A machine learning PROGRAM to identify COVID-19 and other diseases from hematology data

Area of Science:

• Hematology
• Machine Learning
• Medical Informatics

Background:

• Hematology data offers a rich source of information for disease detection.
• Current screening methods may not fully leverage the potential of high-dimensional hematology metadata.
• Personalized medicine approaches require robust analytical tools for individual health assessment.

Purpose of the Study:

• To develop and validate a machine learning (ML) method for screening communicable and noncommunicable diseases using full blood count (FBC) metadata.
• To demonstrate the potential of ML in identifying disease patterns within hematology data.
• To assess the generalizability of the proposed ML method across different disease types.

Main Methods:

• Extraction of high-dimensional hematology metadata from Sysmex analyzers over 11 months (43,761 patients).
• Development of predictive models for age, sex, and patient individuality to establish data personalization.
• Application of supervised and unsupervised ML techniques, utilizing both numeric and flow cytometry data, to predict pneumonia, urinary tract infection, COVID-19, and heart failure.

Main Results:

• Models accurately predicted chronological age (R²: 0.59), sex (AUROC: 0.83), and patient individuality (99.7% accuracy).
• Significant predictive performance was achieved for infectious diseases: pneumonia (AUROC: 0.74), urinary tract infection (AUROC: 0.68), and COVID-19 (AUROC: 0.8).
• The method demonstrated generalizability by predicting heart failure with an AUROC of 0.78.

Conclusions:

• Machine learning applied to hematology data is a viable method for predicting both communicable and noncommunicable diseases.
• The findings suggest the potential for widespread application of this ML approach in disease screening at local and global scales.
• Hematology metadata, analyzed through ML, can provide valuable insights for early disease detection and patient management.