CXR‑Hemo: A Calibrated Logistic Model and Deployable API for Threshold‑linked Hemodynamic Risk

Abstract

Background: Routine chest radiography (CXR) is inexpensive and ubiquitous. Beyond categorical reads, classical CXR metrics—CTR, VPW, and a lung transparency index (LTI)—encode hemodynamic and pulmonary information. We present CXR‑Hemo, a math‑first framework and deployable API that converts these metrics together with bedside physiology (SpO2, ABG/VBG, MAP) into calibrated probabilities linked to clinical decision thresholds via decision‑curve analysis (DCA).
Methods: We formalize a constrained logistic risk function with monotonic feature effects and a reader‑in‑the‑loop design. Image metrics are computed automatically from standardized definitions: CTR from transverse diameters; VPW from mediastinal landmarks with projection‑aware scaling (millimeters via DICOM PixelSpacing); and LTI from normalized lung‑mask intensity. Physiologic inputs include SpO2, ABG/VBG surrogates, lactate, and MAP. Post‑hoc probability calibration uses isotonic regression or temperature scaling with reliability curves and Brier score auditing. Clinical utility is summarized across 0.2–0.6 thresholds using DCA. We expose three JSON endpoints—/cxr/metrics, /physio/ingest, and /risk/hemo—to return calibrated risk, attention weights, and net‑benefit snapshots.
Findings (analytical demonstration): We provide calibration diagnostics and DCA curves using simulated data to illustrate threshold‑linked interpretation, accompanied by a schematic depicting dual encoders, a modality gate α, a constrained logistic head, and post‑hoc calibration to produce an actionable probability.
Interpretation: By explicitly foregrounding calibration and DCA, CXR‑Hemo links probabilities to threshold‑anchored actions and can avert discretionary CT in risk ranges where model‑guided net benefit exceeds ‘CT‑for‑all’ [1,2]. Given the large dose differential between CXR and chest CT in adults (≈0.1 mSv vs. several mSv), the framework has the potential to reduce population radiation exposure while maintaining clinical safety [3,4].
CXR‑Hemo reframes routine CXR and basic gases as a probabilistic, interpretable sensor for hemodynamic risk. By foregrounding calibration and DCA, the approach links probabilities to actions, enabling a transparent, threshold‑aware tool suitable for resource‑variable emergency care.