Development of the CareContinuityAI Algorithm for Detecting Unresolved Patient-Safety Concerns Across Nursing Handoffs Using Natural Language Processing with Comparative Analysis Against Conventional Structured Handoff and Checklist-Based Methods

Abstract

Failures in the continuity of clinical information during nursing handoffs can allow unresolved patient-safety concerns to persist across shifts, increasing the likelihood of delayed intervention, medication error, deterioration, missed follow-up, and incomplete escalation. Conventional structured handoff approaches such as Situation-Background-AssessmentRecommendation and checklist-based transfer protocols improve communication consistency but depend heavily on manual recognition, documentation quality, and the receiving nurse’s ability to interpret fragmented clinical narratives. This study develops CareContinuityAI, a novel natural language processing algorithm for automated detection, tracking, and prioritization of unresolved patient-safety concerns across sequential nursing handoff records. The proposed architecture combines a domain-adapted transformer encoder with clinical named-entity recognition, temporal relation extraction, crosshandoff semantic alignment, concern-resolution state tracking, contradiction detection, and an attention-based continuity risk scoring mechanism. Unlike conventional text classifiers that assess individual handoff notes independently, CareContinuityAI maintains a longitudinal representation of each identified safety concern and determines whether it has been resolved, acknowledged, escalated, deferred, contradicted, or omitted in subsequent handoffs. The model is designed to distinguish high-priority unresolved concerns involving medication administration, abnormal laboratory findings, changes in vital signs, falls risk, infection indicators, device-related complications, pain escalation, pending investigations, and incomplete clinical actions. Comparative evaluation will benchmark CareContinuityAI against structured Situation-Background-AssessmentRecommendation handoff review, checklist-based detection, Term Frequency-Inverse Document Frequency with Support Vector Machine, Bidirectional Long Short-Term Memory networks, Bidirectional Long Short-Term Memory with Conditional Random Field sequence labelling, ClinicalBERT, BioClinicalBERT, and RoBERTa-based classification. Performance will be assessed using precision, recall, F1-score, sensitivity, specificity, area under the receiver operating characteristic curve, area under the precision-recall curve, false-negative rate, concern-resolution accuracy, escalation-detection accuracy, and mean detection latency. Comparative graphs will examine model-level discrimination, unresolved-concern recall, false-negative reduction, detection latency, class-wise safety performance, calibration, and robustness across handoff complexity levels. The central hypothesis is that combining contextual clinical language modelling with explicit longitudinal concern-state tracking will provide stronger continuity-sensitive detection than isolated-note natural language processing models and conventional structured handoff methods. CareContinuityAI therefore provides a technical foundation for intelligent nursing handoff surveillance systems capable of identifying safety concerns that remain clinically unresolved as responsibility transfers between caregivers.