Decision-Architecture Dependence and Patient-Gender-Conditioned Allocation in Large-Language-Model-Assisted Nursing Allocation: A Multi-Model Counterfactual Audit.
5 affiliations
- 1Guangxi Hospital Division of the First Affiliated Hospital, Sun Yat-sen University, Nanning 530022, China.
- 2Faculty of Humanities, Management and Science, Universiti Putra Malaysia Sarawak, Bintulu 97008, Malaysia.
- 3Faculty of Business Administration, University of Macau, Avenida da Universidade, Taipa, Macau 999078, China.
- 4The First Affiliated Hospital, Sun Yat-sen University, Guangzhou 510000, China.
- 5The Peoples Hospital of Guangxi Zhuang Autonomous Region, Nanning 530022, China.
Abstract
Background/Objectives: Large language models (LLMs) may support healthcare workforce allocation, but their decisions can be sensitive to demographic cues and decision architecture. This study examined gender-task congruity, Candidate-A dependence under the tested decision architecture, and patient-gender-conditioned nursing allocation. Methods: Three provider-specific LLM configurations were audited across 24 nursing tasks. Male and female nurses were professionally equivalent and counterbalanced across Candidate A/B; patient gender was varied in six intimate-care tasks. Sixty conditions were repeated three times per model, yielding 540 valid outputs. Results: Male nurses were selected in 50.0% of Technical/Procedural and Relational/Educational allocations. Candidate A was selected in all 288 observations, although 285 (99.0%) assigned identical suitability scores. Because A/B assignment was bundled with labeling, presentation order, evaluation order, and response-template anchoring, this pattern indicates Candidate-A dependence under the tested decision architecture rather than an isolated position effect. Changing patient gender from female to male produced 57 female-to-male nurse switches and no reverse switches (p < 0.001). Conclusions: The audited configurations showed no generalized gender-task pattern but demonstrated decision-architecture dependence and patient-gender-conditioned allocation. Fairness audits should assess equivalent-candidate handling, patient preference, and counterfactual stability.
Related
- Intervention-Conditioned Counterfactual Human Digital Twin Generator
- Counterfactual Constraint-Conditioned On-Policy Distillation for Multi-Constraint Instruction Following
- Neural-DISCO: Source-Conditioned Counterfactual Editing of Neural Population Activity
- Explainable multimodal learning analytics for model-based counterfactual decision support using literature-calibrated synthetic data
- Distributional Counterfactual Dependence in the Special Sciences
BibTeX
@article{li2026decisionarchitecture,
title = {{Decision-Architecture Dependence and Patient-Gender-Conditioned Allocation in Large-Language-Model-Assisted Nursing Allocation: A Multi-Model Counterfactual Audit.}},
author = {Zhiying Li and Wenjie Zhu and Luodi Jiang and Xiuzhen Wei and Jie Zhou and Wei Yang and Baoling Zhong and Qiongyan Hu},
journal = {Healthcare (Basel, Switzerland)},
year = {2026},
month = oct,
doi = {10.3390/healthcare14203549},
url = {https://europepmc.org/article/MED/42940596},
note = {CC BY},
}
RIS
TY - JOUR
TI - Decision-Architecture Dependence and Patient-Gender-Conditioned Allocation in Large-Language-Model-Assisted Nursing Allocation: A Multi-Model Counterfactual Audit.
AU - Zhiying Li
AU - Wenjie Zhu
AU - Luodi Jiang
AU - Xiuzhen Wei
AU - Jie Zhou
AU - Wei Yang
AU - Baoling Zhong
AU - Qiongyan Hu
PY - 2026
DA - 2026/10/21
JO - Healthcare (Basel, Switzerland)
DO - 10.3390/healthcare14203549
UR - https://europepmc.org/article/MED/42940596
AB - Background/Objectives: Large language models (LLMs) may support healthcare workforce allocation, but their decisions can be sensitive to demographic cues and decision architecture. This study examined gender-task congruity, Candidate-A dependence under the tested decision architecture, and patient-gender-conditioned nursing allocation. Methods: Three provider-specific LLM configurations were audited across 24 nursing tasks. Male and female nurses were professionally equivalent and counterbalanced across Candidate A/B; patient gender was varied in six intimate-care tasks. Sixty conditions were repeated three times per model, yielding 540 valid outputs. Results: Male nurses were selected in 50.0% of Technical/Procedural and Relational/Educational allocations. Candidate A was selected in all 288 observations, although 285 (99.0%) assigned identical suitability scores. Because A/B assignment was bundled with labeling, presentation order, evaluation order, and response-template anchoring, this pattern indicates Candidate-A dependence under the tested decision architecture rather than an isolated position effect. Changing patient gender from female to male produced 57 female-to-male nurse switches and no reverse switches (p < 0.001). Conclusions: The audited configurations showed no generalized gender-task pattern but demonstrated decision-architecture dependence and patient-gender-conditioned allocation. Fairness audits should assess equivalent-candidate handling, patient preference, and counterfactual stability.
SN - 2227-9032
ER -