Introduction

The COVID-19 pandemic has generated a global health, economic, and humanitarian crisis, causing, until July 2021, more than 100 million cases and more than 3.9 million deaths [1]. Fatality from COVID-19 was estimated to be less than 3% during the first months of the pandemic [2]; however, later estimates reported rates between 10 and 48%, which varied between locations (continents and countries) [3]. The characteristics of each location (environmental, economic, structural, etc.) probably contribute to the changes in fatality. For instance, temperature, concomitant epidemics of other respiratory infections (tuberculosis, etc.) [4], the type of viral variant and changes in its prevalence and dispersion, as well as the availability of medical resources mainly influence fatality rates.

In view of the high fatality rates due to COVID-19, the World Health Organization [5] proposed multiple initiatives to promote the creation of predictive fatality models to assist decision-makers in the development of strategies, planning, and design of public policies for the prioritization of the most vulnerable groups. Different studies have used population databases to assess whether clinical, sociodemographic, and laboratory characteristics influence the occurrence of negative outcomes among COVID-19 cases [6–8]. However, due to the differential distribution and representation of specific races/ethnicities [9], characteristics of healthcare systems [10], social vulnerability (which entail higher tendencies of crowding and viral exposure) [11], economic characteristics (which bring about inequalities in healthcare access) [12], geographic factors (climatic factors and altitude levels) [13, 14], distribution of comorbidities [15], and genetic factors that could modify the disease burden and its trends [16], conduct that the usefulness of these prediction models may be limited in contexts other than those in which they were developed [17].

Peru has had a variable and dynamic COVID-19 incidence [18]. The different presentations of the disease, the interventions (pharmaceutical and non-pharmaceutical), the adopted treatment protocols, and the spread and predominance of local viral variants have most likely influenced the COVID-19 incidence. These factors differed significantly from those observed in the United States and England [19], for example, and even amongst countries in the same region [20, 21]. Globally, Peru had the highest number of deaths per 100,000 inhabitants during the first and second COVID-19 waves [22] and, until July 2021, was the fifth country with the highest case-fatality ratio from COVID-19 [1]. A number of studies have reported various factors associated with COVID-19 fatality in the Peruvian population [23–27]; however, their results could be inaccurate and non-representative given multiple limitations such as; low sample size, lack of temporality, and selection bias. This study aimed to build multiple prediction models of COVID-19 fatality using population data from the Peruvian surveillance system.

Methodology

Results

The database included a total of 23,742 confirmed cases of COVID-19 with a positive test and symptom onset date up to September 30, 2020, of which 1,404 were excluded for being under 18 years of age, 169 for not having Peruvian nationality, and 71 because they came from a department other than Lima (S3 Fig). After exclusions, the database included a total of 22,098 cases (mean age 45.96 ± 16.82; 53.41% female) eligible for the analysis. The dataset used for the model building and the one used for validation included a sample of 17,678 (mean age 45.99 ± 16.86; 53.25% female) and 4,420 cases (mean age 45.85 ± 16.64; 54.05% female), respectively. (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Characteristics of the COVID-19 confirmed cases in nine provinces of the department of Lima, Peru.

https://doi.org/10.1371/journal.pgph.0002854.t001

The sociodemographic characteristics, the type of test used for diagnosis, the notifying healthcare institution, clinical characteristics and fatality were comparable within datasets (Table 1). Also, survival curves within datasets were comparable (S4 Fig).

Initially, three models were built using the following strategies; strategy 1 (includes variables previously described as predictors elsewhere) with 13 variables, strategy 2 (based on the LASSO method) with nine variables, and strategy 3 (based on significance in the bivariate logistic regression) with 22 variables (Table 2). Interestingly, the three models consistently included three variables: age, sex, and dyspnea (Table 2). In view of this, we defined a fourth strategy to build a post-hoc model including only the previous three variables.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Predictive factors of fatality among COVID-19 confirmed cases using the model-building dataset (N = 17 678).

https://doi.org/10.1371/journal.pgph.0002854.t002

Upon observing this pattern, a post-hoc strategy was implemented. When evaluating the performance of the constructed models, it was identified that for a death probability cutoff of 52%, similar estimates were obtained, with slightly higher sensitivity and specificity for strategy 3 (Sensitivity (S): 83.08%; Specificity (E): 82.30%), followed by strategy 1 (S: 82.58%; E: 82.07%), strategy 2 (S: 80.00%; E: 82.35%), and strategy 4 (S: 80.78%; E: 81.75%) (Table 3). Similarly, similar areas under the curve were observed for the models obtained through the four variable selection strategies (S5 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Performance of the four prediction models for death in COVID-19 confirmed cases (N = 4 420).

https://doi.org/10.1371/journal.pgph.0002854.t003

Although the four variable selection strategies resulted in models with similar performances, models constructed with strategies 1 and 4 were selected for the following reasons. The model from strategy 1 included a smaller number of variables compared to strategy 3 (13 vs. 22), and in comparison, to strategy 2, the performance was slightly better (AUC 0.89 vs. 0.88). On the other hand, the model from strategy 4 is a condensed model with performance comparable to the other strategies. Interestingly, the area under the curve of the selected models in the validation and creation datasets was comparable for the selected model strategies (Strategy 1: 0.89 vs. 0.89 / Strategy 4: 0.89 vs. 0.88; Fig 1). Moreover, using the validation dataset, the performance of the models from strategies 1 and 4 remained stable when conducting a sensitivity analysis including only cases confirmed by RT-PCR (Strategy 1: 0.89 vs. 0.89 / Strategy 4: 0.86 vs. 0.88) and according to pandemic periods (Strategy 1: first period = 0.89 vs. 0.89; second period = 0.89 vs. 0.89; peaks between incidence = 0.88 vs. 0.89 / Strategy 4: first period = 0.88 vs. 0.88; second period = 0.90 vs. 0.88; peaks between incidence = 0.87 vs. 0.88) (S6 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Validation of the selected models for the prediction of death in confirmed cases of COVID-19.

(A) Performance of the model derived from strategy 1 and 4 in the dataset used for the model construction (N = 17 678) and (B) performance of the model derived from strategy 1 and 4 in the dataset for validation (N = 4 420).

https://doi.org/10.1371/journal.pgph.0002854.g001

Discussion

Throughout the pandemic, Peru was one of the countries with the highest number of deaths from COVID-19 per 100,000 inhabitants and with a case fatality rate of approximately 5%, which was higher than countries such as Mexico and the United States [38]. Fatality has been concentrated mainly in the city of Lima, the capital of Peru, because most the population were concentrated there [1]. Lima is surrounded by nine provinces; unfortunately, to the best of our knowledge, despite the great connectivity between the capital and the provinces, fatality and its trends in these provinces have not been explored. The disparities between the city center of Lima and its provinces can be substantial, primarily driven by cultural, ethnic, social, racial, economic, and resource inequalities factors [39], that may contribute to the variations in COVID-19 trends between these locations. In this regard, the results of this study provide an initial characterization of this population. Specifically, our study explored various predictors of death from COVID-19 in the nine provinces of the department of Lima and identified a fatality of 7.5%, corresponding to the first and second pandemic waves. Overall, the high fatality can be attributed to the low preparedness of the Peruvian health system to prevent fatalities. Notably, Peruvian provinces faced much greater challenges due to their limited number of high-capacity hospitals and thus faced a pandemic not only with an unprepared system but also with shortages of intensive care units, medical devices, and specialized personnel [22]. On the other hand, it is necessary to consider that Peru also prioritized the use of serological or rapid tests for the diagnosis of COVID-19, as reflected in our study. This approach resulted in a missed opportunity to diagnosed early and, consequently, in a hindered containment of future infections and prevention of fatalities, especially in provinces–such as those described here–with no available laboratories to perform molecular testing.

Considering that the COVID-19 pandemic was characterized by the disproportionate use of health resources, it is still necessary to structure and articulate prioritization systems to prevent the occurrence of negative outcomes in vulnerable populations. Here, we describe the development of four models to predict death by COVID-19 considering different variable selection strategies. Globally, all models performed adequately; however, based on the criteria described above, we choose the models derived from strategies 1 and 4. Nevertheless, the use of the other two models remains possible and therefore, we recommend further evaluation.

The model constructed with strategy 1 included clinical predictors reported in previous studies [7, 8, 36]. These studies used population-based data from the United States and described models that performed well (AUC = 0.8) [7, 8, 36]. Although we hypothesized that using previously described models in populations other than those for which they were constructed might not be appropriate because of potential different characteristics between populations, health systems, or the epidemiology of COVID-19; we observed that the use of the variables previously described generates a model that performs well for the prediction of COVID-19 fatality in the Peruvian population studied. We even reported an AUC greater than 0.8, which was robust in the validation and sensitivity analysis. Also, the performance described here is comparable to that reported in other studies that use national databases [7, 8, 36] and clinical [7, 36] and laboratory [8] information. However, although the clinical variables of our model were obtained at patient enrollment, is likely that other symptoms prior to care or symptoms that appeared during the clinical course play an important role in the prediction. On the other hand, we considered some variables as a proxy of unmeasured variables. For instance, fever and hypertension were considered as proxies for temperature and arterial pressure at enrollment [7, 8, 36]. Despite this, we consider that this constraint does not significantly affect the performance of the model, due to the close relationship of the proxies used with the unmeasured variables. On the other hand, since the comorbidities were obtained through self-reports, and given that it is conceivable that there may be an underlying under-diagnosis of comorbidities, it is therefore likely that their effect and contribution to the model are underestimated, and consequently, the model performance is underestimated.

The model built based on strategy 4 only comprised three variables that were consistently included in the other three models. Although this model was the most parsimonious and performed similarly to the model constructed with strategy 1, it is important to note that it was built post-hoc and was based on findings from the other models. Therefore, the model 4 could be affected by multiple limitations described elsewhere [40]. It is important to note that the model was robust in sensitivity analyses, and its performance was comparable with other non-complex [10, 41, 42] and complex models described elsewhere [7, 8, 36]. Therefore, based on these observations, we consider that the model derived from strategy 4 could be used with caution.

In order to opt for one of the selected models, we propose a prior assessment of the quality of the data. As an example, if there is no confidence in how the information on comorbidities was measured or collected, the model 4 could be used to predict COVID-19 death. On the other hand, if the quality of the data related to comorbidities is reliable, model 1 could be the most suitable for predicting death, considering that the better marginal performance over model 4 would translate into better predictive power.

The selected models included some variables that were significantly associated with the outcome, such as dyspnea, age, male gender, diabetes, kidney disease, and obesity. These observations are discussed below.

Dyspnea was one of the characteristics that most increased the possibility of death, approximately 7-fold. Artificial intelligence modeling identified dyspnea as one of the most predictive factors of COVID-19 fatality [7]. This is because severe stages of COVID-19 are characterized by hypoxia, which often leads to respiratory failure and causes dyspnea. Previous studies have reported a higher incidence of death as oxygen saturation decreases [43, 44], which has also been corroborated in the Peruvian population [23, 24, 26].

In this study we identified that the higher the age, the greater the possibility of death. This relationship has been previously described [45–47], specifically indicating that fatality of COVID-19 increases by almost 50% in older age groups [47]. This is likely due to the fact that the burden of disease increases in older population, given that increased probability of comorbidities, polypharmacy, and frailty, and consequently a greater risk of negative health outcomes develop. It is important to consider that the relationship between age and death may be different in settings where vaccination coverage is high in the elderly and infection rates are higher in the young adult population [47]. Therefore, the extrapolation of this finding must be carried out with caution.

The chance of death from COVID-19 was approximately doubled in males. Previous studies have suggested that females have a better understanding of their health and potentially quicker access to healthcare compared to males, reducing the probability of death [48]. However, recent studies conducted in the context of COVID-19 have identified that females experience a higher delay in accessing healthcare [49, 50]. Thus, the association may not be related to delayed access, which is understandable given the fear surrounding the disease at the beginning of the pandemic. On the other hand, some studies hypothesized that sex hormones influence the response against COVID-19, suggesting that there is a higher expression of immune-related genes in women, which leads to a higher antibody response and protection against infections [51]. Recently, it has been reported that women maintain a high immune reactivity after viral infections and generate higher humoral responses to vaccines compared with men [52]. For this reason, if we consider the extrapolation of this observation in scenarios with variable COVID-19 vaccine coverage or infection, it would be expected that this association would remain.

Comorbidities such as diabetes, kidney disease, and obesity increased the chance of death from COVID-19. Although, it is known that COVID-19 causes myelopoiesis, T cells and natural killers dysregulation, and uncontrolled production of cytokines; the extent to which underlying comorbidity influences the immune response in the SARS-CoV-2 infection remains unknown [53]. However, previous studies have reported that obesity and diabetes lead to dysregulation of immune cells and hormones, generating decreased immunity and a higher probability of hyperinflammation [53–55], which consequently translates into a boosting of infection and the development of multiple adverse health events. On the other hand, regarding chronic kidney disease, it is known that normal renal function contributes to immune homeostasis given the filtration of circulating cytokines and immunogenic pathogenic components and, thus, inflammation is limited [56]. In this instance, decreased renal function leads to increased activation of innate immune cells and increased cytokine production [57], and the subsequent cytokine storm, thereby generating an exponentiation of the disease severity [53]. Although previous studies have described the relationship between other comorbidities and death [53, 58–60], the uncertainty in the measurement of comorbidities may play an important role in the non-association found.

Conclusion

Herein, we describe four models with different variable selection strategies for the prediction of COVID-19 fatality. Our findings suggest that, regardless of the strategy, the models displayed comparable performances. However, two prediction models, whose performances were optimal in the validation and sensitivity analyses, exhibited slight outperformance due to their plausibility with the outcome and/or their parsimony. Future studies should corroborate the performance and validate the usefulness of the models described under both actual and real-world conditions.

Supporting information