Physician in the Loop
For youSign in
Pulmonary and critical care ·

Graph neural network separates PAH from left heart disease on CT; deep learning predicts long ICU stays in sepsis

An incentive spirometer with a clear graduated chamber and piston, a slim side column holding an amber float, and a mouthpiece on a ribbed hose.
THE SHORT READ

A model combining chest CT and clinical data distinguished pulmonary arterial hypertension from pulmonary hypertension due to left heart disease in 905 patients, Shanghai researchers reported in npj Digital Medicine. Korean researchers reported a model that predicts prolonged intensive care unit (ICU) stays in sepsis from first-day data. Both studies were retrospective.

THE DETAILS
  • AI Oct 8, 2026
    Graph neural network using CT and clinical data separates PAH from left heart disease pulmonary hypertension in 905 patients
    Researchers at Shanghai Pulmonary Hospital reported in npj Digital Medicine on Oct. 8 a model, AHGNN, that combines contrast-enhanced thoracic CT images with noninvasive clinical variables to distinguish pulmonary arterial hypertension from pulmonary hypertension due to left heart disease. In a retrospective study of 905 patients, the mean area under the receiver operating characteristic curve across 100 test folds of repeated cross-validation was 0.946, and sensitivity and specificity were each 0.867 at a 0.50 threshold. The authors said prospective external validation, recalibration and clinical safety evaluation are required before clinical use.
    Source: npj Digital Medicine
  • AI Oct 9, 2026
    Deep learning model predicts ICU stays longer than four days in sepsis from first-day SOFA data
    Researchers at Chungbuk National University Hospital in South Korea reported in npj Digital Medicine on Oct. 9 a deep learning model, FT-TabNet, that uses 17 Sequential Organ Failure Assessment-based features from the first 24 hours of intensive care unit (ICU) admission to predict a stay of more than four days in patients with sepsis. Developed on MIMIC-IV data, it had an area under the receiver operating characteristic curve of 0.848 in internal validation and 0.823 in external validation at Chungbuk National University Hospital; the abstract also reports 0.781 for a second cohort. The study measured prediction accuracy, not patient outcomes.
    Source: npj Digital Medicine

The weekly innovation roundup for pulmonary and critical care, published every Sunday for the week before; every item links to its sources. Latest pulmonary and critical care news · All specialties