A machine learning based exploration of COVID-19 mortality risk

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Epidemiology

Background:

• Effective pandemic response requires early identification of patient mortality risks.
• Efficient resource allocation and timely treatment planning are crucial for reducing mortality rates.

Purpose of the Study:

• To develop and compare machine learning models for predicting patient mortality risk upon admission.
• To evaluate the efficacy of invasive, non-invasive, and combined data features for prognosis prediction.

Main Methods:

• Utilized Support Vector Machine (SVM) algorithms to build three distinct models: invasive data only, non-invasive data only, and combined data.
• Employed SVM-Recursive Feature Elimination (SVM-RFE) and sparsity analysis for feature importance assessment.

Main Results:

• Non-invasive features demonstrated mortality prediction performance comparable to invasive and combined models.
• The non-invasive model achieved superior performance with fewer features, highlighting the predictive power of SPO2, age, and cardiovascular disorders.
• Non-invasive data excelled in predicting mortality for longer-term intervals, while invasive data was better for imminent predictions.

Conclusions:

• Non-invasive machine learning models offer a viable and efficient approach for early mortality risk prediction in pandemics.
• These models, particularly when integrated with wearable technology, can significantly enhance patient triage and intervention strategies.