Artificial Intelligence Decision Support and Nursing Care Quality in Iraqi Teaching Hospitals: A Multicenter Cluster Quasi-Experimental Study

Abstract

Background: Clinical decision-making in nursing is a complex, high-stakes activity, yet evidence on artificial-intelligence-based clinical decision support systems (AI-CDSS) inside Iraqi teaching hospitals remains scarce. Aim: I evaluated the effect of an AI-CDSS intervention on the quality of nursing care in three Iraqi teaching hospitals and identified moderating contextual factors. Methods: A multicenter, ward-level cluster-randomized, quasi-experimental pre–post study was conducted from January to August 2025 across 18 wards (six per hospital) in three teaching hospitals in central Iraq, reporting in line with the CONSORT-AI and DECIDE-AI extensions. The protocol was retrospectively recorded in the the Al-Turath University Research Implementation Registry (ATU-RIR-2025-04). Eighteen wards were randomly allocated 1:1 (nine per arm) using block randomization stratified by hospital, with concealed allocation by an independent statistician; sample-size calculation incorporated an intracluster correlation coefficient (ICC) of 0.07 and a design effect of 2.02. A total of 280 registered nurses received either the AI-CDSS intervention (n = 140) or matched-intensity control content (n = 140) for 16 weeks. The intervention was a gradient-boosted, locally recalibrated AI module embedded in the electronic nursing record (pre-deployment AUROC 0.84, 95% CI 0.81–0.87; week-8 calibration slope 0.93–0.99 across sites). Primary analyses used linear mixed-effects models with random intercepts for hospital and ward; sensitivity analyses used GEE with cluster-robust standard errors. Subgroup fairness audits were performed across sex, age, and ward type. Analyses followed an intention-to-treat principle (3.2% missingness, multiple imputation). Results: In mixed-effects analysis, decision accuracy increased by 12.6 percentage points (95% CI 9.8–15.4) in the AI-CDSS arm relative to controls (Bonferroni-adjusted p < 0.001; ICC 0.06). Medication errors fell by 3.7 events per 1000 patient-days (95% CI 2.9–4.5; ICC 0.08). The QNCI rose by 14.5 points (95% CI 11.7–17.3). Subgroup audits revealed no major effect-size disparities by sex or age. Patient-level safety outcomes (falls, pressure ulcers) showed favorable trends but did not reach significance. Conclusion: AI-CDSS implementation showed initial associations with improved nursing process measures in Iraqi teaching hospitals, with no significant effect on patient-level safety outcomes within the 16-week window. Because primary measurement relied partly on clinical vignettes and the QNCI is investigator-developed, replication with longer follow-up, real-bedside outcomes, external QNCI validation, prospective registration, and full economic evaluation is required before any scaling decision.