Development and Multicenter Validation of an Explainable System for Neonatal ARDS-Compatible Phenotyping

Study design and objective This is a three-center observational cohort study combining retrospective system development and validation with prospective validation of a locked system. Its objective is to develop an explainable electronic health record-based system for identifying a NARDS-compatible clinical phenotype and prioritizing candidate records for specialist review.

Establishment of retrospective cohorts Historical NICU records dated January 1, 2018, through December 31, 2025, will provide the retrospective source population. Records from the lead center dated 2018-2022 will form the development cohort. Records from the lead center dated 2023-2025 will form an independent internal temporal validation cohort. Records from each collaborating center dated 2023-2025 will form two separate geographic external validation cohorts. The planned formal adjudication sample is approximately 1,600 candidate events for development and 800 candidate events in each of the three retrospective validation cohorts. Each infant's first eligible event will be used in the primary analysis.

Candidate screening and sampling Candidate records will be identified using prespecified evidence of respiratory support or escalation, bilateral or diffuse pulmonary imaging abnormalities, acute lung injury triggers, or relevant diagnostic expressions. Diagnostic codes will be used for screening and descriptive analyses.

System development and locking The system will extract traceable evidence from imaging reports and other permitted clinical records, including acute respiratory deterioration, acute lung injury triggers, pulmonary imaging abnormalities, and alternative explanations. It will assign high-priority review, routine review, or insufficient-information status and retain the supporting source and timestamp. The variable dictionary, text-processing rules, feature mappings, decision rules, thresholds, software version, and analysis plan will be locked before validation. No retraining, threshold selection, or center-specific tuning will use validation labels.

Reference standard Two qualified neonatal intensive care specialists will independently assess each evidence package using a prespecified adjudication manual. Disagreement will undergo independent arbitration and, when necessary, structured consensus. Adjudicators will be masked to system outputs, explanations, discharge diagnostic labels, and diagnostic codes. They will assign one of four categories: high-likelihood NARDS-compatible clinical phenotype, non-compatible clinical phenotype, clinically uncertain, or insufficient information. The high-likelihood phenotype requires an acute respiratory course, evidence of an acute lung injury trigger, bilateral or diffuse pulmonary parenchymal imaging abnormalities, and no single alternative explanation that fully accounts for the findings.

Prospective validation After retrospective temporal and geographic validation, the same locked system will be assessed in newly eligible infants at the participating centers. The proposed prospective phase will include consecutive eligible infants and will preserve the same index-event definition, evidence window, reference standard, and output categories. Results will be generated in silent mode and will not be displayed to treating clinicians or used to guide care. Independent adjudicators will remain masked to system outputs. Prospective sample size, center participation, recruitment dates, and consent arrangements will be finalized in the protocol and ethics documentation before this phase begins.

Performance analysis The primary measure will be sensitivity of high-priority review for the reference-standard high-likelihood phenotype in each independent validation cohort. Results will be presented separately for the lead-center temporal cohort, each external center, and the prospective phase, with center-specific prospective results when sample size permits. Ninety-five percent confidence intervals will accompany performance estimates. Sampling weights will be used when estimating performance in the historical candidate population. These categories will remain visible in study flow and sensitivity analyses. The primary analysis uses one event per infant.

Data governance Identifiable records, original free text, images, and identity linkage files will remain within each center's authorized environment. Locked code will preferentially be run locally at collaborating centers. Only approved aggregate results and the minimum permitted de-identified audit information will be transferred to the coordinating center. No research-specific treatment or additional biological sampling is planned.

Time frame: One assessment per first eligible respiratory event, using records from 48 hours before to 48 hours after the first screening trigger (T0).

The investigators will calculate the high-priority proportion among expert-adjudicated NARDS-compatible infants, excluding routine-review and insufficient-information outputs, and report performance with 95% CIs by cohort, with uncertain assessments reported separately.

Time frame: One assessment per first eligible respiratory event, using records from 48 hours before to 48 hours after T0.

Among high-priority infants with determinate reference, report adjudicated high-likelihood NARDS-compatible proportion and 95% CI by cohort. Enriched samples yield sampling-weighted candidate estimates if permitted. It is not NICU-wide incidence.

Time frame: One assessment per first eligible respiratory event, using records from 48 hours before to 48 hours after T0.

Among determinate-reference non-high-priority infants, report non-compatible phenotype proportion and 95%CI by cohort; describe routine-review and insufficient-information outputs separately, to avoid concealment. Weight historical estimates if possible.

Time frame: At reference adjudication of the first eligible event, using the same records from 48 hours before to 48 hours after T0.

Assessed infants with high-likelihood NARDS-compatible or non-compatible phenotype divided by all assessed; uncertain and insufficient-information assessments reported separately with reasons and 95% CIs, not relabeled negative or removed from flow.

Development and Multicenter Validation of an Explainable System for Identifying Clinical Phenotypes Compatible With Neonatal Acute Respiratory Distress Syndrome