction models [8]. However, a systematic review of the literature reveals four persistent, unresolved gaps:

This study directly addresses all 4 gaps through the design, deployment, and 12-month evaluation (6-mo pre- and 6-mo postimplementation) of the explainable AI–enabled nurse scheduling decision support system (XAI-NSDSS) at a 671-bed tertiary referral hospital in Taiwan. The aim of this study was to demonstrate that a single integrated framework can simultaneously achieve operational efficiency, provable algorithmic equity, and clinician trust through explainability. Specific contributions are as follows:

This pragmatic before-after implementation study was conducted at Fu Jen Catholic University Hospital, a 671-bed tertiary referral hospital in New Taipei City, Taiwan, from January to December 2023. The XAI-NSDSS was deployed across 8 nursing departments—medical, surgical, intensive care, emergency, pediatrics, obstetrics, oncology, and geriatrics—serving 156 nurses (42 novice [<3 y experience], 78 midlevel [3‐10 y], and 36 experienced [>10 y]).

We used a mixed methods evaluation design with a 6-month preimplementation baseline (manual scheduling; January-June 2023) compared to a 6-month postimplementation period (AI-assisted scheduling; July-December 2023). Randomization was not feasible due to the hospital-wide nature of the intervention and risk of cross-contamination; the large effect sizes observed (Cohen d>4) and consistency across all 8 departments strengthen confidence in attribution. This study is reported in accordance with the Standards for Quality Improvement Reporting Excellence (SQUIRE 2.0) guidelines.

All registered nurses employed in the participating departments during the study period were eligible for inclusion. Exclusion criteria included (1) nurses on extended leave (>4 consecutive weeks), (2) temporary or agency staff without permanent employment contracts, and (3) nurses in administrative roles without direct patient care responsibilities. Of 162 eligible nurses, 156 (96.3%) participated in the full 12-month study period. Six nurses were excluded due to maternity leave (n=3), resignation (n=2), or transfer to nonparticipating departments (n=1).

The XAI-NSDSS uses a modular 5-layer architecture (Figure 1; detailed in Section S2.1 in ): (1) user interface layer—mobile-responsive dashboards (React.js 18.2.0; Meta Platforms, Inc); (2) AI engine layer—random forest regressors (scikit-learn 1.3.0; scikit-learn developers) for workload capacity prediction; (3) explainability module—SHAP TreeExplainer (shap v0.42.1; Scott Lundberg); (4) optimization layer—hybrid integer programming (IP; Gurobi 10.0.3; Gurobi Optimization, LLC)+binary differential evolution (BDE); and (5) data layer—PostgreSQL (15.3; PostgreSQL Global Development Group) with TimescaleDB (Tiger Data) and advanced encryption standard with a 256-bit key encrypted storage.

Algorithmic fairness was audited quarterly across 3 dimensions: (1) demographic parity, proportional undesirable shift distribution across experience tiers; (2) equal opportunity, preference satisfaction equity (one-way ANOVA); and (3) disparate impact, selection rate ratios against the 0.80‐1.20 threshold (“four-fifths rule”) [11]. Gini coefficients provided complementary inequality measurement. Automated alerts upon threshold violations were reviewed monthly by the AI Scheduling Oversight Committee (ASOC), a multidisciplinary body comprising 8 nursing managers, 2 information technology staff, 1 clinical ethicist, 1 legal advisor, and 1 elected nurse representative (S2.6 in Multimedia Appendix 1 and dashboard: Figure S2 in S1.2 in Multimedia Appendix 1).

System integrity was maintained through weekly automated model drift detection and quarterly retraining (threshold: R² drop >0.03 from baseline). SHAP explanations were scoped to individual nurse data only, preventing disclosure of colleagues’ performance metrics. All 156 nurses completed a 1-hour domain-specific training session covering system usage and SHAP interpretation prior to deployment. All data were stored in databases encrypted using the advanced encryption standard with a 256-bit key and protected by role-based access controls, in compliance with the Taiwan Personal Data Protection Act.

The study was approved under an exemption waiver by the Institutional Review Board of Fu Jen Catholic University Hospital (IRB reference: FJUH-IRB-114‐459). All participating nurses and managers provided written informed consent prior to enrollment. Participation was voluntary, and participants were informed that they could withdraw from the study at any time without penalty or any effect on their employment status, work assignments, or institutional evaluation. All data were deidentified before analysis and reported only in aggregate form. Individual-level scheduling records, interview transcripts, SHAP engagement data, and satisfaction responses were stored in encrypted, access-controlled databases, and no personally identifiable information is disclosed in the manuscript or supplementary materials. No images or other materials containing identifiable participant information are included. No financial compensation or other incentives were provided to participants.

Table 2 summarizes key operational metrics. The XAI-NSDSS reduced monthly scheduling time by 81.2% (from mean 32.0, SD 8.0 to mean 6.0, SD 2.0 h; P<.001; Cohen d=4.33; 95% CI for difference −29.1 to −22.9 h), compressing a 3‐5 workday administrative burden to under 1 day. At the Emergency Department, reduction was most pronounced (45 → 8 h; 82.2%). Schedule error rates decreased by 73.8% (mean 18.3, SD 4.3% to 4.8 ± 1.2%; P<.001; d=4.12), with reductions in rest period violations (91.7% reduction, from 8.4 to 0.7 violations per schedule), hard constraint violations (86.9%), and skill mismatches (78.4%).

System adoption reached 148 out of 156 nurses (94.9%) by Month 3, with subsequent stability through Month 6 (Figure 2A-2D). The largest pre-to-post satisfaction gains were in transparency/explainability (mean 2.6, SD 1.0-mean 4.3, SD 0.6; P<.001; d=2.08) and fairness perception (mean 2.9, SD 1.0-mean 4.2, SD 0.7; P<.001; d=1.50). Trust in AI recommendations improved from a mean of 2.8 (SD) 0.9 to mean of 4.1 (SD 0.7; P<.001; d=1.62). Of 156 nurses, 82 (52.6%) regularly engaged with SHAP explanations, with a significant positive association between engagement frequency and overall satisfaction (Pearson r=0.456; P<.001; Figure S3 in ). Full satisfaction subscale results are reported in Table S4 in Multimedia Appendix 1.

Global SHAP analysis identified cumulative fatigue score (mean |SHAP|=0.42), years of experience (0.38), and recent overtime hours (0.35) as the 3 most influential scheduling determinants (Figure 3). Survey responses confirmed that SHAP explanations helped 122 out of 156 (78.2%) nurses understand shift assignments, increased trust for 111 out of 156 (71.2%) nurses, and identified actionable strategies to improve scheduling priority for 100 out of 156 (64.1%) nurses.

The hybrid IP+BDE approach significantly outperformed either component in isolation (Table 3). BDE-only achieved only 95.2% hard constraint compliance versus IP’s guaranteed 100%. IP-only achieved only 78.3% preference satisfaction versus the hybrid’s 88.1%. Computation time averaged 12.7 minutes per monthly schedule versus 32.0 hours for manual scheduling (99.3% reduction). Emergency rescheduling resolved within 5‐8 minutes.

Preimplementation, experienced nurses received significantly higher preference satisfaction than novice nurses (76.2% vs 68.4%; gap=7.8 percentage points; P=.009). Postimplementation, this disparity was eliminated (ANOVA P=.38; all pairwise comparisons P>.30), with all 3 experience tiers achieving satisfaction rates ≥87%. Equity was achieved through a “leveling up” strategy—absolute gains >13 percentage points for all groups (P<.001)—rather than redistributing the burden from experienced to novice nurses.

Hawthorne effect analysis (Months 1‐3 vs 4‐6) confirmed sustained benefits, scheduling time remained stable (mean 6.2, SD 2.1 vs mean 5.8, SD 1.9 h; P=.34); satisfaction continued its upward trend (mean 4.2, SD 0.7-mean 4.6, SD 0.5; P=.03). Workload CV remained below the preimplementation level across all 6 months (range 0.08‐0.10), confirming sustained improvement rather than novelty effect. Seasonal adjustment revealed no significant interaction between time and holiday period (P=.41). Detailed results are provided in Table S5 in Multimedia Appendix 1.

Beyond the prespecified sensitivity analyses reported above, a formal WSA was conducted to assess whether scheduling outcomes depended on the choice of objective-function weights. WSA was conducted across all 7 prespecified configurations (Table 4). One-way repeated-measures ANOVA revealed a significant main effect of weight configuration on workload equity CV (F_{₆, ₂₈₂}=18.43; P<.001; η²=0.28), preference fulfillment rate (F_{₆, ₂₈₂}=22.17; P<.001; η²=0.32), and composite schedule quality score (F_{₆, ₂₈₂}=9.84; P<.001; η²=0.17). No significant main effect was observed for the disparate impact ratio range (F_{₆, ₂₈₂}=1.72; P=.12), indicating that the fairness-auditing module maintained equitable shift distribution regardless of weight settings.

Figure 2 presents monthly trend data across all 4 primary performance domains during the postimplementation period. The 2-month adaptation period followed by sustained stability aligns with the “implementation cliff” described in implementation science literature [18] and supports the recommendation of a 3-month intensive support period for similar deployments. Detailed monthly data are provided in Table S6 in Multimedia Appendix 1.

Semistructured interviews (n=24 nurses; n=8 managers; mean duration 38 min) yielded 5 major themes. A full participant characteristics table is provided in Table S3 in Multimedia Appendix 1.

This study demonstrates that a fairness-audited, XAI scheduling system can simultaneously reduce administrative burden by 81.2%, eliminate preexisting experience-based algorithmic bias, and achieve 94.9% (148/156 nurses) system adoption in a real-world hospital setting. Three findings warrant particular emphasis: (1) preexisting experience-based bias was formally documented and completely eliminated; (2) SHAP engagement was significantly associated with user satisfaction; and (3) sustained improvements confirmed by sensitivity analyses ruled out novelty effects.

The most significant finding is the systematic identification and complete elimination of preexisting experience-based bias. Preimplementation, 69 out of 156 novice nurses were over-represented in all 4 undesirable shift categories by 26%‐46% (disparate impact ratios 1.35‐1.56), exceeding the established four-fifths rule threshold for systemic discrimination [11]. This trend aligns with documented informal seniority hierarchies in nursing workplaces [19] but had not previously been quantified with formal fairness metrics in the AI scheduling literature.

The complete normalization of all disparate impact ratios postimplementation (1.01‐1.04) demonstrates that embedding W_imbalance and F_total into the objective function can mitigate institutionalized bias. The W_imbalance term minimizes variance in shift allocation across all nurses, and post hoc disparate impact analysis confirmed this proxy effectively eliminated experience-based disparities, validating the chosen surrogate metric. The Gini coefficient improvement (0.24-0.13) further confirms equity advancements. Future implementations should treat preexisting bias characterization as a mandatory predeployment audit step.

A central hypothesis was that SHAP-based explainability would serve as an active trust-building mechanism. Our data show a significant positive association between SHAP engagement and satisfaction (r=0.456; P<.001). While causality cannot be established from this observational correlation—satisfied nurses may be more inclined to engage with explanations—the qualitative data consistently suggest that explanations functionally mitigate organizational conflict when used, consistent with a bidirectional reinforcing relationship. This extends work by Liu et al [20] on “integrability” as a determinant of AI acceptance, demonstrating that explanations must be actively used—not merely available—to generate trust benefits.

Critically, while 82 out of 156 nurses (52.6%) regularly viewed SHAP explanations, this study measured engagement frequency rather than comprehension accuracy. A formal assessment of whether nurses correctly interpreted feature importance rankings was not conducted, leaving a “comprehension gap” unquantified. Future research should use validated XAI literacy instruments to assess whether engagement translates to understanding.

The 52.6% rate of voluntary SHAP engagement surpasses engagement rates reported for passive transparency features in health care AI [21]. The 1-hour domain-specific training investment appears to be an effective and scalable strategy for facilitating XAI adoption.

Our 73.8% error rate reduction exceeds the 40%‐60% range reported in comparable scheduling optimization studies [22,23], attributable primarily to the human-in-the-loop verification enabled by SHAP transparency. The hybrid IP+BDE approach outperformed both single-method alternatives, consistent with the demonstration by Ben Said et al [24] that hybrid constraint-learning pipelines outperform single-method approaches in health care scheduling. Regarding the role of XAI versus IP optimization, XAI in this system is not merely a technical overlay but a functional organizational mechanism. Without SHAP explanations, the system would generate algorithmically optimal schedules that clinicians cannot audit, verify, or trust, making sustained adoption unlikely regardless of algorithmic performance. The 94.9% adoption rate at Month 3 versus typical rates for opaque AI systems substantiates this claim [9,10].

The WSA addresses a fundamental challenge in multiobjective optimization for health care applications, the selection of objective weights is inherently value-laden and context-dependent; yet, the robustness of outcomes to weight perturbations has rarely been formally evaluated in prior nurse scheduling studies. Our findings contribute 3 key insights to this literature.

First, the consensus-derived FP-Default configuration (w₁=0.30; w₂=0.25) achieved the highest composite schedule quality score and was the only configuration meeting prespecified noninferiority criteria on both workload equity and preference fulfillment simultaneously. This validates the stakeholder co-design process as an effective method for identifying near-optimal weight configurations in institutional contexts, consistent with findings from participatory AI scheduling design studies [10,19].

Second, the analysis revealed an inherent trade-off between workload equity and preference satisfaction that cannot be eliminated through algorithmic design alone. Increasing w₁ beyond 0.30 (FP-High) improved equity marginally but reduced preference fulfillment by 5.3 percentage points—a clinically meaningful reduction given that preference accommodation is a primary driver of nurse satisfaction and retention [5,22]. Conversely, maximizing preference satisfaction (PP-High) degraded workload equity, potentially reintroducing the experience-based disparities that the system was designed to eliminate. This trade-off surface highlights the irreducible nature of the fairness–preference tension and underscores the importance of transparent weight disclosure to affected stakeholders.

Third, and most critically for regulatory compliance, the disparate impact ratios remained within the 0.80‐1.25 fairness threshold across all 7 weight configurations (P=.12 for between-configuration differences). This structural invariance demonstrates that the fairness-auditing module provides a hard constraint on algorithmic bias that operates independently of optimization priorities, a design principle with direct implications for responsible AI deployment in health care workforce governance [11,25]. Institutions adopting the XAI-NSDSS framework can therefore customize weight configurations to reflect local priorities without compromising the fairness guarantees central to equitable workforce governance.

The differential engagement patterns across experience levels suggest that explainability features are particularly valuable for less-experienced staff who may have less institutional power to challenge unfair scheduling decisions. This finding aligns with qualitative research by Gerlach et al [19] identifying transparency as a critical adoption factor for junior nurses, and extends it by quantifying the relationship between explanation engagement and satisfaction outcomes.

The initial satisfaction dip (Month 1) followed by sustained improvement (Months 3‐6) is consistent with the “implementation cliff” [18]. Our data suggest a 2‐3 month stabilization window as the critical support period. The ASOC governance structure—with monthly fairness dashboard reviews and multidisciplinary oversight—addresses the requirements for organizational accountability in AI deployment reported by Saeed et al [25], distinguishing this implementation from most published systems that lack ongoing governance postlaunch.

Single-site implementation in Taiwan limits generalizability to other health care systems, labor regulations, and cultural contexts; multisite validation is needed before broad adoption recommendations can be made. The 6-month postimplementation period does not address long-term system drift or multiyear organizational embedding. The quasiexperimental before-after design inherently limits causal inference: despite large effect sizes (Cohen d>3.7 across all primary outcomes), secular trends, regression to the mean, or co-occurring organizational changes cannot be fully excluded. SHAP comprehension accuracy was not formally assessed—only viewing frequency—leaving a “comprehension gap” unquantified. Patient care outcomes (eg, medication error rates and fall incidence) were not measured; connecting scheduling equity to downstream clinical outcomes remains an important research direction. The SHAP-satisfaction correlation is observational and directional causality cannot be established without experimental manipulation.

The WSA was conducted within a single institutional context; the consensus-derived FP-Default weight configuration (w₁=0.30, w₂=0.25) reflects this institution’s scheduling culture, labor regulations, and workforce demographics. While the structural robustness of fairness metrics across all 7 configurations is likely to generalize, the absolute weight values may require site-specific calibration in institutions with different operational priorities. Future multisite replications should include systematic WSA to determine whether FP-Default represents a transferable standard or an institution-specific optimum.

Priority directions include (1) multisite validation across diverse health care systems, (2) randomized trials comparing explainable versus nonexplainable scheduling systems to establish the causal impact of XAI on trust and adoption, (3) integration with real-time patient acuity data, and (4) federated learning for privacy-preserving multi-institutional model development [26].

Regarding optimization robustness, adaptive weight optimization methods—such as Bayesian preference elicitation or reinforcement learning from stakeholder feedback—could automate the weight configuration process and reduce reliance on expert consensus, potentially improving outcomes in institutions lacking access to multidisciplinary co-design teams. The WSA framework developed in this study should be replicated in multisite contexts to determine whether the FP-Default weight configuration generalizes across institutions with different staffing cultures and regulatory environments or whether site-specific calibration is required.

This study presents a prospectively validated, fairness-audited XAI implementation framework for nurse scheduling delivering measurable operational improvements while maintaining interpretability and human oversight. Three key findings are noteworthy:

The XAI-NSDSS framework demonstrates that algorithmic decision support can simultaneously achieve efficiency, equity, and transparency. The framework is replicable, scalable, and aligned with emerging standards for responsible AI governance in clinical settings.