Comparison of Two Statistical Models for Predicting Mortality in COVID-19 Patients in Iran

Abstract

Background: Today, the COVID-19 pandemic is ever-increasingly challenging healthcare systems globally with many uncertainties and ambiguities regarding disease behavior and outcome prediction. Thus, machine learning (ML) algorithms could be potentially demanding to tackle these challenges. Objectives: The present study aimed to construct and compare two prediction models based on statistical and computational ML algorithms to predict mortality in COVID-19 hospitalized patients and, finally, adopt the best-performing algorithm, accordingly. Methods: Having considered a single-center registry, we scrutinized 482 records of laboratory-confirmed COVID-19 hospitalized patients admitted from February 9, 2020, to December 20, 2020. The most important clinical parameters for COVID-19 mortality prediction were identified using the Phi coefficient technique. In the next step, two statistical and computational ML models, ie, logistic regression (LR) and artificial neural network (ANN), were evaluated through the metrics derived from the confusion matrix. Results: Predictive models were trained using 16 validated features. The results indicated that the best performance pertained to the ANN classifier with a positive predictive value (PPV) of 0.96, a negative predictive value (NPV) of 0.86, the sensitivity of 0.94, specificity of 0.94, and accuracy of 0.93. Conclusions: According to the results, ANN predicted mortality in hospitalized patients with COVID-19 with an acceptable level of accuracy. Therefore, it would be extremely reasonable to develop intelligent decision support systems to early detect high-risk patients, helping clinicians come up with proper interventions.