AI-assisted identification of potential treatment injury from inpatient discharge summaries: an early-stage sociotechnical evaluation in a New Zealand district hospital
- 1Health New Zealand
Abstract
ObjectivesTo assess the feasibility and early performance of an AI-assisted workflow screening inpatient discharge summaries for potential treatment injury requiring clinician review and possible Accident Compensation Corporation (ACC) consideration in New Zealand.
DesignSingle-site, six-month quality-improvement/service-evaluation pilot using a large language model as an index screening tool with clinician adjudication.
SettingWairarapa Hospital, New Zealand.
ParticipantsAll inpatient discharge summaries produced during the pilot period (n=2,642).
InterventionsThe model generated structured candidate harm signals with supporting text and certainty labels. Clinicians reviewed all flagged cases and determined whether escalation was appropriate.
Main outcome measuresScreening yield, positive predictive value (PPV) for clinician-adjudicated iatrogenic harm, unsupported outputs, overlap with incident reporting and ACC claims, and clinician workload.
ResultsThe workflow generated 237 candidate signals (yield 9.0%, 95% CI 7.9% to 10.1%). Clinicians classified 102 as confirmed iatrogenic harm, 43 as possible harm, 81 as no harm and 11 as unsupported outputs. PPV was 43.0% for confirmed harm (95% CI 36.8% to 49.4%) and 61.2% for confirmed-or-possible harm (95% CI 54.9% to 67.1%); unsupported outputs occurred in 4.6%. Seven ACC claims were lodged following AI-assisted identification. Of 68 organisational ACC claims during the same period, 17 overlapped with AI-identified cases. The workflow required approximately 25-30 minutes of clinician time per week.
ConclusionsAI-assisted discharge-summary screening was operationally feasible and increased visibility of potential treatment injuries for clinician review. Further prospective evaluation should use an independent chart-review reference standard to estimate diagnostic accuracy and assess equity, workload and implementation outcomes.
RegistrationNot applicable.
Key messagesO_ST_ABSWhat is already known on this topicC_ST_ABSHealthcare-associated harm is systematically under-detected, with incident reporting and administrative data capturing only a fraction of treatment injuries. Structured record review methods such as the Global Trigger Tool identify substantially higher rates of adverse events but are resource-intensive and difficult to sustain at scale.
What this study addsThis study demonstrates that an AI-assisted workflow can feasibly screen inpatient discharge summaries to identify potential treatment injuries for clinician review within routine workflows. The findings suggest that AI can function as a form of visibility infrastructure, surfacing harm signals not consistently captured by existing reporting systems while preserving clinician judgement and governance processes.
How this study might affect research, practice or policyAI-assisted screening could strengthen patient safety surveillance and support more consistent consideration of treatment injury within compensation systems. Future research should evaluate diagnostic accuracy, equity impacts, and workflow design, including whether carefully governed automation of administrative steps can improve escalation processes without displacing clinical oversight.
DisclaimerThe AI system generated candidate signals of potential treatment injury only. Final determination of treatment injury status and ACC eligibility remained the responsibility of clinicians and the Accident Compensation Corporation through standard statutory processes.
Related
- Computational identification and evaluation of Adenium obesum phytochemicals as potential EGFR inhibitors for targeted therapy against non-small-cell lung cancer
- Heterogeneity in non-suicidal self-injury ideation and treatment response among inpatient female adolescents with mood disorders: Insights from ecological momentary assessment
- PharmaAgent: a real-world evaluation of high- and low-autonomy agentic systems for inpatient vancomycin dosing
- Mixed-Methods Analysis of a National Survey of Hospitals With Inpatient Hepatitis C Treatment Initiatives
- MAP: evaluation and multi-agent enhancement of large language models for inpatient pathways
BibTeX
@misc{bhatnagar2026aiassisted,
title = {{AI-assisted identification of potential treatment injury from inpatient discharge summaries: an early-stage sociotechnical evaluation in a New Zealand district hospital}},
author = {P. Bhatnagar and W. Fearnley},
howpublished = {medRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.09.26365175},
url = {https://www.medrxiv.org/content/10.64898/2026.10.09.26365175v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - AI-assisted identification of potential treatment injury from inpatient discharge summaries: an early-stage sociotechnical evaluation in a New Zealand district hospital
AU - P. Bhatnagar
AU - W. Fearnley
PY - 2026
DA - 2026/10/09
PB - medRxiv
DO - 10.64898/2026.10.09.26365175
UR - https://www.medrxiv.org/content/10.64898/2026.10.09.26365175v1
AB - ObjectivesTo assess the feasibility and early performance of an AI-assisted workflow screening inpatient discharge summaries for potential treatment injury requiring clinician review and possible Accident Compensation Corporation (ACC) consideration in New Zealand. DesignSingle-site, six-month quality-improvement/service-evaluation pilot using a large language model as an index screening tool with clinician adjudication. SettingWairarapa Hospital, New Zealand. ParticipantsAll inpatient discharge summaries produced during the pilot period (n=2,642). InterventionsThe model generated structured candidate harm signals with supporting text and certainty labels. Clinicians reviewed all flagged cases and determined whether escalation was appropriate. Main outcome measuresScreening yield, positive predictive value (PPV) for clinician-adjudicated iatrogenic harm, unsupported outputs, overlap with incident reporting and ACC claims, and clinician workload. ResultsThe workflow generated 237 candidate signals (yield 9.0%, 95% CI 7.9% to 10.1%). Clinicians classified 102 as confirmed iatrogenic harm, 43 as possible harm, 81 as no harm and 11 as unsupported outputs. PPV was 43.0% for confirmed harm (95% CI 36.8% to 49.4%) and 61.2% for confirmed-or-possible harm (95% CI 54.9% to 67.1%); unsupported outputs occurred in 4.6%. Seven ACC claims were lodged following AI-assisted identification. Of 68 organisational ACC claims during the same period, 17 overlapped with AI-identified cases. The workflow required approximately 25-30 minutes of clinician time per week. ConclusionsAI-assisted discharge-summary screening was operationally feasible and increased visibility of potential treatment injuries for clinician review. Further prospective evaluation should use an independent chart-review reference standard to estimate diagnostic accuracy and assess equity, workload and implementation outcomes. RegistrationNot applicable. Key messagesO_ST_ABSWhat is already known on this topicC_ST_ABSHealthcare-associated harm is systematically under-detected, with incident reporting and administrative data capturing only a fraction of treatment injuries. Structured record review methods such as the Global Trigger Tool identify substantially higher rates of adverse events but are resource-intensive and difficult to sustain at scale. What this study addsThis study demonstrates that an AI-assisted workflow can feasibly screen inpatient discharge summaries to identify potential treatment injuries for clinician review within routine workflows. The findings suggest that AI can function as a form of visibility infrastructure, surfacing harm signals not consistently captured by existing reporting systems while preserving clinician judgement and governance processes. How this study might affect research, practice or policyAI-assisted screening could strengthen patient safety surveillance and support more consistent consideration of treatment injury within compensation systems. Future research should evaluate diagnostic accuracy, equity impacts, and workflow design, including whether carefully governed automation of administrative steps can improve escalation processes without displacing clinical oversight. DisclaimerThe AI system generated candidate signals of potential treatment injury only. Final determination of treatment injury status and ACC eligibility remained the responsibility of clinicians and the Accident Compensation Corporation through standard statutory processes.
ER -