Early Stage Identification of COVID-19 Patients in Mexico Using Machine Learning: A Case Study for the Tijuana General Hospital

Abstract

Background: The current pandemic caused by SARS-CoV-2 is an acute illness of global concern. SARS-CoV-2 is an infectious disease caused by a recently discovered coronavirus. Most people who get sick from COVID-19 experience either mild, moderate, or severe symptoms. In order to help make quick decisions regarding treatment and isolation needs, it is useful to determine which significant variables indicate infection cases in the population served by the Tijuana General Hospital (Hospital General de Tijuana). An Artificial Intelligence (Machine Learning) mathematical model was developed in order to identify early-stage significant variables in COVID-19 patients. Methods: The individual characteristics of the study subjects included age, gender, age group, symptoms, comorbidities, diagnosis, and outcomes. A mathematical model that uses supervised learning algorithms, allowing the identification of the significant variables that predict the diagnosis of COVID-19 with high precision, was developed. Results: Automatic algorithms were used to analyze the data: for Systolic Arterial Hypertension (SAH), the Logistic Regression algorithm showed results of 91.0% in area under ROC (AUC), 80% accuracy (CA), 80% F1 and 80% Recall, and 80.1% precision for the selected variables, while for Diabetes Mellitus (DM) with the Logistic Regression algorithm it obtained 91.2% AUC, 89.2% accuracy, 88.8% F1, 89.7% precision, and 89.2% recall for the selected variables. The neural network algorithm showed better results for patients with Obesity, obtaining 83.4% AUC, 91.4% accuracy, 89.9% F1, 90.6% precision, and 91.4% recall. Conclusions: Statistical analyses revealed that the significant predictive symptoms in patients with SAH, DM, and Obesity were more substantial in fatigue and myalgias/arthralgias. In contrast, the third dominant symptom in people with SAH and DM was odynophagia.

1. Introduction

A novel coronavirus, known as Severe Acute Respiratory Syndrome (SARS-CoV-2), was identified in December 2019 as the cause of a respiratory illness called Coronavirus Disease 2019, or COVID-19 [1]. The origin of this virus is not yet confirmed, but an analysis of its genetic sequence suggests it is phylogenetically related to bat viruses similar to SARS (severe acute respiratory syndrome), making bats a possible key reservoir [2]. Symptoms of COVID-19 infection appear after an incubation period of approximately 5.2 days [3]. The period from the onset of COVID-19 symptoms to death ranges from 6 to 41 days with a median of 14 days [4]. This period depends largely on the age and the state of the patient’s immune system [4].

The infection is transmitted through droplets generated by symptomatic patients when coughing or sneezing, but it can also occur through asymptomatic patients and even before the onset of symptoms [5].

The clinical features of COVID-19 are diverse, from an asymptomatic state to acute respiratory distress syndrome and multiorgan failure [5]. The most common early symptoms of COVID-19 illness are fever, cough, and fatigue, while other symptoms include headache, sputum production, hemoptysis, dyspnea, diarrhea, and lymphopenia [6]. Advanced age, cardiovascular disease, diabetes, chronic respiratory disease, hypertension, and cancer are said to increase the risk of death for people diagnosed with COVID-19.

Regarding COVID-19, as of 15 August 2021, there were 207,784,507 confirmed cases (410,464 new cases) and 4,370,424 deaths, while 4,462,336,040 vaccine doses have been reported worldwide [7,8]. Most estimates of fatality ratios have been based on cases detected through surveillance and calculated using crude methods, giving rise to widely variable estimates of CFR depending on the country—from less than 0.1% to over 25% [9].

Currently, in Mexico (August 2021), there are 3,310,989 estimated positives, with 261,384 estimated deaths, and 133,866 estimated actives. However, there are 3,108,438 confirmed cases, 5,527,343 negative, 477,811 suspected and 248,652 accumulated deaths. Of the confirmed cases, 53.56% have been women and 46.44% men. A total 5.6% of patients have been hospitalized, and 94.4% have been outpatients. Among the main comorbidities are hypertension (10.34%), Obesity (8.989%), diabetes (7.31%) and smoking (8.03%), with information updated on 16 August 2021. On the same date, the state of Baja California had 913 active cases, with 54,453 accumulated cases, and 8979 deaths. The state’s capital, Mexicali, is the city with the highest number of cases in the state, with 21,778 accumulated cases, followed by Tijuana, our case study [10].

2. Background

Several authors have addressed the issue of SARS-CoV-2 from a technological point of view, with the development of artificial intelligence algorithms. There are models to predict the mortality rate [11]. Some studies present the detection of severely ill patients with COVID-19 from those with mild symptoms using clinical information and data from blood and urine tests [12]. Several artificial intelligence models have been used with Machine Learning and Deep Learning methods which have been used intensively for COVID-19. Although Machine Learning and Deep Learning methods show successful results in the COVID-19 cases tested, there are accounting challenges that can be considered to improve the quality of the research in that direction [13].

Table 1. Minireview of papers.

Authors.	Year	Objective	Learners	Metrics	Novelties
Li WT et al. [14].	2020	Classification	XGBoost	sensitivity of 92.5% and a specificity of 97.9%	Novel associations between clinical variables, including correlations between being male and having higher levels of serum lymphocytes and neutrophils. We found that COVID-19 patients could be clustered into subtypes based on serum levels of immune cells, gender, and reported symptoms.
Guan X et al. [15]	2020	Prediction	XGBoost	>90% precision and >85% sensitivity, as well as F1 scores >0.90	Proposed disease severity, age, serum levels of hs-CRP, LDH, ferritin, and IL-10 as significant predictors for death risk of COVID-19, which may help to identify the high-risk COVID-19 cases.
Delafiori J et al. [16]	2021	Diagnosis and risk assesment	gradient tree boosting (GDB) ADA tree boosting	96.0% of specificity and 83.1% of sensitivity 80.3% of specificity and 85.4% of sensitivity	Propose machine learning techniques to determine from databases the five main challenges in responding COVID-19 and how to overcome these challenges to save lives.
Allam M et al. [17]	2020	Diagnosis and prediction	Neural Networks	100% sensitivity and 99.9% specificity	The Abbott antibody test (SARS-CoV-2 IgG assay) has shown 100% sensitivity and 99.9% specificity thus far. The Abbott test finds whether the patient has IgG antibodies for COVID-19, which can stay for months to years after a person has recovered.
Assaf D et al. [18]	2020	Prediction	Classification and Regression Tree (CRT) model	Sensitivity, specificity, PPV, NPV and accuracy of 88.0%, 92.7%, 68.8%, 97.7% and 92.0%, respectively, with ROC AUC of 0.90.	The analysis of the database in this study found that most contributory variables to the models were APACHE II score, white blood cell count, time from symptoms to admission, oxygen saturation and blood lymphocytes count. Machine-learning models demonstrated high efficacy in predicting critical COVID-19 compared to the most efficacious tools available.
Naseem M et al. [19]	2020	Detection	Neural Networks	sensitivity of 90% and specificity of 96% respectively	Results were synthesized and reported under 4 themes. (a) The need of AI during this pandemic: AI can assist to increase the speed and accuracy of identification of cases and through data mining to deal with the health crisis efficiently, (b) Utility of AI in COVID-19 screening, contact tracing, and diagnosis: Efficacy for virus detection can a be increased by deploying the smart city data network using terminal tracking system along-with prediction of future outbreaks, (c) Use of AI in COVID-19 patient monitoring and drug development:
Arga KY [20]	2020	Prediction and diagnosis	Apache, Gleason and PASI	-	Machine learning is considered to help reduce diagnostic errors and unnecessary use of diagnostic tools through the development of rational algorithms. Indeed, the COVID-19 pandemic showed that digital health is invaluable, feasible, and not too far.
Majhi R et al. [21]	2020	Prediction	Nonlinear Regression (NLR), Decision Tree (DT) based regression, and random forest (RF) models	Evaluation metrics obtained using the Mean Absolute Percentage Error (MAPE). NLR = 0.24% DT = 0.18% RF = 0.02%	The algorithm predict the number of positive cases in India. In essence, the paper proposes a machine learning model that can predict the number of cases well in advance very effectively and also suggest some key inputs.
van der Schaar M et al. [22]	2020	Prediction	-	.	This paper summarizes the use of machine learning techniques in different studies to manage limited healthcare results, developing personalized treatment, informing policies and able effective collaboration and expediting clinical trials.
Das AK et al. [23]	2020	Prediction	Linear Regression	For liner regression (area under ROC curve = 0.830), calibration (Matthews Correlation Coefficient = 0.433; Brier Score = 0.036).	In this study, according to the random forest algorithm, age was the most important predictor followed by exposure, sex and province, whereas this order was sex, age, province and exposure as per logistic regression
Swapnarekha H et al. [24]	2020	Prognosis	Support Vector Machine (SVM), Rannom Forest RF, K-Means, XGBoost and linear regression	0.933 true positive rate, 0.74 true negative rate and 0.875 accuracy.	This article obtained good metrics for COVID-19 prediction. On the other hand, mentioned machined learning techniques used for classification and prediction to reduce the spread of coronavirus and understand the limitation of machine learning analysis, being: lack of information, accuracy of predictions, usage of advanced approaches, providing feasible solutions for developing countries and necessity of advance intelligent systems on symptom based identification of COVID-19.

summarizes the articles included in this minireview, together with some characteristics of these studies.

3. Materials and Methods

This article is based on a study of COVID-19 patients at the Tijuana General Hospital, a public hospital that serves a very particular low-income population. Tijuana is a border city in northern Mexico next to San Diego, California, in the United States. Including neighboring Rosarito, the greater Tijuana region has a population of around 1,900,000 inhabitants, the majority Mexican nationals. However, there are also many migrants from other Central American, South American, and Caribbean countries living in unaccounted shantytowns, seeking to enter the United States in any way, either legal or illegally. While doing so, they temporarily live in Tijuana without a permanent job, a fixed salary, a steady place of work, or a regular postal address, and therefore, do not have access to social security or health services, and eventually fall ill, often due to a myriad of causes. The range of pathologies and diseases found in the city of Tijuana is much broader than those found in other cities with more homogeneous or steady populations, which further complicates medical and healthcare services for this specific segment of people. Tijuana General Hospital is one of the few public health institutions that serve this marginalized segment of the Tijuana population, hence the size and complexity of the challenges faced by its medical staff every day, as well as the diversity of pathologies met by the health professionals who treat them. In this article, we evaluate a group of COVID-19-diagnosed patients, who were treated at this Tijuana hospital during 2020.

3.1. Sample Size

The required sample size for this study was estimated considering the expected prevalence in studies carried out using bioimpedance analysis, 17% [25], assuming a margin of error of 5% and a confidence interval of 95%. According to these criteria, a total of 185 patients were needed in order to obtain the desired results. The average age of the studied population was 55 years, while the average hospital length of stay duration was six days.

Patients arrived mainly from the Tijuana and Rosarito urban and suburban areas. This research included patient medical history, pharmacology, PCR testing and biochemical data.

3.2. Database

Information on 185 patients with 99 variables was collected for each of them to create the database, the description of the variables can be found in Appendix A. Table 1 shows the gender criteria for evaluating patients at the Tijuana General Hospital, which serves a low-income population in the Tijuana and Rosarito areas of Baja California, with a higher percentage of men than women as seen from the

Table 2. Evaluation Criteria.

Gender	%	Kg-m2	SpO
Women	39.46%	<6.1 kg/m2	>95%
Men	60.54%	<8.5 kg/m2	>95%

.

3.3. Bedford’s Law

Benford’s Law validation method was used in order to make sure the data was consistent, in order to develop an efficient study. Benford’s Law, or the Law of First Digits, is a tool used in different fields of science, with a method to suggest a mathematical pattern in the distribution of the first digits in a dataset that does not display a uniform distribution, but rather are arranged in such a way that the digit “1” is the most frequent, followed by “2”, then “3”, and so on, down to “9”. This model suggests that, within a random set of data, the first digit of approximately 30.10% of the numbers will be “1”. Several studies have used this technique to validate and evaluate veracity in databases with information about COVID-19 [26,27].

By using Benford’s Law as a validation method, it was demonstrated that there is consistency in the data collected from the Tijuana General Hospital, as the curve of our current information is close to the curve generated by the percentages established by Benford’s Law. For this comparison, “length of stay” data were used, as within the database that was used, most information is described by binary numbers (0 and 1), while the variable “length of stay” is a defined variable. For this, in addition to the graph function, two Excel functions were used:

• Left: This function allowed the first digit to be taken to the left of the number within the “length of stay” column.
• Countif: This function allows to count the frequency of each of the digit numbers without considering 0.

Figure 1 shows the comparison results between Bedford’s Law and “length of stay” which suggest that the data is consistent.

The distribution by first digit numbers in the data can be seen in

Table 3. Data compared with Bedford’s Law.

Benford’s Law	Actual	R.E. *
	Tijuana cases 1
30.10%	25.57%	−15.01%
17.61%	14.20%	−19.30%
12.49%	14.00%	9.12%
9.69%	12.50%	22.89%
7.92%	10.23%	29.16%
6.69%	7.95%	18.83%
5.80%	6.25%	7.75%
5.12%	6.82%	33.20%
4.58%	2.84%	1.73%

¹ Considering “length of stay” for validation. * Relative Error.

. Invalidating the data, the comparison suggests that according to Benford’s law, the results obtained from this analysis for the Tijuana Hospital database are accurate.

3.4. Machine Learning Analysis

For the evaluation of each dataset, the following classifiers were used: Decision Tree (DT), Support Vector Machine (SVM), Random Forest (RF), Multi-Layer Perceptron Neural Network (MLPNN), Naive Bayes (NB), Logistic Regression (LR) and AdaBoost (AB). The configuration for each algorithm were suggested by Orange. These configurations are shown in

Table 4. Configurations used for each learner in the analysis.

Learner	Configuration for Learners
RF	A number of trees: 10, minimum subsets split: 5, maximum tree depth: unlimited.
kNN	Number of neighbors: 3, metric: Euclidean, weight: uniform
SVM	Type: SVM Regression, C = 1, ε = 0.1, Kernel= Radial Basis Function (RBF), exp (−auto|x − y|2), numerical tolerance: 0.001
MLPNN	Hidden layers: 100, activation: ReLu, solver: Adam, alpha: 0.0001, maximum iterations: 200, replicable training: True.
NB
AB	Base estimator: tree, number of estimators: 50, algorithm (classification): Samme. r, loss (regression): Linear
DT	Type: binary tree, internal nodes < 5, maximum depth: 100, splitting: 95%.

. The algorithms were evaluated by stratified k-fold cross-validation, where the data was iteratively examined by ten folds, using nine folds for training and 1-fold for testing; with this method, the data can be divided into equal parts, which gives better results, avoiding generalization which means possible errors when new data is used for predictions with the trained models.

The datasets for each analysis were created by feature selection; for this, HAS, DM, and Obesity were selected as targets due to their importance in the development of COVID-19. After selecting the targets, the complete database was analyzed iteratively with each of the chosen targets. The process of selecting the most relevant features was done by using the DT and comparing the results with the rank, using the following scoring methods: gain ratio, Gini, X², ReliefF and Fast Correlation Based Filter (FCBF). The algorithm followed for the scoring techniques is shown in

Table 5. Rank Scoring Algorithms.

Method	Algorithm
Gain Ratio	$IG\left(Ex,a\right)=H\left(Ex\right)-{\displaystyle {\displaystyle \sum }_{v \in values \left(a\right)}}\left(\frac{\left|\left\{x\in Ex|value\left(x,a\right)=v\right\}\right|}{\left|Ex\right|}\ast H\left(\left\{x\in Ex|value\left(x,a\right)=v\right\}\right)\right)$
Gini	$G=\frac{{{\displaystyle \sum }}_{i=1}^n\ast {{\displaystyle \sum }}_{j=1}^n\ast \left|x_i-x_j\right|}{2{{\displaystyle \sum }}_{i=1}^n\ast {{\displaystyle \sum }}_{j=1}^n\ast x_j}$
X2	$X=\frac{m-Np}{\sqrt{Npq}}$
ReliefF	$W_i=W_i-{(x_i-nearHi{t}_i)}^2+{(x_i-nearMis{s}_i)}^2$
FCBF	$H\left(X\right)=-{\displaystyle {\displaystyle \sum }_i}P\left(x_i\right)lo{g}_2(P\left(x_i\right))$

.

4. Results

Given that the comorbidities prevalent in patients of the analyzed database are SAH with 49% of patients, DM with 34% of patients, and Obesity with 11% of patients, it was decided to take these three diseases independently as a target to find the variables that were most related to these diseases present in patients with COVID-19, determining the factors involved where the information had to be pre-processed. The Decision Tree and ranking methods were used to determine the variables with the most significant impact.

Table 6. Datasets for HAS as a target (variable names in Spanish).

Dataset	Target HAS
Dataset 1	Edad, género, grupo etario, DM, ECV, Hepáticas, SNC, Neumopatía, Cáncer, Inmunosupresión, Obesidad, Otros (1), Comorbilidades, Fiebre, Mialgias/artralgias, Fatiga, Odinogafia/ardor faringeo, Tos, Disnea, Dolor Toracico, Congestión Nasal, Rinorrea, Expectoración Diarrea, Náusea, Anorexia, Vómito, Cefalea, Mareo, Hispomia/Anosmia, Ageusia, Conjuntivitis, Saturación >90, Saturación 80–90%, Saturación < 80%, Leucopenia, Leucocitosis, Neutropenia, Neutrofilia, Linfopenia, Linfocitosis, Eosinopenia, Trombocitopenia, Trombocitosis, TP normal, TP alargado, INR normal, INR Alto, TTPa normal, TTPa alargado, Creatinina normal, Creatinina alta, Ferritina normal, Ferritina alta, Dímero D normal, Dímero D Alto, Fibrinógeno normal, Fibrinógeno Alto, PCR normal, PCR alta, Procalcitonina normal, Procalcitonina alta, Troponina normal, Troponina alta, CPK normal, CPK alta, CK-MB normal, CK-MB alta, Albúmina baja, Albúmina normal, Bilirrubina total normal, Bilirrubina total alta, ALT/TGP normal, ALT/TGP alta, AST/TGO normal, AST/TGO alta, DHL normal, DHL alta, DHL > 1000, Fosfatasa alcalina normal, Fosfatasa alcalina alta, Muestra, POSITIVA, Mayor 50%, Moderado, Grave, Oseltamivir, Ceftriaxona, Claritromicina, Azitromicina, Levofloxacino, Otros (2), Hidroxicloroquina/Cloroquina, Tocilizumab, Esteroides, Pronación, Respondedor, Respondedor parcial, No respondedor, Alta por mejoría, Defunción, Días de estancia hospitalaria.
Dataset 2	Comorbilidades, Edad, Tos, CK-MB Normal, INR Alto, DM, Cefalea, Neutrofilia, Dímero, Leucocitos, Neumopatía, Obesidad, Días de estancia, CPK alta, Saturación 90, Eosinopenia, TP Alargado y Odinofagia
Dataset 3	Comorbilidades, Edad, CK-MB Normal, CK-MB Alto, DM, Neutrofilia, Dímero D Alto, Leucocitosis, Neumopatía, Obesidad, Dias de estancia, Otros (1)

,

Table 7. Datasets for DM as a target (variable names in Spanish).

Dataset	Target DM
Dataset 1	Edad, género, grupo etario, HAS, ECV, Hepáticas, SNC, Neumopatía, Cáncer, Inmunosupresión, Obesidad, Otros (1), Comorbilidades, Fiebre, Mialgias/artralgias, Fatiga, Odinogafia/ardor faringeo, Tos, Disnea, Dolor Toracico, Congestión Nasal, Rinorrea, Expectoración Diarrea, Náusea, Anorexia, Vómito, Cefalea, Mareo, Hispomia/Anosmia, Ageusia, Conjuntivitis, Saturación > 90, Saturación 80–90%, Saturación < 80%, Leucopenia, Leucocitosis, Neutropenia, Neutrofilia, Linfopenia, Linfocitosis, Eosinopenia, Trombocitopenia, Trombocitosis, TP normal, TP alargado, INR normal, INR Alto, TTPa normal, TTPa alargado, Creatinina normal, Creatinina alta, Ferritina normal, Ferritina alta, Dímero D normal, Dímero D Alto, Fibrinógeno normal, Fibrinógeno Alto, PCR normal, PCR alta, Procalcitonina normal, Procalcitonina alta, Troponina normal, Troponina alta, CPK normal, CPK alta, CK-MB normal, CK-MB alta, Albúmina baja, Albúmina normal, Bilirrubina total normal, Bilirrubina total alta, ALT/TGP normal, ALT/TGP alta, AST/TGO normal, AST/TGO alta, DHL normal, DHL alta, DHL > 1000, Fosfatasa alcalina normal, Fosfatasa alcalina alta, Muestra, POSITIVA, Mayor 50%, Moderado, Grave, Oseltamivir, Ceftriaxona, Claritromicina, Azitromicina, Levofloxacino, Otros (2), Hidroxicloroquina/Cloroquina, Tocilizumab, Esteroides, Pronación, Respondedor, Respondedor parcial, No respondedor, Alta por mejoría, Defunción, Días de estancia hospitalaria.
Dataset 2	Edad, Neutropenia, Comorbilidades, Cáncer, Claritromicina, HAS, linfocitosis, Ferritina normal, Hepáticas, SNC, Leucopenia, Inmunosupresión, eosinofilia, ferritina alta, Troponina normal, vómito, INR alto, CM-KB alta, Disnea, TTP alargado, Levofloxacino, Fatiga, AST/TGO alta, bilirrubina total alta, fiebre, creatinina alta, INR normal, Diarrea, Augesia.
Dataset 3	Edad, Género, HAS, Obesidad, Otros (1), Comorbilidades, Leucocitosis, Creatinina normal, Creatinina alta, Procalcitonina alta, Levofloxacino, Hidroxicloroquina

and

Table 8. Datasets for Obesity as a target (variable names in Spanish).

Dataset	Target Obesity
Dataset 1	Edad, género, grupo etario, HAS, DM, ECV, Hepáticas, SNC, Neumopatía, Cáncer, Inmunosupresión, Otros (1), Comorbilidades, Fiebre, Mialgias/artralgias, Fatiga, Odinogafia/ardor faringeo, Tos, Disnea, Dolor Toracico, Congestión Nasal, Rinorrea, Expectoración Diarrea, Náusea, Anorexia, Vómito, Cefalea, Mareo, Hispomia/Anosmia, Ageusia, Conjuntivitis, Saturación > 90, Saturación 80–90%, Saturación < 80%, Leucopenia, Leucocitosis, Neutropenia, Neutrofilia, Linfopenia, Linfocitosis, Eosinopenia, Trombocitopenia, Trombocitosis, TP normal, TP alargado, INR normal, INR Alto, TTPa normal, TTPa alargado, Creatinina normal, Creatinina alta, Ferritina normal, Ferritina alta, Dimero D normal, Dimero D Alto, Fibrinogeno normal, Fibrinogeno Alto, PCR normal, PCR alta, Procalcitonina normal, Procalcitonina alta, Troponina normal, Troponina alta, CPK normal, CPK alta, CK-MB normal, CK-MB alta, Albumina baja, Albumina normal, Bilirrubina total normal, Bilirrubina total alta, ALT/TGP normal, ALT/TGP alta, AST/TGO normal, AST/TGO alta, DHL normal, DHL alta, DHL > 1000, Fosfatasa alcalina normal, Fosfatasa alcalina alta, Muestra, POSITIVA, Mayor 50%, Moderado, Grave, Oseltamivir, Ceftriaxona, Claritromicina, Azitromicina, Levofloxacino, Otros (2), Hidroxicloroquina/Cloroquina, Tocilizumab, Esteroides, Pronación, Respondedor, Respondedor parcial, No respondedor, Alta por mejoría, Defunción, Días de estancia hospitalaria
Dataset 2	Saturación 80–90, Edad, Cefalea, Género, Levoflaxina, GRAVE, Hisponia/asmonia, Linfopeni, a Neumopatía, Fosfatasa alcalina normal, Creatinina normal, días de estancia, PCR alta
Dataset 3	Saturación 80–90, Edad, Género, Levoflaxina, GRAVE, Hisponia/asmonia, Linfopenia Neumopatía, Fosfatasa alcalina normal

show the resulting datasets after the analysis. Dataset 1 in each table was determined by using all of the existing variables in the database; then, for dataset 2, the selection of features was determined by choosing the ones closer to the root within the first ten levels; finally, to assess dataset 3, ranking results were considered, regardless of the variables “Tos” (cough), “Fiebre” (fever), “Disnea” (dyspnea), and “Dolor de cabeza/Cefalea (headache)” as these were established by the World Health Organization as official COVID-19 symptoms, and by using them for analysis they performed better obtaining an increase in the dataset scores, therefore by dismissing them the research found other significant variables.

After the analysis of each dataset, the obtained results are shown in

Table 9. Highest scores from a dataset with HAS as a target.

Model	AUC	CA	F1	Precision	Recall
Tree Decision	0.814	0.784	0.784	0.784	0.784
SVM	0.867	0.762	0.762	0.762	0.762
Random Forest	0.866	0.784	0.784	0.784	0.784
Neural Network	0.876	0.773	0.773	0.773	0.773
Naive Bayes	0.832	0.757	0.755	0.761	0.757
Logistic Regression	0.910	0.800	0.800	0.801	0.800
AdaBoost	0.860	0.811	0.811	0.811	0.811

,

Table 10. Highest scores from a dataset with DM as a target.

Model	AUC	CA	F1	Precision	Recall
Tree	0.867	0.881	0.878	0.882	0.881
SVM	0.934	0.886	0.885	0.886	0.886
Random Forest	0.877	0.849	0.847	0.847	0.849
Neural Network	0.871	0.816	0.812	0.813	0.816
Naive Bayes	0.849	0.838	0.838	0.838	0.838
Logistic Regression	0.912	0.892	0.888	0.897	0.892
AdaBoost	0.827	0.843	0.844	0.844	0.843

Best model: Logistic Regression with 0.934 of AUC, 0.886 of CA, 0.885 of F1, 0.886 of Precision, and 0.886 of Recall.

and

Table 11. Highest scores from a dataset with Obesity as a target.

Model	AUC	CA	F1	Precision	Recall
Tree	0.700	0.876	0.869	0.864	0.876
SVM	0.643	0.838	0.831	0.824	0.838
Random Forest	0.807	0.903	0.872	0.912	0.903
Neural Network	0.834	0.914	0.899	0.906	0.914
Naive Bayes	0.793	0.865	0.845	0.833	0.865
Logistic Regression	0.778	0.881	0.856	0.851	0.881
AdaBoost	0.697	0.881	0.879	0.876	0.881

Best model: Neural Network with 0.834 of AUC, 0.914 of CA, 0.899 of F1, 0.906 of Precision and 0.914 of Recall.

; where the best scores are shown in bold font for each method.

As shown in Table 8, the highest score using as target HAS was obtained with the Logistic Regression algorithm, using the default parameters with a cost strength of 1.00 (C-1) and Ridge-type regularization (L2), which shows 91.0% Area under ROC (AUC), 80% Classification Accuracy (CA), 80% F1, and 80.1%, Precision and Recall for selected variables in Dataset 3 in Table 6 as possible variables essential to consider for a more accurate determination of vulnerability in people with HAS.

On the other hand, when targeting DM patients, the best-functioning model using the default parameters already mentioned above was found to be Logistic Regression, obtaining 91.2% AUC, 89.2% CA, 88.8% F1, 89.7% precision, and recall 89.2% for selected variables from dataset 3 in Table 7.

While the Neural Network algorithm showed better results for patients with Obesity, obtaining 83.4% AUC, 91.4% CA, 89.9% F1, 90.6% Accuracy, and 91.4% Recall (See Table 11).

The prevalent symptoms present in people with different comorbidities, and statistical analysis was performed without considering cough, fever, dyspnea, and headache as part of the symptoms for the same reason mentioned above. In people with HAS, DM, and Obesity, the presence of fatigue and myalgias/arthralgias was greater; while the third dominant symptom in people with HAS and DM was odynophagia, instead of people with Obesity, this symptom was positioned with the eighth, occupied the third position with chest pain (See Figure 2, Figure 3 and Figure 4).

5. Discussion

Obesity, diabetes, and hypertension are high-prevalence comorbidities in the Mexican population. In 2018, approximately 1/4 of the Mexican people had high blood pressure, while 71.2% of people over the age of 20 had a prevalence of obesity and overweight. These two conditions are considered to be significant risk factors for developing other diseases such as diabetes since 90% of people with diabetes in Mexico are overweight and obese [28].

In processing the information in this study, the prevalence of these three diseases considered as risk factors for COVID-19 infection could be noted. For this reason, it was decided to carry out an analysis using Artificial Intelligence (Machine Learning) techniques, aiming for each of these diseases, looking for any relationship between the symptoms and conditions of people suffering from these comorbidities.

After reviewing the medical literature, it was found that COVID-19 causes chronic inflammation in patients with obesity, along with other diseases considered as risk factors, such as lack of vitamin D and intestinal dysbiosis, which result in deficiencies in the functioning of the immune system in the face of infections. On the other hand, Obesity has a negative impact on respiratory mechanics as it is affected due to the resistance generated by lack of elasticity in the chest box [29,30]. Both diabetes and high blood pressure are diseases more present in the elderly, one of the reasons why these diseases are thought to be closely related to COVID-19, due to the fact that people with advanced age are highly vulnerable to COVID-19. The reason diabetes could have a high impact on the condition of patients with COVID-19 may be due to a disruption generated in the endocrine system where the COVID-19 virus affects angiotensin-converting Enzyme 2 (ECA2), which is responsible for anti-inflammatory regulation, vasodilators, and the process of releasing sodium into the urine. Another function of ECA2 is to offer protection to different organs, including those that are part of the pulmonary and cardiovascular systems, which is why the absence of this enzyme is related to the involvement of the lungs and the development of hypertension [31,32].

Different authors mentioned in a report of the analysis carried out with a database of 3894 patients in Italy [33], obtaining values for a death risk target with an accuracy rating of 83.4%, F1 value of 90.4%, specificity of 30.8%, and Recall of 95.2% when using Random Forest over a dataset with variables such as Glomerular Filtration Rate (eGFR), C- Reactive Protein (CRP), Age, Diabetes, Sex, Hypertension, Smoking, Lung Disease, Myocardial Infection, Obesity, Heart Failure, and Cancer, demonstrating this as vulnerable to those who exhibit the above-mentioned comorbidities. To achieve a comparison between the variables established with the results of this analysis, the scores of the three targets used (HAS, Diabetes, and Obesity) were taken into account and the average was calculated, obtaining values of 84.50% for CA, 83.43% of F1, 84.76% Precision, and 84.53% of Recall.

On the other hand, a study of Brazilian patients shows their scores at 86% for CA, 92% for AUC, 28% of Precision, 86% of Recall, and 42% of F1 for the Logistic Regression model, where the results show that there is a high relationship between patients over 60 years of age with breathing difficulties, fever, cough, rhinorrhea, odynophagia, diarrhea, headache, heart disease, pneumopathies, kidney disease, diabetes, smoking, Obesity, and hospitalization with those who come down with COVID-19 [34].

This analysis provides a complete and accurate perspective of the current situation in Tijuana since having information from a single source allows us to know the current situation of those served at the Tijuana General Hospital, who are mostly part of low-income or economically disadvantaged family groups. Since 41.9% of the Mexican population is economically disadvantaged [35], it is important to understand the comorbidities and symptoms present in this group, which results in more vulnerable people and increases expenditures in public sector hospitals regarding space and medical equipment, which often is extremely limited. Since the study was done considering typical human patients, regardless of their geographical and ethnic origin, a similar genomic situation can be assumed with migrant patients from both parts of Mexico and regions such as Central America, South America, and the Caribbean.

6. Conclusions

The values of the prevalent comorbidity found in this study were as follows: 49% in patients with HAS; 34% in patients with DM; and 11% in patients with Obesity, corresponding to the population segment that was served at the General Hospital of Tijuana during the aforementioned dates.

The deaths reported during the evaluated period of the total number of patients evaluated in this study were 42 subjects, from a population universe of 185 patients evaluated, with a fatality rate of 22%, a non-representative sample but within the data that places Mexico with a fatality rate (Case fatality) of 8.8% compared to other fatality rates in Latin America, such as Peru (3.5%), Colombia (2.6%), Chile (2.5%), Brazil (2.4%), and 2–3% globally [36].

A total of 52 different medicines were prescribed, where Steroids (16.1%), azithromycin (10.9%), enoxaparin (8.2%), levofloxacin (7.3%), hydroxyquinone (6.6%), omeprazole (6.2%), and acetaminophen (6.2%) were most commonly used. The rest, presenting themselves with a <6.0% per drug, correspond as a whole to 38.5% and are classified as “other”.

The values obtained for each dataset evaluating Obesity, DM, and HAS in this study using Artificial Intelligence, specifically with various Machine Learning techniques, have been compared with values presented in other similar academic/medical publications, particularly those for HAS were obtained with the Logistic Regression algorithm obtaining 91.0% AUC, CA, F1 and Recall 80%, and accuracy 80.1%. For Diabetes, the Logistic Regression algorithm obtained 91.2% AUC, 89.2% CA, 88.8% F1, 89.7% accuracy, and recall 89.2%, obtaining similar results for Hypertension and Diabetes, while improved results were obtained in the case of Obesity using the Neural Network algorithm, obtaining 83.4% AUC, 91.4% CA, 89.9% F1, 90.6% Accuracy and 91.4% Recall.

The results presented here confirm the convenience of using Logistic Regression algorithms in the dataset presented when evaluating HAS as a target; Similarly, Logistic Regression’s algorithms were the most successful in evaluating DM as a target; while finally, Neural Network algorithms showed the best results for the case of Obesity as a target, in the specific sample case of 185 patients with limited or non-existing financial resources, who suffered these medical conditions, and who were served at the Tijuana General Hospital during the year 2020.

With these results, we can conclude that the use of Artificial Intelligence using Machine Learning techniques has been effectively used to identify the early stages of COVID-19 in patients in Baja California.