The AI-derived Congestion Index in the CDSS-HF model improved adverse event prediction in ADHF patients, with a significant increase in event rates from 1.7% to 18.9% across risk quartiles.
Does integrating an AI-derived Congestion Index from chest radiographs with clinical data improve the prediction of adverse outcomes in patients hospitalized with acute decompensated heart failure?
Integrating AI-derived radiographic biomarkers of pulmonary congestion with routine clinical data significantly improves risk stratification for patients hospitalized with acute decompensated heart failure.
Absolute Event Rate: 0% vs 0%
Abstract Background/Introduction The prognosis of patients with acute decompensated heart failure (ADHF) is influenced by demographic, hemodynamic, and biochemical parameters. Chest radiography remains a routine diagnostic modality in these patients, yet its interpretation often lacks objectivity. Advances in deep learning have enabled the extraction of quantitative imaging biomarkers, offering new opportunities to improve risk stratification. Purpose This study aimed to develop and validate a machine learning-based Clinical Decision Support System (CDSS-HF) for predicting adverse outcomes in patients hospitalized with ADHF. The model integrates routinely available clinical variables with a novel AI-derived digital biomarker—Congestion Index (CIx)—quantified from chest radiographs. Methods We retrospectively analyzed data from 9,286 patients hospitalized with ADHF between May 2003 and April 2022. A previously validated convolutional neural network was used to analyze frontal chest radiographs and derive the Congestion Index (CIx), which quantifies the degree of pulmonary congestion. We developed three machine learning models using the PyCaret platform. The first model included basic demographic and hemodynamic variables such as age, sex, systolic blood pressure, and heart rate. The second model added laboratory parameters including hemoglobin and estimated glomerular filtration rate (eGFR). The third model incorporated the CIx and lung volume in addition to the variables from the first two models. Model performance was evaluated by area under the receiver operating characteristic curve (AUROC), and the best-performing algorithm was identified through internal validation. Results The CIx demonstrated predictive performance comparable to NT-proBNP for adverse outcomes (AUROC 0.646 vs. 0.620, P = 0.516). When patients were categorized into quartiles based on CIx values, event rates increased progressively: 3.5% in Q1, 6.3% in Q2, 9.1% in Q3, and 14.6% in Q4 (P 0.001). Among the machine learning classifiers tested, the Extra Trees Classifier showed the highest accuracy. The AUROC of Model 1 was 0.665 (95% CI, 0.629–0.700), that of Model 2 was 0.704 (95% CI, 0.670–0.738), and Model 3 achieved the best performance with an AUROC of 0.750 (95% CI, 0.719–0.780). When the final CDSS-HF score was stratified into quartiles, the incidence of adverse outcomes significantly increased across groups: 1.7% in Q1, 5.3% in Q2, 7.9% in Q3, and 18.9% in Q4 (P 0.001). Conclusions The integration of AI-analyzed chest radiograph features with clinical and laboratory parameters significantly improves prognostic prediction in patients with ADHF. The proposed CDSS-HF model may serve as a valuable tool for personalized risk assessment and clinical decision-making in heart failure care.Performance of Prediction modelsFeature importance
Jeon et al. (Thu,) reported a other. The AI-derived Congestion Index in the CDSS-HF model improved adverse event prediction in ADHF patients, with a significant increase in event rates from 1.7% to 18.9% across risk quartiles.