Source-linked AI summary
Use of Machine Learning and Artificial Intelligence to predict SARS-CoV-2 infection from Full Blood Counts in a population
Abhirup Banerjee, Surajit Ray, Bart Vorselaars, Joanne Kitson, Michail Mamalakis, Simonne Weeks, Mark Baker, Louise S. Mackenzie
TL;DR
The study asks whether SARS-CoV-2 infection can be screened early from full blood counts despite overlapping symptoms and limited access to rt-PCR testing. Using anonymized patient data, it compares machine-learning, neural-network, and simple statistical models. The models achieved AUC values up to 95% in regular-ward patients and 86% in community patients, while a four-count formula reached 85% AUC in the community.
Problem
Overlapping symptoms and limited access to rt-PCR testing motivate screening for SARS-CoV-2 using accessible clinical data.
Method
The study trains random forest, shallow-learning, artificial neural-network, and simple statistical models using anonymized, normalized full blood counts without symptoms or patient history.
Results
A four-count formula achieved 85% AUC for early-stage community patients, while the broader models achieved up to 86% AUC in community patients and 95% in regular-ward patients.
Takeaways & Limitations
Full blood count profiles provided an initial screen distinguishing SARS-CoV-2-positive from negative patients, including patients later diagnosed with other pathogens.
Takeaways & Limitations
Further validation is required to determine whether the model can fully distinguish SARS-CoV-2 from other pathogens, and ICU patients were excluded.
Abstract
from arXiv · showhide
Since December 2019 the novel coronavirus SARS-CoV-2 has been identified as the cause of the pandemic COVID-19. Early symptoms overlap with other common conditions such as common cold and Influenza, making early screening and diagnosis are crucial goals for health practitioners. The aim of the study was to use machine learning (ML), an artificial neural network (ANN) and a simple statistical test to identify SARS-CoV-2 positive patients from full blood counts without knowledge of symptoms or history of the individuals. The dataset included in the analysis and training contains anonymized full blood counts results from patients seen at the Hospital Israelita Albert Einstein, at São Paulo, Brazil, and who had samples collected to perform the SARS-CoV-2 rt-PCR test during a visit to the hospital. Patient data was anonymised by the hospital, clinical data was standardized to have a mean of zero and a unit standard deviation. This data was made public with the aim to allow researchers to develop ways to enable the hospital to rapidly predict and potentially identify SARS-CoV-2 positive patients. We find that with full blood counts random forest, shallow learning and a flexible ANN model predict SARS-CoV-2 patients with high accuracy between populations on regular wards (AUC = 94-95%) and those not admitted to hospital or in the community (AUC=80-86%). Here, AUC is the Area Under the receiver operating characteristics Curve and a measure for model performance. Moreover, a simple linear combination of 4 blood counts can be used to have an AUC of 85% for patients within the community. The normalised data of different blood parameters from SARS-CoV-2 positive patients exhibit a decrease in platelets, leukocytes, eosinophils, basophils and lymphocytes, and an increase in monocytes.
1. Introduction
The study addresses the need for early SARS-CoV-2 screening when symptoms overlap with common infections and access to rt-PCR testing is limited. It evaluates whether full blood counts can identify infection without symptoms or patient history.
- SARS-CoV-2 symptoms can be difficult to distinguish from other common infections, motivating early screening to identify, isolate, and track infected patients.
- Wide-scale rt-PCR testing is constrained by resources, and the standard test has reported 80% accuracy compared with chest CT results.
- Specialized CRISPR and biosensor tests require equipment and resources that may be unavailable in less affluent areas.
- The study predicts SARS-CoV-2-positive or negative status from full blood counts using machine learning and artificial neural networks.
- The analysis uses anonymized full blood counts and rt-PCR results from 598 patients at Hospital Israelita Albert Einstein in São Paulo, without released symptoms or patient history.
2. Methods
The study analyzes anonymized, standardized full blood counts and compares several classification models for SARS-CoV-2 status. It evaluates performance with metrics suited to imbalanced community data.
- The dataset contains anonymized patient data standardized to mean zero and unit standard deviation, including full blood counts and rt-PCR outcomes.
- Of 5,644 released records, 598 complete full blood count results were analyzed, while 5,046 lacked full blood count data.
- Patients in semi-intensive and intensive care were excluded, and age and neutrophils were omitted from modeling.
- Community data contained 8% SARS-CoV-2-positive and 92% negative patients, so sensitivity and specificity were assessed alongside accuracy.
- The models included random forest, regularized logistic regression using glmnet, and an artificial neural network evaluated with stratified 10-fold cross-validation.
- Performance was measured using AUC, sensitivity, specificity, and accuracy because accuracy alone can be misleading in an imbalanced dataset.
3. Results
Full blood counts showed SARS-CoV-2-associated immune-cell differences, and multiple models discriminated positive from negative patients in regular-ward and community groups. A four-count linear formula also provided community-level prediction.
- Significant differences were found in 9 of 15 blood count parameters among regular-ward patients testing positive or negative for SARS-CoV-2.
- 90% average accuracy and AUC 0.95 ± 0.08 were achieved by the ANN for regular-ward patients over stratified 10-fold cross-validation.
- The community ANN achieved 89% average accuracy, with AUC 0.77 ± 0.08 before SMOTE and AUC 0.80 ± 0.05 after class balancing.
- Random forest and glmnet produced average AUC 94% in regular-ward patients and 84–86% in community patients.
- Community SARS-CoV-2-positive patients showed higher monocytes and lower leukocytes, while regular-ward patients showed decreased leukocytes and eosinophils.
- A normalized four-count formula adding monocytes and subtracting leukocytes, eosinophils, and platelets achieved AUC 85% in community patients and AUC 81% in regular-ward patients.
- Among tested patients, 188 of 366 tested for other pathogens were diagnosed with other infections, and only one patient was positive for SARS-CoV-2 and another pathogen.
- The study also observed changes in red blood cells and platelets in regular-ward patients.
4. Discussion
The study developed blood-count-based models that distinguish SARS-CoV-2-positive patients across community and regular-ward settings, while identifying characteristic immune-cell and platelet changes. A four-count additive formula also provided an early community-screening approach, although normalized data complicate biological interpretation.
- Model performance: AUC up to 86% in community patients and 95% in regular-ward patients was achieved using normalized full blood counts with multiple independent models.The models included statistical, random forest, and shallow-learning approaches.
- Model performance: The models distinguished SARS-CoV-2-positive patients from negative patients using altered blood profiles, including cases later diagnosed with other pathogens.The analysis focused on early disease stages and excluded ICU patients.
- Simple screening formula: 85% AUC was obtained for early-stage community patients using normalized monocytes - leukocytes - eosinophils - platelets.The authors describe this as a simple calculation requiring further validation against other pathogens.
- Biological interpretation: The study links monocyte involvement to a possible SARS-CoV-2-associated immune response and emphasizes monocytes as important for prediction.The discussion connects this interpretation with prior observations involving monocytes and IL6.
- Biological interpretation: Decreased platelets may distinguish SARS-CoV-2 from Influenza A, but normalized measurements can mislead interpretation of platelet count and size.The discussion notes apparently conflicting evidence involving thrombosis and platelet production.
Funding
The research received no specific grant funding from public, commercial, or not-for-profit funding agencies.
- No specific grant supported the research.
- The study received no funding from public-sector agencies.
- The study received no funding from commercial or not-for-profit sectors.