Clinical Trial Details
— Status: Completed
Administrative data
| NCT number |
NCT04525287 |
| Other study ID # |
2020-12020001231 |
| Secondary ID |
|
| Status |
Completed |
| Phase |
|
| First received |
|
| Last updated |
|
| Start date |
February 20, 2020 |
| Est. completion date |
August 20, 2020 |
Study information
| Verified date |
June 2020 |
| Source |
Second Affiliated Hospital, School of Medicine, Zhejiang University |
| Contact |
n/a |
| Is FDA regulated |
No |
| Health authority |
|
| Study type |
Observational
|
Clinical Trial Summary
Clinical observation has found that COVID-19 patients often present inconsistency of clinical
features, nucleic acid of the SARS-CoV-2 and imaging findings, which brings challenges to the
management of patients.The quantitative assessment of patients' pulmonary lesions of chest
CT, combined with the basic information, epidemiological history, clinical symptoms, basic
diseases and other information of patients, will quickly establish a reliable prediction
model for the severe COVID-19. This model will greatly contribute to the effective diagnosis
and treatment of COVID-19.
Description:
1. Research purpose
The research team collected the clinical and chest CT of 1,000 COVID-19 patients from
multiple hospitals. We plan to use these data to explore the imaging features of the
COVID-19 and develop a convenient, easy-to-use, highly reliable imaging AI model for
detecting and predicting the severe COVID-19. The model is used for imaging evaluation
of COVID-19 patients, in order to achieve the purpose of early diagnosis, reasonable
management of patients and prediction of severe COVID-19.
2. Research design and methods:
This research is a retrospective study. The project research period have 6 months. Start
time: the date of ethics approval.
End time: August 20, 2020.
2.1 Establish an AI model for the detection of COVID-19 chest CT lesions Based on existing
models and data, rapid detection of lesions on the chest high-resolution CT (HRCT) images of
COVID-19, identification of the character of the lesions including the volume of the lesions.
2.1.1 Research data COVID-19 Group: 1,000 cases of COVID-19 patients who were tested positive
for nucleic acid of the SARS-CoV-2 in more than ten designated hospitals within and outside
Zhejiang Province. all underwent chest HRCT examination and relatively complete clinical and
laboratory data.
Control group: patients with other viral and bacterial pneumonia. A collection of 1000
patients with other types of pneumonia in multiple centers, all underwent chest HRCT
examination and relatively complete clinical data.
2.1.2 Research methods
1. Lesion detection, segmentation and quantification. Based on the artificial intelligence
analysis function developed by Yitu Technology, the patient's chest CT image data is
analyzed, including: a. Detecting lung lesions; b. Quantitative and radiomics analysis
of key imaging features such as the shape, extent, and density of the lesions, and
accurately calculating the cumulative pneumonia burden of the disease; c. For focal
lesions, diffuse lesions, quantitative analysis of the severity of various pneumonia
diseases involving the entire lung.
2. Based on the above-mentioned lesion segmentation detection results, analyze the
characteristic manifestations of the COVID-19. Observe the differences in the number,
shape, range, density, and radiomics characteristics of lung lesions in patients with
COVID-19 and other pneumonia, and quantify their unique lung characteristics. Compare
and analyze the correlation of lung characteristics with clinical symptoms and guideline
classification characteristics, and clarify the value of quantitative mathematical
characteristics in auxiliary diagnosis classification.
2.2 Establish an AI model for predicting severe COVID-19 Establish a reliable AI model for
predicting severe COVID-19 through chest CT imaging data of the patient, basic patient
information, epidemiological history, clinical symptoms, and underlying diseases. It is
planned to establish a severe COVID-19 risk assessment system that can not only assist
doctors in the critical evaluation of patients in hospital, but also warn the severe risk of
patients in home isolation. The system will include simple and accurate models. The former
only uses patient images, basic demographic characteristics, symptoms and other
easy-to-collect information. This model may be used in Wuhan City that has basic data for
mild cases but has no treatment conditions. Allow them to be isolated at home or other
patients without medical resources, and early warning of the risk of severe transformation in
the hospital system for preliminary testing, so as to facilitate subsequent patient
management and treatment. The complex model will incorporate more complex and detailed
information, such as multiple images of patients and blood test data, to establish a more
accurate predictive model, which can be used to provide reference for the diagnosis and
treatment strategies of patients in the hospital, and can prompt more COVID-19 Factors
related to pneumonia.
2.2.1 Research object 1,000 patients with mild cases of novel coronavirus pneumonia were
diagnosed, and were divided into severe group and non-severe group based on subsequent
clinical outcomes.
Definition of mild illness: The patient only showed symptoms such as fever and respiratory
tract in general, and no severe symptoms occurred during the visit and follow-up.
Definition of severe illness: Patients who have one of the following conditions during
treatment: 1. Respiratory distress, RR≥30 beats/min; 2. In resting state, mean oxygen
saturation≤93%; 3. Arterial oxygen partial pressure ( PaO2)/Inhalation Oxygen Concentration
(FiO2) ≤300mmHg (1mmHg=0.133kPa); 4. Respiratory failure occurs and mechanical ventilation is
required; 5. Shock occurs; 6. ICU monitoring and treatment is required for combined other
organ failure.
2.2.2 Research methods This study uses artificial intelligence technology to predict the
severity of patients with mild illness. The data required for its modeling comes from
multiple sources: (1) Using radiomics analysis technology, extract the CT image of the
patient including the chest CT value, shape and size of the lesion high-dimensional imaging
radiomics features such as texture and wavelet features to obtain more accurate and
comprehensive image data information. Yitu's existing imaging raidomics feature extraction
tool can extract up to 5,900 CT image features to make lung state more accurate assessment
and prediction It becomes possible; (2) Collecting information on the subjective evaluation
of CT image signs by imaging doctors, as well as multi-dimensional information such as basic
patient information, disease history, laboratory test results, and clinical symptoms; (3)
Based on the developed chest CT image analysis Function to extract quantitative parameters
such as pneumonia load index, patch semi-quantitative information and so on using deep
learning technology.
The total collected data set is divided into training set and internal verification set.
Firstly, the information is analyzed by traditional medicine and multi-dimensional AI
algorithm, comprehensively and quantitatively analyze whether the data column is included in
the prediction model and the weight in the model, and look for strongly related factors. Try
to use machine learning, deep learning and other AI algorithms to establish a risk prediction
model for severe COVID-19. And the results output the probability of the patient's severity,
and classify the risk of the patient's severity. Evaluate the model's ability to identify
high-risk and low-risk patients with indicators such as sensitivity, specificity, and
preliminarily verify the stability of the model.
2.2.3 Clinical application verification After the prediction model is established, the
prediction model will continue to be used in the subsequent multi-center collection of
supplementary clinical patient data, use patient follow-up data to verify its sensitivity and
specificity, and continuously incorporate the newly collected data into the model training
set to continuously improve the prediction model. Improve the application efficiency of risk
assessment models. In the end, it will reduce the conversion rate of severe patients, reduce
the management pressure of mild patients, and better assist doctors in clinical diagnosis and
treatment decision-making and patient management.
