Introduction

The emergence of the novel coronavirus SARS-CoV-2 has triggered an unprecedented global health crisis, leading to the COVID-19 pandemic and posing immense challenges to healthcare systems worldwide1. In response to this ongoing crisis, the scientific and medical communities have demonstrated extraordinary adaptability and collaboration, striving to accelerate the development of diagnostic tools, treatments, and vaccines2. One study analyzed factors influencing the spread and mortality of COVID-19, finding that unemployment rate and population density significantly impacted the infection rate, while diabetes prevalence and hospital bed availability were linked to mortality3. Additionally, spatial statistical models have been utilized to examine the spatiotemporal distribution of cumulative incidence across Europe, identifying poverty and an aging population as key factors contributing to the severity of the pandemic4.

Amid the multifaceted challenges presented by the pandemic, the search for effective oral antiviral drugs has become a research priority. Currently, several antiviral drugs have been approved for the treatment of COVID-19, including nirmatrelvir-ritonavir (Paxlovid), monoclonal antibodies, remdesivir, and molnupiravir5. However, during the outbreaks caused by the Omicron variant in late 2022 and early 2023, the most widely used antiviral drugs in China were Paxlovid and Azvudine. Azvudine, a reverse transcriptase inhibitor, was approved by the China National Medical Products Administration in July 2022 for the treatment of COVID-196. Azvudine significantly reduced the risk of disease progression compared to Paxlovid in hospitalized COVID-19 patients with underlying conditions, particularly among males, patients under the age of 65, and those requiring antibiotic treatment, as reported in a study7. Similarly, another study found that Azvudine was more effective than nirmatrelvir-ritonavir (Paxlovid) in reducing disease progression and all-cause mortality, especially in patients under the age of 65 and those with severe COVID-19 manifestations8. Moreover, it was reported that Azvudine effectively reduced the rate of disease progression and hospitalizations among high-risk non-hospitalized patients with mild-to-moderate COVID-19, particularly when administered within three days of symptom onset, which also shortened the duration of fever9. In contrast, a different study compared the antiviral effects of Azvudine and nirmatrelvir-ritonavir in hospitalized COVID-19 patients, finding that while the latter achieved faster reductions in viral load, both drugs showed no significant difference in disease progression rates10. As a result, the choice of antiviral drugs for different COVID-19 patients remains controversial. The dynamic nature of the pandemic and the evolving treatment guidelines underscore the urgency of further research11.

Building on the findings of [7–10], this study retrospectively collected real-world clinical data from a specific period, encompassing patient cohorts with varying underlying conditions. The study integrates statistical modeling techniques and personalized prediction tools to rigorously assess the therapeutic efficacy of the oral antivirals Azvudine and Paxlovid in the treatment of COVID-19. Additionally, it examines key determinants—including age, severity stratification, and drug type—that affect viral clearance time. The overarching objective is to develop a comprehensive understanding of these antivirals’ effectiveness across diverse patient populations, thereby providing valuable evidence to inform clinical decision-making and guide public health policy.

Materials and methods

Study population

To assess the therapeutic efficacy of COVID-19 antiviral drugs, we conducted a stratified analysis based on the type of drug administered, examining differences across gender, age, disease classification, initial CT values, and prognosis, while evaluating their association with COVID-19 conversion to negative status. It is important to note that hospital discharge does not necessarily equate to full recovery, specifically in terms of COVID-19 symptom resolution or viral clearance. A chi-square test (cross-analysis) was employed to quantify the impact of each factor on viral clearance, revealing that the initial CT value had a significant effect.

Our findings underscore the complexity of COVID-19 treatment and recovery, highlighting the crucial role of individual patient characteristics—such as initial CT value—in shaping treatment outcomes. Notably, p-values for all features related to drug types were above 0.05, indicating no significant correlation among these treatment groups. The overall research methodology is depicted in Fig. 1. This study was approved by the Institutional Review Board (IRB) of The Second Xiangya Hospital of Central South University (approval number: LYF-2023015). All methods were carried out in accordance with relevant guidelines and regulations. Due to the retrospective nature of the study, the IRB of The Second Xiangya Hospital of Central South University waived the need for obtaining informed consent.

Data collection

We retrospectively recruited COVID-19 patients from the inpatient department of the Second Xiangya Hospital of Central South University. The inclusion criteria were: (a) age 18 years or older; (b) patients diagnosed with COVID-19 according to the World Health Organization (WHO) case definition (2022). Patients who received both Azvudine and Paxlovid therapies were excluded from the study.

As detailed in Table 1, information was collected from electronic medical records, covering the following aspects:

  1. 1.

    Demographic Data: Including gender, age, and ethnicity.

  2. 2.

    COVID-19 Disease Course Data: Including the time of onset of COVID-19 symptoms (e.g., cough, expectoration, shortness of breath), the time of COVID-19 nucleic acid or antigen positivity, the time of initial oral antiviral drug treatment, the type of antiviral drugs used, the cycle threshold (CT) value of each COVID-19 detection during hospitalization, and the time of conversion to negative status.

  3. 3.

    Medical History: Including whether patients had pre-existing chronic conditions such as hypertension, coronary heart disease, diabetes, chronic liver disease, chronic respiratory diseases, and chronic kidney disease, as well as smoking history.

As shown in Fig. 1, a total of 127 patients were recruited based on the inclusion criteria. We performed data preprocessing to ensure data quality, excluding records of patients with missing key information or clinically unreasonable treatment details. Among them, 10 patients who underwent only a single antigen/nucleic acid test, with missing data on conversion to negative status, were excluded. Additionally, some patients classified as mild cases (n = 1) or critical cases (n = 7) were also excluded from the study. Ultimately, 109 patients were included in the final study cohort, all of whom had recorded data for the first positive antigen result and the conversion to negative, along with a comprehensive medical history, as detailed in Table 1. Of these patients, 35 received Paxlovid treatment, 33 received Azvudine, and 41 did not receive any antiviral treatment.

Statistic analysis

All statistical analyses were conducted using Python (version 3.10.9) and R (version 4.3.1), employing a range of specialized libraries. In Python, Pandas was used for data manipulation and cleaning, while NumPy handled numerical operations and calculations. Matplotlib and Seaborn were utilized for data visualization, including kernel density estimations12and box-and-whisker plots13. Statsmodels was used for chi-square tests and regression analyses, and lifelines for Cox proportional hazards regression analysis14. In R, the Survival package facilitated survival analysis and Cox regression modeling, ggplot2 was used for visualizing results such as schoenfeld residual plots, and rms was used for creating nomograms15 and conducting model calibration.

We employed several statistical methods for comprehensive data analysis. Descriptive statistics were used to summarize demographic information, presenting categorical variables as counts and percentages. Chi-square tests were applied to evaluate differences between categorical groups. Kernel density estimation was used to analyze the distribution of recovery times across subgroups, and box-and-whisker plots helped to identify data distributions and potential anomalies. The Cox proportional hazards model was applied to assess the influence of various factors on the likelihood of transitioning to a negative COVID-19 status, with Schoenfeld residual plots used to evaluate the proportional hazards assumption. Finally, nomograms were developed to predict individual recovery probabilities, providing personalized insights into patient outcomes.

Fig. 1
figure 1

The flowchart describing the overall research process.

Full size image
Table 1 Demographic statistics.
Full size table

Results

Antiviral treatment effects on recovery

We conducted an in-depth analysis, as shown in Fig. 2, to assess the impact of antiviral agents on recovery times among COVID-19 patients. Our findings indicate that patients receiving antiviral treatment had a significantly greater probability of recovery, with Azvudine demonstrating pronounced efficacy during the initial six days, followed by a more substantial effect observed with Paxlovid in subsequent stages. This trend is illustrated by the kernel density plots16 (Fig. 2-c).

Furthermore, when evaluating the relationship between age and recovery time across the three patient cohorts, we observed a notable clustering effect within the 50 to 80-year age range. Patients treated with Azvudine exhibited the shortest recovery times, followed by those receiving Paxlovid, whereas untreated patients experienced the longest recovery durations. These findings underscore the benefits of antiviral interventions in expediting recovery, with distinct patterns emerging based on both drug type and patient age.

Fig. 2
figure 2

Assessing Different COVID-19 Drugs on Recovery. (a) Shows the chance of achieving a negative COVID-19 status for patients who took medication compared to those who did not. (b) Compares the likelihood of turning COVID-19 negative after using two drugs, Paxlovid and Azvudine, using the Kaplan-Meier model. (c) Examines connections between drug types, patient age, and recovery time in COVID-19 patients. This helps us understand potential links between these factors.

Full size image

Influence of age and CT values on recovery outcomes

We investigated the influence of age and CT values on recovery outcomes, as illustrated in Fig. 3, which presents recovery times for patients treated with Paxlovid (Fig. 3-a), Azvudine (Fig. 3-b), and those without antiviral therapy (Fig. 3-c). Our results indicate that Azvudine provided considerable clinical benefits for patients aged 50–80, although findings for patients aged 35–50 and those over 80 years should be interpreted with caution due to limited sample sizes.

Additionally, our analysis revealed differential effects of the antiviral drugs on improving lung CT values and symptom relief. Paxlovid exhibited notable efficacy, particularly in patients with moderate disease severity. Patients classified as moderate showed substantial benefits from both Azvudine and Paxlovid, whereas the impact was markedly diminished in those with severe disease. These findings underscore the pivotal role of antiviral agents in mitigating disease severity and enhancing lung function, particularly in targeted patient subgroups.

Fig. 3
figure 3

Analysis of the impact of CT values and classifications at different ages and periods on the effectiveness of taking different drugs. (a), (b), and (c) Shows how different age groups influence recovery times when using different drugs. This helps us understand age-related recovery variations with various medications. (d) Illustrates how different drugs affect CT values, a critical diagnostic measure in COVID-19. This reveals the impact of drug choices on diagnostic indicators. (e) Examines how different drugs impact recovery times among different symptom classifications. It highlights variations in drug efficacy for distinct groups of COVID-19 patients.

Full size image

Cox proportional hazards analysis

The Cox proportional hazards model is primarily used in survival analysis to evaluate the impact of multiple covariates on time-to-event outcomes, such as patient survival or recovery time17. A key feature of this model is its ability to handle censored data without requiring assumptions about the specific distribution of survival times18. We used the cph function in R to fit the Cox model, where the survival object contains both a time variable and an event indicator (i.e., whether the patient turned negative). The model includes covariates such as drug type, age, initial CT value, stratification, CT1, and CT2.

To validate whether the covariates in the Cox model satisfy the proportional hazards assumption, we generated Schoenfeld residual plots19. As shown in Fig. 4, the residuals for the six covariates—drug type, age, initial CT value, stratification, CT1, and CT2—were randomly distributed along the horizontal axis, and all individual test p-values exceeded 0.05 (drug type p = 0.85, age p = 0.33, initial CT value p = 0.74, stratification p = 0.52, CT1 p = 0.54, and CT2 p = 0.88). These results indicate that each covariate satisfies the proportional hazards assumption. Additionally, the global Schoenfeld test p-value was 0.88, further supporting a good fit between the Cox model and the dataset. These findings suggest that there are no significant discrepancies between model predictions and actual observations, thus validating the robustness and appropriateness of the model.

Fig. 4
figure 4

Schoenfeld residual plots to Assess Variable Influence and Model Suitability.

Full size image

Prognostic analysis using nomograms

Figure 5provides a comprehensive evaluation of how six key variables—drug regimen, age, initial CT value, stratification, and follow-up CT value—collectively influence recovery probability, utilizing a nomogram20. This figure also integrates calibration curve analysis, which is crucial for assessing the accuracy of the nomogram. The calibration plots compare predicted survival probabilities (x-axis) with observed survival rates (y-axis), thereby evaluating the concordance between model predictions and actual outcomes. Such an evaluation yields critical insights into the precision and reliability of our recovery probability estimates.

To illustrate the practical utility of the nomogram, we considered a scenario involving a 70-year-old male COVID-19 patient undergoing Paxlovid treatment, classified as having moderate disease with specific CT values. Figures 5-b and 5-c present calibration plots indicating the probability of achieving negative status within 15 and 30 days, respectively. According to these plots, a total score of 60 points corresponds to an 80% probability of recovery within 15 days, increasing to 90% within 30 days. This nomogram-based approach supports personalized recovery and treatment outcome assessments, thereby enhancing the precision of clinical predictions tailored to individual patient characteristics.

Fig. 5
figure 5

Nomogram Analysis and Validation - Understanding Factors Affecting Recovery Probability.

Full size image

Discussion

Current clinical guidelines prioritize the use of Paxlovid and Azvudine in the treatment of COVID-19 patients; however, comprehensive real-world evidence evaluating the efficacy of these antiviral agents across diverse patient populations remains relatively sparse.

In this study, we employed the Kaplan-Meier model to assess the therapeutic effects of Azvudine and Paxlovid on COVID-19 recovery, with a particular focus on the time to viral clearance. Notably, Azvudine demonstrated a pronounced effect during the initial six-day period, facilitating a more rapid transition to negative status, whereas Paxlovid exhibited superior efficacy during later stages of recovery, particularly among patients in the mid-symptomatic phase. These findings indicate distinct, stage-dependent therapeutic roles for the two drugs. Additionally, Azvudine was most efficacious in patients aged 50 to 80 years, leading to the shortest recovery durations, which underscores significant variability in drug response across different age cohorts. Conversely, due to limited sample size, the interpretation of outcomes for patients aged 35–50 and over 80 years should be approached with caution.

Furthermore, we evaluated recovery times stratified by symptom severity (mild, moderate, severe) across the treatment groups. Both Azvudine and Paxlovid conferred substantial benefits in promoting recovery among patients with moderate symptoms, though their effects were less pronounced in severe cases, indicating differential efficacy based on disease severity, with the greatest benefits observed in moderate cases.

To further validate our findings, we utilized Schoenfeld residual plots to evaluate the impact of key covariates on recovery time within the Cox proportional hazards model, confirming the validity of the proportional hazards assumption. The model demonstrated a satisfactory fit, suggesting its robustness in predicting COVID-19 recovery trajectories. We also employed a nomogram to quantify the integrated influence of six key variables (treatment group, age, initial CT value, stratification, and follow-up CT value) on the probability of recovery, supported by calibration curves that assessed model accuracy. Collectively, these analyses demonstrate the high applicability and reliability of the Cox proportional hazards model and nomogram for predicting recovery outcomes, providing critical support for personalized treatment decisions and optimizing therapeutic strategies.

Conclusion

Our study demonstrates that oral antiviral agents significantly shorten the time to viral clearance in COVID-19 patients with moderate disease severity. We specifically assessed the differential effects of Azvudine and Paxlovid on viral clearance dynamics, providing a nuanced understanding of their respective therapeutic roles. Furthermore, we constructed a statistical model to facilitate prognosis assessment and predictive analysis, which underwent rigorous evaluation and validation. However, given the limited sample size of this investigation, caution is warranted when interpreting the findings, particularly across different age cohorts. As this study is based on data from a single-center, small-sample validation, the generalizability of these outcomes to other populations and regions remains contingent upon further validation with larger, more diverse datasets.