AI Chest X‑Ray + Clinical Data Predicts ICU Admission (AUC 0.92) in Hospitalized COVID‑19 Patients

Read the paper →

AI-Driven Chest X-Ray and Clinical Data Integration Predicts Severe COVID-19 and Mortality With High Accuracy

A study of 2,552 hospitalized COVID-19 patients demonstrates that combining automated chest radiograph analysis with key clinical variables can reliably predict ICU admission risk (AUC 0.92) and in-hospital mortality (AUC 0.81).

Introduction

Since the emergence of COVID-19, healthcare systems worldwide have faced an urgent challenge: rapidly identifying which patients are at greatest risk of progressing to severe disease or death. Timely and accurate risk stratification is critical for allocating intensive care resources and optimizing patient management. However, current approaches often rely either solely on clinical data or on chest imaging interpreted by radiologists, both of which have notable limitations. Clinical scoring systems sometimes lack desired accuracy, while radiological assessment is susceptible to subjective variability, especially in settings without specialized radiologists.

This study presents an innovative artificial intelligence (AI) model that automates the interpretation of chest radiographs integrated with readily available clinical variables to predict critical outcomes in COVID-19 patients. By leveraging deep learning for image analysis and advanced statistical methods for clinical data, the model delivers robust prediction of ICU admission and mortality. Key findings include an impressive area under the curve (AUC) of 0.92 for ICU admission prediction and 0.81 for mortality, outperforming models using clinical or imaging data alone.

Study Partnership & Context

The research was conducted by a multidisciplinary team led by Nicolás Munera and Luis F. Reyes, including institutions such as Arkangel AI (Bogotá), Universidad de la Sabana and Clínica Universidad de La Sabana (Chía, Colombia), and collaborators from Spain and the UK. The study harnessed data from the Latin American Intensive Care Network’s LIVEN COVID-19 registry, reflecting a diverse patient population hospitalized across 22 hospitals in eight Latin American countries during the first year of the pandemic (March 2020 – January 2021).

This Latin American context is significant because healthcare resource variability and radiology expertise disparities are common, underscoring the value of an AI tool that can facilitate objective chest X-ray interpretation without requiring expert radiologists.

Study Design and Methodology

The study was a prospective diagnostic test evaluation including 3,007 patients with confirmed SARS-CoV-2 infection via RT-PCR. After exclusions for missing clinical or imaging data, 2,552 patients were analyzed for ICU admission and mortality based on clinical variables, while 582 patients with high-quality frontal chest radiograph images comprised the imaging cohort for model development and validation.

Data encompassed socio-demographics, comorbidities, symptoms, vital signs, laboratory results, and chest X-ray images captured upon hospital admission. Obesity was recorded based on physician diagnosis (BMI>30). Clinical data were collected using REDCap in a standardized manner.

AI model architecture combined two main components:

• Image model: A convolutional neural network (CNN) was fine-tuned using transfer learning with ImageNet-pretrained backbones (e.g., DenseNet121, InceptionV3) to extract radiographic features from chest X-rays. Images were pre-processed to normalize contrast and exclude low-quality scans.
• Clinical model: Key clinical variables were identified via random forest and logistic regression analyses, including age, fraction of inspired oxygen (FiO2), dyspnoea, obesity, blood pressure, oxygen saturation, and Glasgow Coma Scale. These formed the inputs to a simple perceptron model.

A combined model integrated outputs from both CNN imaging and clinical models by linking their sigmoid output layers to generate a final risk prediction. Training employed an 70/12/18% split for training, validation, and testing (images), with cross-validation to mitigate overfitting. Outcomes of interest were ICU admission and in-hospital mortality.

Key Results

• ICU Admission Prediction:
  • Image-only model AUC: 0.88 ± 0.05
  • Clinical-only model AUC: 0.90 ± 0.04
  • Combined model AUC: 0.92 ± 0.04 (statistically superior, p < 0.0001 compared to individual models)
  • Combined model sensitivity: 91%; specificity: 78%
• Hospital Mortality Prediction:
  • Image-only model AUC: 0.75 ± 0.07
  • Clinical-only model AUC: 0.81 ± 0.06
  • Combined model AUC: 0.81 ± 0.06 (no significant improvement over clinical alone; p=0.13)
  • Combined model sensitivity: 74%; specificity: 75%
• Important clinical predictors included age, FiO2 on admission (strongest association), dyspnoea, obesity, blood pressure, oxygen saturation, Glasgow Coma Scale, male sex, and hypertension.

Interpretation & Implications

This study establishes that AI-driven automated interpretation of chest radiographs, when combined with key clinical variables, substantially improves the prediction of severe COVID-19 requiring ICU admission compared to using either data source alone. While imaging features alone held predictive value, clinical variables contributed critical context improving overall model accuracy. For hospital mortality, clinical data remained the primary driver, and imaging added limited additional predictive power.

For clinicians and health systems, this means: automated AI support tools could help rapidly identify high-risk patients on admission, guiding resource allocation and timely intervention. The AI model reduces dependency on radiologist expertise by providing objective, reproducible chest X-ray analysis — an advantage in resource-limited or overburdened settings common in Latin America and beyond.

Limitations include the relatively smaller subset of patients with adequate imaging data and some variability in image quality. Additionally, deep learning models remain “black boxes,” potentially limiting clinician trust without enhanced interpretability. Future work should focus on external validation, integration into clinical workflows, and improving model transparency.

Deployment & Scalability

The model has potential for deployment within hospital information systems to provide real-time risk scores upon admission. Implementation requires integration with existing radiology imaging and electronic health records to automatically input patient data and chest X-rays for analysis. Barriers include variability in image acquisition protocols, hardware availability, regulatory approvals, and clinician acceptance.

However, once integrated, the approach could be adapted for other respiratory diseases where chest radiographs and clinical indicators predict severity, such as pneumonia or ARDS. Moreover, use of open-source AI backbones and cloud-based processing can facilitate deployment in resource-constrained environments.

Conclusion & Next Steps

This study advances the use of AI in COVID-19 by demonstrating that combining automated chest radiograph interpretation with clinical data yields highly accurate prediction of severe disease and mortality. The results support the integration of imaging AI models into clinical decision-making to assist frontline providers worldwide, especially in settings lacking radiology expertise.

Future research priorities include external validation in diverse populations, prospective clinical trials assessing impact on patient outcomes, and enhancing model explainability to foster clinician trust. Sustained efforts toward scalable AI tools will help health systems respond more effectively to current and future respiratory pandemics.

References

Munera N, Garcia-Gallo E, Gonzalez Á, et al. A novel model to predict severe COVID-19 and mortality using an artificial intelligence algorithm to interpret chest radiographs and clinical variables. ERJ Open Res. 2022; 8(2): 00010-2022. https://doi.org/10.1183/23120541.00010-2022