Abstract
The rapid spread of coronavirus 2019 disease (COVID-19) has manifested a global public health crisis, and chest CT has been proven to be a powerful tool for screening, triage, evaluation and prognosis in COVID-19 patients. However, CT is not only costly but also associated with an increased incidence of cancer, in particular for children. This study will question whether clinical symptoms and laboratory results can predict the CT outcomes for the pediatric patients with positive RT-PCR testing results in order to determine the necessity of CT for such a vulnerable group. Clinical data were collected from 244 consecutive pediatric patients (16 years of age and under) treated at Wuhan Children's Hospital with positive RT-PCR testing, and the chest CT were performed within 3 days of clinical data collection, from January 21 to March 8, 2020. Advanced decision tree based machine learning models were developed for the prediction of CT outcomes. Results have shown that age, lymphocyte, neutrophils, ferritin and C-reactive protein are the most related clinical indicators for predicting CT outcomes for pediatric patients with positive RT-PCR testing. Our decision support system has managed to achieve an AUC of 0.84 with 0.82 accuracy and 0.84 sensitivity for predicting CT outcomes. Our model can effectively predict CT outcomes, and our findings have indicated that the use of CT should be reconsidered for pediatric patients, as it may not be indispensable.
Introduction
While chest Computed Tomography (CT) is a highly sensitive tool for diagnosing and monitoring COVID-19, it poses significant risks for pediatric populations. Children are more sensitive to ionizing radiation, which is linked to an increased lifetime risk of cancer. Furthermore, studies have shown that roughly 35% of children with a positive RT-PCR test for COVID-19 actually have "normal" chest CTs, meaning many scans are medically unnecessary. This study sought to determine if easily obtainable data—like age, fever, and blood work—could serve as a reliable proxy for CT findings, creating a clinical decision support system to help doctors decide which children truly need imaging.
Methods
The researchers employed an explainable machine learning pipeline centered on Gradient Boosting Decision Trees (GBDT). The architecture, illustrated in Figure 1, involves a multi-stage process starting with tabular data processing through a deep feature extractor and ending with ensemble selection.
Key components of the methodology included:
- TF-IDF Embedding: Borrowing from Natural Language Processing (NLP), clinical feature values were treated like "words" to identify the most influential indicators.
- Target and Frequency Encoding: These methods transformed categorical data into numerical values for model processing.
- Bayesian Optimization: This was used to fine-tune the model’s hyperparameters, ensuring optimal performance on the specific dataset.
- Focal Loss: To handle the inherent imbalance in medical data, focal loss prevented the model from becoming biased toward the majority class.
Datasets
The study utilized a retrospective dataset of 244 consecutive pediatric patients (aged 16 and under) treated at Wuhan Children’s Hospital between January 21 and March 8, 2020. All patients had confirmed positive RT-PCR tests. The dataset included 32 clinical symptoms and various laboratory markers. Chest CTs were performed within three days of clinical data collection to ensure synchronization.
Key Insights and Biomarker Synergy
Out of dozens of variables, five emerged as the most critical predictors: age, C-reactive protein (CRP), neutrophils, lymphocytes, and ferritin. The researchers found a non-linear relationship between age and CT abnormalities; children under 4 were more likely to have abnormal CTs, while those over 14 were less likely.
As shown in the correlation matrix (Figure 2), there are significant relationships between these biomarkers. While single lab results provided mediocre predictive power, the combination of these markers within the GBDT model significantly boosted accuracy. Unlike "black-box" deep learning models, this decision-tree approach allowed researchers to use Odds Ratios (OR) to quantify exactly how much a specific factor increased the likelihood of a positive CT.
Key Results
The proposed model outperformed standard machine learning techniques across all major metrics:
- AUC (Area Under the Curve): 0.8412 (compared to 0.7131 for standard XGBoost).
- Accuracy: 0.8191.
- Sensitivity (Recall): 0.8597, ensuring few abnormal cases are missed.
- F1 Score: 0.8464.
An ablation study confirmed that the combination of specialized encoding and embedding techniques was responsible for a roughly 14% improvement in AUC over a baseline GBDT model.
Limitations and Future Work
Several limitations must be addressed before clinical adoption. The dataset consisted entirely of Asian pediatric patients from a single hospital, requiring validation on more diverse global populations. Additionally, while 244 patients is a solid start, larger cohorts would improve robustness. Future work will involve prospective clinical trials where the model is used in real-time to guide imaging decisions and refining the "negative" group classification to account for potential RT-PCR false negatives.
Summary updated