COVID-19 is a disease caused by SARS CoV-2 virus that has recently gained attention due to the morbidity, mortality, lifestyle changes, economic changes and psychological issues it has caused in the general population. Being an RNA virus, SARS CoV-2 mutates frequently, thereby evading specific immune responses and necessitating frequent vaccination. It is because of these mutations that we have not been able to develop specific drugs targeting the virus. As a result, management of symptoms and associated complications has been the mainstay of treating COVID-19 so far. Hence, early diagnosis and prediction of complications is critical in reducing COVID-19 related morbidity and mortality.
The Real-Time Reverse Transcriptase Polymerase Chain Reaction (RT-PCR) test is the current gold standard for diagnosis of COVID-19. However, limited testing infrastructure and long turnaround times for RT-PCR are important limiting factors. Medical imaging is a feasible way of aiding diagnosis and predicting the severity of the disease. Computed Tomography (CT) of the chest, though highly specific and sensitive, has limited availability at primary health centres. High patient volumes and portability are other factors that affect its usage at primary health centres. Because of its low cost, wider breadth of coverage, and portability, the chest X-ray has the potential to be a first line triage tool in resource-poor settings.
As the recent COVID-19 waves exposed the gaps in India’s diagnostic infrastructure, the concept of deep learning based artificial intelligence (AI) model was proposed. It was decided to develop an AI model that predicted the severity of the disease using chest X-rays, and relevant clinical data. The objectives of the study were to analyse clinical parameters and radiological findings in relation to patient outcome, as well as to evaluate a prognosis model that takes X-ray and clinical data as input and predicts whether the patient is at high risk (mortality) or low risk.
Figure 1: Chest X-ray images indicating subtle haziness in the bilateral lung base in COVID-19 pneumonia, demarcated by an AI model using (a) heat map and outline view, and (b) bounding boxes and outline view.
Data from patients who had a positive RT-PCR test and at least one X-ray imaging post-admission were used in the retrospective study. The data used for AI modeling included radiological findings in Chest X-rays (CXRs), clinical findings, symptoms (coughing, difficulty in breathing, vomiting, and diarrhoea), comorbidities, number of days between the onset of symptoms and the date of the X-ray. The patients’ demographic characteristics included their age and gender. Hypertension, diabetes mellitus, asthma, hypothyroidism and other conditions, were among the comorbidities.
The data was pre-processed (data cleaning and dimensionality reduction) before training a random forest model. The hyperparameter grid search optimization method was used to find the right parameters for optimal functioning of the model. Two features were added from the X-ray detection model: probability of prediction and area of the predicted mask in relation to the image size. Both features were normalized within the range [0,100] and a boolean output was retained, with discharge assigned 0 and succumbing to the infection assigned 1. Relevant demographic characteristics, symptoms, comorbidities and other clinically relevant data were also converted to features. The presence or absence of a feature was represented in a binary format (0/1).
A retrospective analysis of clinical data of 201 patients from a tertiary care centre in Pune was performed. The clinical data based model produced a sensitivity of 0.83 and a specificity of 0.61. When chest X-rays were also included in addition to the clinical parameters, the specificity of the model was increased to 0.7. Moreover, the area under the receiver operating characteristic (AUC-ROC) curve was also increased from 0.74 to 0.79. This infers that X-rays are an important parameter in determining the severity of COVID-19.
Figure 1: Receiver operating characteristic (ROC) plots for COVID-19 mortality prediction with and without X-Rays as a parameter.
The features used in AI decision making were ranked according to their importance. The age of the patient was among the top predictors of mortality, followed by AI prediction of CXRs, comorbidities, number of days between the date of scan and the date of onset of symptoms, clinical condition of the patient at the time of admission, and symptoms such as shortness of breath (dyspnea), vomiting, cough, and diarrhoea.
Figure 2: Random forest model ranking top 10 clinical parameters for predicting COVID-19 associated mortality.
As the pandemic progresses, there is a possibility of the spread of COVID-19 to the rural areas of India. Because X-ray imaging is the primary mode of radiological investigation in these areas, the likelihood of its use for diagnostic purposes is high. Our model demonstrates that the severity of lung involvement as revealed by the X-ray is a valuable prognostic factor that can be used to triage patients, and provide adequate care and vigilance for those at a higher risk of mortality. It can also assist the physician in modifying the treatment based on the predicted needs of the patient. This could be valuable for triage and proactive management for high risk patients. The full article is available at: https://www.medrxiv.org/content/10.1101/2021.09.22.21263956v1