Abstract
Risk management and compliance: projects around governance, compliance and/or ethics and societal impact Misplaced nasogastric tubes can cause serious patient harm, including aspiration, pneumothorax, delayed feeding and death.Although recent artificial intelligence (AI) systems have shown that feeding tube position can be robustly assessed on chest radiographs, important gaps remain between technical feasibility and safe clinical integration.This study examined how AI for feeding tube assessment should be designed and introduced within existing care pathways to improve clinical utility, safety and acceptability.A qualitative human-centred study was undertaken using contextual inquiry and semi-structured interviews with 15 stakeholders involved in feeding tube placement, chest radiograph acquisition, image interpretation and feeding decisions.The sample included intensive care clinicians and nurses, stroke clinicians and reporting radiographers.Five prospective AI roles were explored: early misplacement alerting, radiology worklist prioritisation, cross-checking of human assessment, visual overlay guidance and draft report generation.Data were analysed using reflexive thematic analysis, with additional review of documentation practices relevant to future data set curation and outcome measurement.Challenges extended beyond image interpretation alone and arose across the end-to-end workflow.Participants described delays related to staffing pressures, communication inefficiencies, task fragmentation, image upload dependencies and multiple safeguarding steps before feeding could begin.Perceived AI usefulness depended on timing of intervention, target user, accountability, verification burden and interaction with existing safeguarding processes.Participants generally favoured AI that strengthened human review rather than stand-alone automation.The study also identified important data challenges, including edge cases, data bias, variable documentation practices and uncertainty around suitable labels and clinically meaningful evaluation metrics.Clinically useful radiology AI requires more than high model accuracy.Early workflow-centred design can identify safer and more acceptable AI roles, clarify benefit-risk and feasibility trade-offs, and better guide data set curation, implementation strategy and prospective evaluation of responsible AI in practice.