Hospital-acquired infections (HAIs) remain a major patient-safety challenge because infection risk develops dynamically from interactions among patient vulnerability, clinical interventions, nursing observations, device exposure, medication use, mobility status, hygiene-related care activities, and evolving physiological abnormalities. Conventional HAI prediction approaches frequently depend on static demographic variables, diagnosis codes, laboratory results, or isolated clinical measurements, thereby underutilizing the temporal and contextual information contained in nursing assessments and patient-care activities. This study proposes a novel NurseCare Infection Risk Temporal Fusion Network (NCIR-TFN) for early prediction of hospital-acquired infection through integrated analysis of nursing assessments, patient-care activities, and longitudinal electronic health record data. The proposed framework combines temporal feature encoding, transformer-based longitudinal representation learning, gated attention, clinical-event embeddings, missingness-aware feature processing, and dynamic risk calibration to model changing infection susceptibility throughout hospitalization. Structured nursing observations including skin integrity, consciousness, mobility, continence, wound condition, nutrition, temperature, device status, and infection-related assessment findings are synchronized with care activities such as catheter management, vascular-line care, wound dressing, repositioning, hygiene interventions, medication administration, and invasive procedures. These variables are integrated with laboratory measurements, vital signs, comorbidities, medication histories, and encounter-level EHR trajectories. NCIR-TFN is comparatively evaluated against Logistic Regression, Random Forest, XGBoost, Long Short-Term Memory networks, and conventional Transformer-based prediction models using AUROC, AUPRC, sensitivity, specificity, F1-score, calibration error, false-negative rate, and prediction lead time. Comparative graphs, receiver-operating characteristic curves, precision-recall plots, calibration plots, temporal risk trajectories, feature-attribution charts, and ablation analyses are used to evaluate predictive and clinical performance. The proposed model is designed to provide superior discrimination and earlier risk identification by explicitly capturing interactions between nursing-care processes and changing patient states that are often absent from existing prediction systems. The resulting framework supports continuous HAI surveillance, interpretable early-warning generation, targeted preventive nursing intervention, and data-driven infection-control decision support within hospital electronic health record environments.