Potential limitations in COVID-19 machine learning due to data source variability: A case study in the nCov2019...

Area of Science:

• Machine Learning
• Epidemiology
• Data Science

Background:

• Reliable machine learning for coronavirus disease 2019 (COVID-19) is hampered by a lack of representative data.
• Insufficient data sharing and variable data quality, particularly source variability, are significant bottlenecks.
• Understanding data source variability is critical for mitigating bias in COVID-19 research.

Purpose of the Study:

• To investigate potential biases introduced by data source variability in COVID-19 machine learning.
• To explore the discovery and classification of COVID-19 severity subgroups using patient symptoms and comorbidities.
• To highlight the impact of data source variability on model generalizability.

Main Methods:

• Utilized the publicly available nCov2019 dataset containing patient-level data from multiple countries.
• Analyzed cases from countries with the highest prevalence, dividing them into subgroups based on distinct severity manifestations.
• Focused on symptom and comorbidity data for subgroup discovery and classification.

Main Results:

• Identified distinct severity subgroups within COVID-19 cases from different countries.
• Demonstrated that data source variability can reduce the representativeness of training data.
• Highlighted the risk of increased model complexity and overfitting due to data variability.

Conclusions:

• Data source variability is a key factor contributing to bias in distributed research networks for COVID-19.
• Systematic assessment and reporting of data source variability and quality are essential for COVID-19 data sharing.
• Improving data quality is paramount for developing reliable and generalizable machine learning models in COVID-19 research.