A model based on CT radiomic features for predicting RT-PCR becoming negative in coronavirus disease 2019 (COVID-19) patients

Quan Cai, Si-Yao Du, Si Gao, Guo-Liang Huang, Zheng Zhang, Shu Li, Xin Wang, Pei-Ling Li, Peng Lv, Gang Hou, Lina ZhangBMC Medical Imaging 2020 · DOI

TL;DR

This paper introduces a novel model that combines CT radiomic features with clinical data to predict when COVID-19 patients are likely to test negative via RT-PCR. This approach aims to optimize RT-PCR retesting schedules, reduce medical waste, and shorten hospital stays for recovering patients.

What Problem Does This Paper Address?

Discharging COVID-19 patients in China relies on negative RT-PCR tests. However, repeated testing is resource-intensive and can prolong hospital stays, especially for patients in recovery. Doctors need a more objective and efficient way to predict when a patient is likely to clear the virus, as indicated by a negative RT-PCR result. While chest CT scans are useful for diagnosing and monitoring COVID-19 pneumonia, distinguishing between patients who are still infectious (RT-PCR positive) and those who are likely recovering (RT-PCR negative) based solely on CT images can be challenging. This study aims to bridge this gap by leveraging advanced image analysis techniques to predict RT-PCR negativity.

Methods

The researchers developed a predictive model using a two-stage approach:

  1. Image Segmentation and Feature Extraction: They employed a deep learning algorithm to automatically segment lung lobes and identify lesions within chest CT scans. This segmentation process was crucial for accurately extracting a comprehensive set of radiomic features.
  2. Model Development:
    • Feature Selection: After extracting 120 radiomic features and collecting 20 clinical characteristics, they used univariate logistic regression to identify features significantly associated with RT-PCR negativity. Highly correlated features were then removed using Spearman correlation analysis.
  3. Model Building: A multivariate logistic regression model was constructed using the selected features. This model aimed to predict the probability of a patient being RT-PCR negative.
  4. Evaluation: The model's performance was assessed using Receiver Operating Characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA) on both training and testing datasets.

Datasets

The study retrospectively analyzed data from 203 mild COVID-19 patients treated at a Fangcang Shelter Hospital in Wuhan, China, between February 10 and March 10, 2020. The dataset was split into a training cohort (n=141) and a testing cohort (n=62).

Key Insights and Ideas

  • Radiomics Beyond Visual Inspection: The study highlights that while visual assessment of CT scans can be difficult for distinguishing RT-PCR positive from negative cases, deep learning-based segmentation and radiomic feature extraction can reveal subtle patterns indicative of viral clearance.
  • Time is a Key Clinical Predictor: The time interval from symptom onset to CT examination emerged as a significant clinical predictor of RT-PCR negativity. This aligns with the understanding that viral shedding decreases over time.
  • Radiomic Features Capture Deeper Information: The model's strong performance, driven primarily by radiomic features, suggests that these quantitative descriptors capture more detailed information about the internal heterogeneity of lung lesions than standard CT quantitative parameters.
  • Optimizing RT-PCR Testing: The developed model has the potential to guide clinical decisions on when to perform RT-PCR retesting, moving beyond a fixed schedule and potentially reducing unnecessary tests.

Key Results

The developed model, combining CT radiomic features and the time interval from symptom onset to CT exams, demonstrated good predictive performance:

  • AUC: The Area Under the Curve (AUC) for the ROC analysis was 0.811 for the training dataset and 0.812 for the testing dataset.
  • Sensitivity/Specificity: The model achieved a sensitivity of 0.765 and specificity of 0.625 in the training set, and 0.784 sensitivity and 0.600 specificity in the testing set.
  • Clinical Parameter: The time interval from symptom onset to CT exams was the only clinical parameter selected for the final model, with a higher interval associated with a higher odds ratio (OR=2.84) for RT-PCR negativity.
  • Radiomic Features: Nine out of the ten features in the final model were radiomic features, underscoring their importance.

Limitations and Future Work

The authors acknowledge several limitations:

  • Retrospective Study: The study design is retrospective, which inherently has limitations.
  • Mild Cases Only: The model was developed and validated on mild COVID-19 cases, meaning its applicability to severe or critical cases is unknown.
  • Single-Center Study: The data was collected from a single center, suggesting the need for multi-center validation to ensure generalizability.
  • Limited Laboratory Data: Not all relevant laboratory tests were performed on all patients, and the model did not include laboratory indicators.
  • Model Type: Only one model type (multivariate logistic regression) was used, and comparative analyses with other machine learning models (e.g., decision trees, random forests) were not conducted.
  • Biological Interpretation: The biological interpretation of the specific radiomic features used in the model was not fully explored.

Future work should focus on multi-center validation, exploring other model types, and investigating the biological underpinnings of the identified radiomic features.

Summary updated

BibTeX
@article{cai2020model,
  author = {Quan Cai and Si-Yao Du and Si Gao and Guo-Liang Huang and Zheng Zhang and Shu Li and Xin Wang and Pei-Ling Li and Peng Lv and Gang Hou and Lina Zhang},
  title = {{A model based on CT radiomic features for predicting RT-PCR becoming negative in coronavirus disease 2019 (COVID-19) patients}},
  journal = {BMC Medical Imaging},
  year = {2020},
  doi = {10.1186/s12880-020-00521-z},
}
Figure