Artificial intelligence to predict the need for mechanical ventilation in cases of severe COVID-19

Mariana Frizzo de Godoy, José Miguel Chatkin, Rosana Rodrigues, Gabriele Carra Forte, Edson Marchiori, Nathan Gavenski, Rodrigo C. Barros, Bruno HochheggerRadiologia 2023 · DOI

TL;DR

This paper introduces an AI model using deep learning on CT scans to predict which COVID-19 patients will need mechanical ventilation. The model achieves high specificity, meaning it's good at correctly identifying patients who won't need ventilation, which is crucial for resource allocation during a pandemic.

What Problem Does This Paper Address?

The COVID-19 pandemic placed immense strain on healthcare systems, particularly intensive care units (ICUs). A critical challenge was predicting which patients with severe COVID-19 would require mechanical ventilation (MV), a resource-intensive intervention. Early and accurate prediction of MV needs is vital for efficient allocation of ventilators, ICU beds, and staff, especially in situations with limited resources. While RT-PCR tests are used for diagnosis, they have varying sensitivity. Chest CT scans, however, have shown high sensitivity in detecting COVID-19 related lung changes. This paper aims to leverage AI, specifically deep learning applied to CT imaging, to provide a reliable prediction of MV need, thereby aiding in proactive patient management and resource planning.

Methods

The researchers developed an Artificial Intelligence (AI) model based on Convolutional Neural Networks (CNNs), a type of deep learning architecture well-suited for image analysis. They employed a transfer learning approach, using a model pre-trained on the ImageNet dataset and then fine-tuning it on their COVID-19 CT data. The model was trained to analyze individual CT slices, treating each as an input. For validation, they used k-fold cross-validation (k=5) to assess performance metrics across different data splits. The model's output is a score between 0 and 1, indicating the predicted risk of requiring MV.

Datasets

The study utilized a retrospective cohort of CT scans from patients hospitalized with severe acute respiratory syndrome and confirmed COVID-19.

  • Training Set: 823 patients with COVID-19, of whom 93 required MV. This set comprised 189,290 CT slices.
  • Test Set: 114 patients with COVID-19, of whom 67 required MV. This set comprised 28,500 CT slices. The data was collected from two tertiary hospitals in Brazil between April 1, 2020, and May 31, 2020.

Key Insights and Ideas

  • CT Imaging as a Predictive Tool: The study reinforces the value of CT imaging beyond just diagnosis, demonstrating its potential for predicting disease severity and resource needs.
  • High Specificity is Key for Resource Management: The AI model's high specificity (correctly identifying those who don't need MV) is highlighted as its most valuable attribute for practical application. This allows for confident exclusion of patients from intensive resource allocation, preventing overestimation of needs.
  • AI for Pandemic Preparedness: The research suggests that AI can be a powerful tool for hospital management during pandemics, enabling proactive planning for critical care resources.
  • Data-Driven Prediction without Clinical Data: A significant advantage noted is the model's ability to predict MV needs using only CT imaging, which is particularly useful in real-world scenarios where complete clinical data might be scarce or difficult to obtain.

Key Results

The AI model demonstrated the following performance metrics:

  • Overall Sensitivity: 0.417
  • Overall Specificity: 0.860
  • Area Under the ROC Curve (AUC) for the test set: 0.68

The paper emphasizes that the high specificity (0.860) is crucial, indicating that the model is reliable in identifying patients who will not require invasive ventilation.

Limitations and Future Work

The authors acknowledge several limitations:

  • Retrospective Design: The study design is retrospective, which inherently carries a risk of selection bias.
  • Lack of Transparency (Black Box Problem): Deep learning models, including CNNs, can be "black boxes," making it difficult to fully interpret which specific imaging features the AI is using to make its predictions.

Future work suggested includes:

  • Validation in Multiple Cohorts: The authors stress the need to validate their AI model in diverse patient populations and clinical settings to confirm its generalizability and robustness.
  • Integration with Clinical Data: While the current model focuses solely on CT imaging, future iterations could potentially benefit from incorporating clinical data for even more comprehensive prediction.

Summary updated

Abstract

The high specificity of our AI model makes it able to reliably predict which patients will and will not need invasive ventilation. That makes this approach ideal for identifying high-risk patients and predicting the minimum number of ventilators and critical care beds that will be required.
BibTeX
@inproceedings{godoy2023artificial,
  author = {Mariana Frizzo de Godoy and José Miguel Chatkin and Rosana Rodrigues and Gabriele Carra Forte and Edson Marchiori and Nathan Gavenski and Rodrigo C. Barros and Bruno Hochhegger},
  title = {{Artificial intelligence to predict the need for mechanical ventilation in cases of severe COVID-19}},
  booktitle = {Radiologia},
  year = {2023},
  doi = {10.1590/0100-3984.2022.0049},
}
Figure