Bridging Clinical Sleep Reports and AI: An Open, Reproducible Polysomnography Metrics Framework

Polysomnography (PSG) is the reference examination for diagnosing sleep-disordered breathing, but its results are delivered as static, human-readable PDF reports that lock clinically scored metrics away from automated reuse, hindering reproducibility and the development of artificial-intelligence (AI) tools. Mainstream public sleep repositories such as SHHS and Sleep-EDF distribute only raw physiological signals, not the report-level summary metrics that laboratories actually produce, leaving a distinct and unfilled gap. We developed a reproducible, anonymization-first framework that converts a full-night, attended Type I PSG report into an open, machine-readable resource. Numerical tables are extracted through a semi-automated procedure and normalized into a tidy schema (domain, metric, segment, value, unit), then serialized into mirrored CSV and JSON files with a complete data dictionary. Identifiers are removed, automated consistency checks verify coherence, the dataset is deposited on IEEE DataPort with a persistent DOI, and the scripts are released on GitHub and archived on Zenodo. The framework produced an open dataset of 76 summary metrics across six domains (sleep architecture, respiratory, oximetry, cardiac, movement). In a preliminary estimate with three operators, the pipeline reduced curation of these metrics from 24.0 ± 6.0 min (fully manual) to 4.7 ± 1.3 min including verification (an ~81% reduction), while the automated serialization-and-validation step alone ran in under one second; consistency checks passed (apnea components = 135; per-stage durations = 268-minute total sleep time). In the demonstration case, the pipeline reproduced severe obstructive sleep apnea (apnea–hypopnea index 30.2 events/h; REM 48.5; oxygen desaturation index 24.4 events/h; minimum SpO₂ 67%), fully recomputable from the CSV. By capturing report-level summary metrics absent from raw-signal repositories such as SHHS and Sleep-EDF, the framework delivers a replicable, privacy-preserving, AI-ready data template, supporting benchmarking, education, and future wearable and edge-computing screening. The contribution is methodological and technological, demonstrated as a single-subject proof of concept.