Artificial neural networks to predict COVID-19 progression of moderate hospitalized patients using early clinicalparameters and blood tests

ABSTRACT

Background: The early prediction of Covid-19 progression could improve patient's treatment. It is important to develop mathematical models to perform this task using simple blood tests. Aims: To obtain a neural network (ANN) to predict the progression (death vs discharge and intubation vs discharge) of Covid-19 in patients with confirmed diagnosis. Methods: The patients included in this work were diagnosed with Covid-19 by RT-PCR. All data were collected from hospitalized patients admitted to Anhembi Field Municipal Hospital (São Paulo-Brazil), a hospital set up for initial care to patients with moderate symptoms during the pandemic, between June/2020 and October/2020. Blood was collected at the patient's admission. The inputs considered were sex, age, ethinicity, body mass index, tabagism, ex-tabagism, alveolar infiltrate, arterial hypertension, diabetes, heart rate, respiration rate, body temperature, oxygen saturation, D-dimer, activated partial thromboplastin time, prothrombin time, levels of hemoglobin, platelet, leukocytes, lymphocytes, monocytes, neutrophils, lactate dehydrogenase, C-reactive protein, and creatinine. Two ANNs were proposed, as shown at Table 1. The best ANN was defined by a 5-fold cross-validation scheme. Finally, a test step was performed to verify the ANN performance. ANNs with one and two hidden layers were tested. The number of neurons ranged from 5 to 35. Results: The main results are shown at Table 2. The best models were obtained with different ANN's structures, which show the influence of the different outcome. The models presented high ACC, AUC, PPV, NPV, and TNR. The ANN 2 presented better performance than ANN 1. We believe that this may be due the data homogeneity that rises from the inclusion criteria adopted in the study. Conclusions: The results showed that the ANNs could be used to predict the progression of Covid-19 based on simple blood tests. The models could be used in the future after an external validation with high number of patients.