Source-linked AI summary
COVID-19 diagnosis by routine blood tests using machine learning
Matjaž Kukar, Gregor Gunčar, Tomaž Vovko, Simon Podnar, Peter Černelč, Miran Brvar, Mateja Zalaznik, Mateja Notar, Sašo Moškon, Marko Notar
TL;DR
The study addresses COVID-19 diagnosis from routine blood tests, which may reveal data hidden from clinicians and support early diagnosis. It develops an ML pipeline and reports strong cross-validated discrimination, while identifying important scope limitations.
Problem
Routine blood-test data may contain COVID-19 signals hidden from experienced clinicians and could help when RT-PCR misses early symptomatic infection.
Method
The study uses the Smart Blood Analytics CRISP-DM pipeline to train and evaluate an XGBoost COVID-19 diagnostic model from routine blood tests.
Results
81.9% sensitivity, 97.9% specificity, and AUC 0.97 were achieved on the cross-validated training group.
Takeaways & Limitations
The model is presented as a widely available diagnostic aid that is complementary to standard protocols and can support routine blood-test assessment.
Takeaways & Limitations
The study was single-center and retrospective, included only 160 COVID-19-positive patients, and lacked data on presymptomatic diagnosis.
Abstract
from arXiv · showhide
Physicians taking care of patients with coronavirus disease (COVID-19) have described different changes in routine blood parameters. However, these changes, hinder them from performing COVID-19 diagnosis. We constructed a machine learning predictive model for COVID-19 diagnosis. The model was based and cross-validated on the routine blood tests of 5,333 patients with various bacterial and viral infections, and 160 COVID-19-positive patients. We selected operational ROC point at a sensitivity of 81.9% and specificity of 97.9%. The cross-validated area under the curve (AUC) was 0.97. The five most useful routine blood parameters for COVID19 diagnosis according to the feature importance scoring of the XGBoost algorithm were MCHC, eosinophil count, albumin, INR, and prothrombin activity percentage. tSNE visualization showed that the blood parameters of the patients with severe COVID-19 course are more like the parameters of bacterial than viral infection. The reported diagnostic accuracy is at least comparable and probably complementary to RT-PCR and chest CT studies. Patients with fever, cough, myalgia, and other symptoms can now have initial routine blood tests assessed by our diagnostic tool. All patients with a positive COVID-19 prediction would then undergo standard RT-PCR studies to confirm the diagnosis. We believe that our results present a significant contribution to improvements in COVID-19 diagnosis.
Materials and methods
The study built and evaluated a COVID-19 diagnostic model from routine blood tests, using retrospective COVID-19-positive and infectious-disease control populations. The workflow included blood-parameter selection, t-SNE visualization, machine-learning modeling, calibration, and cross-validation.
- Blood parameters used for model building: 35 blood parameters were selected after excluding measurements available in fewer than 25% of COVID-19-positive patients and removing non-blood and arterial-blood parameters.Parameter values were centered and scaled using reference ranges and group medians.
- Visualization of blood parameter space: t-SNE was applied to visualize the arrangement of patients in the high-dimensional space of 35 blood parameters.The method was used for exploratory visualization, while the authors cautioned that plot distances can be misinterpreted.
- Smart Blood Analytics machine learning algorithm: The Smart Blood Analytics pipeline acquired, filtered, preprocessed, modeled, evaluated, and deployed blood-test data using machine-learning algorithms including XGBoost.XGBoost was selected for COVID-19 diagnosis after initial experiments with random forests and deep neural networks.
- Evaluation of predictive models: Repeated stratified ten-fold cross-validation and separate control-group testing evaluated sensitivity, specificity, precision, AUC, and ROC performance.The operational ROC point was calibrated, and model evaluation used standard classification measures.
Results
The XGBoost model identified distinctive routine blood-parameter patterns and accurately discriminated COVID-19 from other infections in cross-validated evaluation. Severe COVID-19 cases occupied blood-parameter space closer to bacterial infections than to other viral infections.
- Among 160 COVID-19-positive patients, 17 required intensive care, 14 required invasive ventilation, 44 developed respiratory failure, and 10 died.
- MCHC, eosinophil count, albumin, INR, and prothrombin activity % had the highest XGBoost feature-importance scores.
- COVID-19 patients differed significantly from bacterial and other viral infections across multiple blood parameters, including urea, hemoglobin, erythrocyte count, hematocrit, leukocyte count, and MCHC-related measures.
- The COVID-19 group overlapped with both bacterial and viral infections in t-SNE space, while severe cases and those with acute respiratory failure had medoids closer to bacterial infections.
- Sensitivity was 81.9% ± 6, specificity was 97.9% ± 0.4%, and AUC was 0.97 in ten-fold stratified cross-validation.The operational ROC point was selected at sensitivity = 0.819 and specificity = 0.979.
Discussion
The model achieved strong cross-validated discrimination of COVID-19 from other infectious diseases using routine blood tests, while offering potentially complementary diagnostic information. Its interpretation is constrained by limited presymptomatic evidence, single-center data, and only 160 COVID-19-positive patients.
- 81.9% sensitivity, 97.9% specificity, and 0.97 AUC were achieved on the cross-validated training group.The authors characterize the low false-positive proportion and AUC above 0.90 as quantitatively excellent.
- The model’s performance was reported as probably complementary to RT-PCR and chest CT, whose sensitivities in a cited study were 59% and 88%, respectively.The authors recommend use alongside standard protocols rather than as an isolated replacement.
- The most discriminating parameters included prothrombin activity percentage, INR, MCHC, eosinophils, and albumin, with coagulation-related findings indicating accelerated clot formation.The analysis also observed raised CRP and MCHC, reduced eosinophils, low albumin, and lymphopenia.
- The t-SNE visualization shifted severe-COVID-19 patients toward the bacterial-infection medoid, supporting early testing for bacterial infection or severe inflammation.The authors also identify this pattern as a possible basis for prognostication of COVID-19 course using machine learning.
- Generalizability is limited by the single-center retrospective design and 160 COVID-19-positive patients; the authors state that at least 1,000 positives could improve results.The model’s ability to diagnose presymptomatic COVID-19 was not assessed and may be inefficient before systemic effects occur.