Evaluation of an Established Semi-Quantitative Chest CT Scoring System for Assessing the Severity of COVID-19 Pneumonia: What is Its Diagnostic Value Regarding Patient Outcomes?

1. Introduction

In March 2020, the World Health Organization (WHO) officially declared coronavirus disease 2019 (COVID-19), caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), a global pandemic. By the end of February 2026, more than 779,000,000 cases of COVID-19 and approximately 7,000,000 deaths had been reported worldwide [1]. Although mortality has steadily declined over time due to vaccination programmes as well as public health and hygiene measures, COVID-19 is still considered endemic in Germany. The high level of population immunity has contributed to a reduction in severe disease courses and long-term sequelae. Nevertheless, further waves of infection are expected, which will predominantly affect elderly individuals and those with pre-existing conditions [2].

Since the onset of the pandemic, chest imaging – particularly computed tomography (CT) – has played a crucial role. It is not only essential for monitoring disease progression and identifying complications, but also for supporting the diagnosis in suspected cases, for example with polymerase chain reaction (PCR) test results that are inconsistent with a clear clinical picture including typical COVID-19 symptoms [3].

On CT imaging, COVID-19 pneumonia most commonly presents with multifocal, predominantly bilateral ground-glass opacities (GGO). These typically exhibit a patchy or geographic distribution and are mainly located in the peripheral and posterobasal regions of the lungs. Consolidations and the so-called crazy-paving pattern are observed less frequently and tend to occur in more advanced stages of the disease [4]. In addition, interlobular septal thickening, vascular enlargement and the halo sign are frequently reported parenchymal findings associated with COVID-19 pneumonia [5].

Quantitative evaluation of lung parenchymal involvement has been shown to be an important factor in predicting disease outcome [6]. One approach is the semi-quantitative CT score proposed by Pan et al., in which each lung lobe is graded according to the extent of involvement (from 0 for no involvement to 5 for > 75 % involvement). A total score of ≥ 18 – calculated as the sum of all five lobes – is associated with a significantly increased risk of a fatal outcome [7,8]. However, this scoring system considers only the extent of lung involvement and does not account for the specific type of parenchymal abnormalities.

A previous study has already shown that inter-rater agreement on the scores is high, even among readers with varying levels of experience, and that fully automated, AI-based evaluations also produce reliable results [9]. It has also been displayed that the score correlates positively with inflammatory parameters, the age of the patients, and the severity of the disease [7].

The aim of this study was to evaluate the diagnostic value of the Pan score with regard to disease severity, i. e. the need for intensive care admission and mechanical ventilation, as well as the patients’ outcomes, and to investigate the extent to which it correlates with various laboratory parameters, the pulmonary function, and the age of the patients.

Our hypothesis was that a higher score would be associated with increased inflammatory parameters, a higher likelihood of intensive care admission, a greater probability of a fatal outcome, and poorer pulmonary function – independent of the patients’ age.

2. Materials and Methods

For this retrospective study, we included 351 consecutive patients (223 males [63.5%], 128 females [36.4%]; mean age 67.0 years) with polymerase chain reaction (PCR)-confirmed COVID-19, who underwent chest CT at our hospital between March 21, 2020, and December 27, 2021. The first case of COVID-19 in our district was confirmed in March 2020.

Patients were included based on the following criteria: (a) positive PCR test for COVID-19, (b) chest CT scan performed not more than three days prior or three days after the positive PCR test, (c) age ≥ 18 years. 38 of the original 607 patients were excluded owing to motion/breathing artifacts, prior extensive lung surgery, incomplete CT data sets, or other technical difficulties that prevented AI-based evaluation. Furthermore, a further 218 patients were excluded due to incomplete clinical and laboratory data for the study period. Approval for this study was granted by the local ethics committee (Medical Association Westphalia-Lippe and University of Münster) and conducted following the principles of the Declaration of Helsinki.

CT images were obtained from the hospital’s picture archiving and communication system (PACS; Centricity^TM Universal Viewer, Version 6.0, GE HealthCare Technologies Inc., Chicago, USA) and organised in a dedicated list within in the radiological information system (RIS; Centricity^TM RIS-i7, Version 7.0.4.4, GE HealthCare Technologies Inc., Chicago, USA).

Laboratory parameters, the results of blood gas analyses and pulmonary function tests, as well as the need for and duration of intensive care unit (ICU) admission and mechanical ventilation (both non-invasive [NIV] and invasive), were retrieved from the hospital information system (IS-H, SAP Deutschland SE & Co. KG, Walldorf, Germany).

Chest CT scans were performed using either a 128-slice (SOMATOM Definition Flash, SIEMENS Healthineers AG, Forchheim, Germany) or a 64-slice CT scanner (SOMATOM go.Top, SIEMENS Healthineers AG, Forchheim, Germany). Images were acquired during a single breath-hold from the lung bases to the apices. Scans performed to confirm (or exclude) suspected COVID-19 pneumonia were obtained without i. v. contrast media, while patients scanned for other clinical indications, such as suspected pulmonary embolism or oncologic evaluation, received 40–80 ml of an iodine-containing contrast agent (ACCUPAQUE^TM-300 or -350, GE HealthCare Buchler GmbH & Co. KG, Braunschweig, Germany) i. v. All scans utilised an automatic exposure control for tube voltage and current. Images were reconstructed using a sharp (pulmonary) and a soft (mediastinal) kernel, with lung images generated at slice thicknesses of 0.8, 1, 2, 3, or 5 mm.

CT images were independently reviewed by a radiology specialist (senior physician), an experienced assistant physician, and a medical student. Readers assessed typical COVID-19 pulmonary features, including ground-glass opacities (GGO), consolidations, and crazy-paving patterns, as well as other abnormalities such as nodules or halo signs. All evaluations were performed blinded to clinical data (laboratory values, blood gas analyses, intensive medical care, etc.) of the patients and to each other’s assessments.

No formal calibration or consensus session was conducted prior to the actual image assessment. The medical student received focused training to accurately identify lung lobes, recognise relevant pathologies, and quantify their extent. The two radiologists, with 14 and 6 years of chest CT experience, were briefed only on the specific pathologies to assess (according to the Fleischner Society’s glossary of terms) and the application of the CT scoring system. As the evaluation of the CT data began in 2023, the definitions of the radiological terms are based on the 2008 version of the Fleischner glossary [10].

Based on the semi-quantitative scoring system described by Pan et al., the extent of parenchymal involvement was assessed separately for each of the five lung lobes. Each lobe was assigned a subjective visual score ranging from 0 to 5 (Table 1). The individual lobar scores were summed to yield a total score ranging from 0 to 25.

In addition to the assessments performed by the three human readers, a fully automated score was generated for each patient using a dedicated AI-based software tool (ADVANCE chest CT, contextflow GmbH, Vienna, Austria; Figure 1). Consequently, four total scores per patient were available for comparative analysis.

The software employs a deep learning-based algorithm applied to lung kernel images to automatically segment the lungs and quantify parenchymal abnormalities. No manual adjustments to the AI-derived results were performed. According to the manufacturer, the algorithm was trained on a multicentric CT dataset; however, detailed information regarding the training process and model architecture is not publicly disclosed.

All statistical analyses and visualisations were performed using R (version 4.x; R Core Team, 2024) [11]. Descriptive statistics were calculated for all variables. Continuous variables are presented as the mean ± standard deviation (SD) or median with interquartile range (IQR), depending on their distribution, while categorical variables are reported as counts and percentages. Group comparisons between survivors and non-survivors were performed using the Wilcoxon rank sum test for continuous variables and the chi-square test or Fisher’s exact test for categorical variables, as appropriate.

Receiver operating characteristics (ROC) curve analyses were conducted to evaluate the discriminatory performance of the Pan score in predicting mortality and the need for mechanical ventilation. The area under the curve (AUC), along with the corresponding confidence intervals, was calculated using the pROC package [12]. Associations between clinical variables, mortality, mechanical ventilation, and ICU length of stay were assessed using correlation analyses based on Pearson’s correlation coefficient.

Logistic regression models were used to evaluate the association between the Pan score and clinically relevant outcomes, including mechanical ventilation and prolonged ICU stay (≥ 7 days). Both univariate and multivariable models were fitted, adjusting for potential confounders including age, sex, and laboratory parameters. Results are reported as odds ratios (OR) with 95% confidence intervals (CI).

To further explore the relationship between the Pan score and outcome probabilities, generalised linear models with a binomial distribution were fitted. To account for the potentially non-linear effect of age on the outcome, age was modelled using restricted cubic splines (with three degrees of freedom) in all multivariable regression models. This approach provides greater flexibility than assuming a linear relationship between age and the outcome and reduces the risk of residual confounding due to misspecification of the functional form. Predicted probabilities with 95% CI were derived and visualised across the range of Pan scores for cumulative outcome thresholds (e. g. ICU stay ≥ 1, ≥ 2, ≥ 3, and ≥ 7 days; ventilation duration ≥ 24, ≥ 48, ≥ 72, and ≥ 168 hours).

Data handling and transformation were performed using functions from the tidyverse collection [13]. Model outputs were processed using the broom package [14]. Visualisations were generated using ggplot2 [15], and tables were created using the knitr package [16]. Additional functionality was provided by the splines2, MASS, and nnet packages [17,18,19].

A two-sided p-value of < 0.05 was considered statistically significant.

3. Results

A total of 351 patients were included in the analysis: 223 males [63.5%], 128 females [36.5%], comprising 276 survivors and 75 non-survivors, corresponding to an overall mortality rate of 21.4%. The mean age of the cohort was 67.0 ± 15.9 years, with non-survivors being older on average than survivors (74.0 ± 11.8 vs. 65.1 ± 16.4 years; p < 0.001).

The mean Pan score (averaged across all raters) was 9.4 ± 5.3 points and was significantly higher in non-survivors than in survivors (11.7 ± 6.0 vs. 8.8 ± 5.0; p < 0.001). Consistent differences were observed across individual raters as well as the AI-based assessment.

Mechanical ventilation was required in 132 of 351 patients (37.6%), while 102 patients (29.1%) were admitted to the intensive care unit (ICU). Among those treated in the ICU, the mean length of stay was 7.0 days [3.2–12.0] and was significantly longer in non-survivors compared with survivors (10.5 [6.0–20.0] vs. 6.0 [3.0–10.0] days; p < 0.003).

Laboratory parameters showed only modest differences between survivors and non-survivors (Table 2). Laboratory parameters were not available for all patients. Specifically, erythrocyte and leukocyte counts were available for only 349 patients each, CRP values for 297 patients, and D-dimer values for only 155 patients.

Pulmonary function tests, which could only be performed in selected cases due to hygiene regulations during the COVID-19 pandemic, were conducted in only 16 patients within our cohort. Consequently, the statistical validity of these functional parameters is insufficient for meaningful analyses. Therefore, pulmonary function parameters are not presented in tabular form.

Pan scores of all four readers were aggregated to derive a single mean Pan score per patient and subsequently stratified by outcome (Figure 2). Non-survivors (n = 75) exhibited significantly higher Pan scores than survivors (n = 276), with mean values of 11.7 ± 6.0 and 8.8 ± 5.0, respectively (Wilcoxon rank sum test, p < 0.001).

However, the ability to discriminate all-cause mortality was limited, as demonstrated by ROC analysis. The Pan score (overall mean) achieved an area under the curve (AUC) of 0.639, indicating only modest predictive performance for all-cause mortality (Figure 3).

Correlation analysis demonstrated that mortality was only weakly associated with the Pan score (r = 0.22, p < 0.01), age (r = 0.23, p < 0.01), and selected laboratory parameters (leukocytes: r = 0.11, p < 0.045; CRP: r = 0.16, p = 0.007). In contrast, certain variables, such as D-dimer, showed no meaningful correlation.

The Pan score was markedly higher in patients requiring mechanical ventilation (n = 132) compared with those who were not ventilated (n = 219) (Figure 4). In the correlation matrix, the strongest association was observed between the Pan score and mechanical ventilation (r = 0.54, p < 0.01), followed by ICU length of stay (r =0.39, p < 0.01). In contrast, other markers such as CRP (r =0.12, p = 0.036) and D-dimer (0.2, p = 0.013) demonstrated considerably weaker correlations with the need for ventilation (Table 3).

ROC analysis for the prediction of mechanical ventilation using a logistic model based on the Pan score demonstrated good discriminatory performance, with an AUC of 0.804. The optimal probability threshold, determined by Youden’s index, was 0.51, yielding a sensitivity of 0.58, a specificity of 0.89, and an overall accuracy of 0.77 for predicting the need for mechanical ventilation (Figure 5).

Patients with prolonged ICU treatment (≥ 7 days) exhibited markedly higher Pan scores than those with shorter or no ICU stay (Figure 6). The Pan score demonstrated good predictive performance for an ICU stay of ≥ 7 days, with an AUC of 0.803 (95% CI: 0.733–0.873). The optimal cutoff, determined by Youden’s index, was approximately twelve points (11.9), corresponding to a sensitivity of 0.71, a specificity of 0.81, and an overall accuracy of 0.79 for identifying patients with prolonged ICU stays.

In univariate logistic regression, each one-point increase in the Pan score was associated with a 31% higher odds of requiring mechanical ventilation (p < 0.001). This association remained robust after adjustment for age, sex, CRP, D-dimer, leukocytes, and erythrocytes (p < 0.001).

Similarly, for an ICU stay ≥ 7 days, the univariate odds ratio per one-point increase in Pan score was 1.27 (95% CI: 1.19–1.36; p < 0.001), and the association persisted after multivariate adjustment (adjusted OR = 1.26, 95% CI: 1.12–1.46; p < 0.001) (Table 4).

The Pan score demonstrated a strong and consistent association with both the need for mechanical ventilation and the duration of ventilatory support. Patients requiring invasive ventilation had significantly higher Pan scores than those who were not ventilated.

In age-adjusted logistic regression analyses, the Pan score emerged as an independent predictor of mechanical ventilation across all examined duration thresholds. For ventilation lasting ≥ 24 hours, each one-point increase in Pan score was associated with a 33% increase in the odds of requiring ventilation (OR = 1.33, 95% CI: 1.24–1.44; p < 0.001). Comparable effect sizes were observed for longer ventilation durations (≥ 48 hours: OR = 1.31, 95% CI: 1.22–1.42; ≥ 72 hours: OR = 1.31, 95% CI: 1.22–1.42; ≥ 7 days: OR = 1.30, 95% CI: 1.20–1.42; all p < 0.001).

Discriminatory performance was consistently high across all thresholds, with areas under the curve ranging from 0.853 to 0.857. The highest performance was observed for ventilation ≥ 24 hours (AUC 0.857, 95% CI: 0.799–0.916), with only minimal variation across longer ventilation durations.

In an ordinal regression model treating ventilation duration as an ordered outcome (no ventilation [n = 219], < 24 hours [n = 79], ≥ 24 hours [n = 6], ≥ 48 hours [n = 5], ≥ 72 hours [n = 13], ≥ 7 days [n = 29], the Pan score showed a significant positive trend (β = 0.276, p < 0.001), indicating a stepwise increase in ventilation duration with rising Pan scores (Figure 7).

Overall, higher Pan scores at baseline CT were associated not only with a greater likelihood of subsequent mechanical ventilation but also with progressively longer durations of ventilatory support.

Predicted probabilities of ICU admission across all duration thresholds (≥ 1, ≥ 2, ≥ 3, and ≥ 7 days) increased monotonically with the baseline Pan score. Adjusted for the cohort’s median age of 69 years, a Pan score of 15 corresponded to a predicted probability of > 50% for an ICU stay of ≥ 7 days. As the duration thresholds increased, the probability curves showed a consistent rightward shift (Figure 8).

To further characterise these findings, a bivariate correlation analysis was performed between the mean Pan score across all readers and selected clinical, laboratory, and outcome parameters (Figure 9). Significant positive monotonic correlations were observed between the mean Pan score and markers of systemic inflammation and tissue injury, markedly LDH (ρ = 0.41, p < 0.001) and CRP (ρ = 0.29, p < 0.001). In addition, higher Pan scores were associated with increased healthcare resource utilisation, as reflected by longer ICU stays (ρ = 0.43, p < 0.001) and prolonged duration of mechanical ventilation (ρ = 0.43, p < 0.001). In contrast, significant negative correlations were found with respiratory parameters, including oxygen saturation (ρ = -0.24, p < 0.001) and arterial pO₂ (ρ = -0.13, p = 0.002). Moreover, higher Pan scores were associated with reduced overall survival (death after PCR: ρ = -0.37, p = 0.001). No significant correlations were identified for D-dimer levels, leukocyte count, or estimated glomerular filtration rate (eGFR).

4. Discussion

The aim of this study was to investigate the extent to which the semi-quantitative CT scoring system proposed by Pan et al. [8] for assessing the severity of COVID-19 pneumonia can predict the clinical course of affected patients. This involved assessing the correlation of the score with various laboratory parameters and the patients’ pulmonary function, as well as evaluating its ability to predict the need for mechanical ventilation, admission to intensive care, and survival.

The Pan score was determined independently by three different readers with varying levels of radiological experience, as well as by an AI-based software application, as described previously [9]. For the present study, the mean score of all four readers was used for comparison with other parameters in each case. Compared with other studies investigating the outcome of COVID-19 patients or the correlation of SARS-CoV-2-associated changes on chest CT with various clinical parameters [6,7,20,21,22,23], the patient cohort examined in our study (n = 351) was considerably larger.

The overall mortality rate in our cohort was 21.4%, which is consistent with the rates reported in other studies, ranging from 15.4 to 33.1% [6,7,22,23]. The COVID-positive patients who died were, on average, significantly older than those who were discharged from hospital alive. In addition, the deceased patients showed significantly higher Pan scores (11.7 ± 6.0 vs. 8.8 ± 5.0). These findings are consistent with the observations of Li et al., who demonstrated that the mortality rate in COVID-19 infections increases with higher CT scores [23]. Although they used an earlier CT scoring system to semi-quantitatively assess the extent of pulmonary involvement, it employs the same threshold values as the Pan score and, in terms of its assessment – based on separate evaluation of different lung segments – is also comparable to it [24].

Of the laboratory parameters examined, only erythrocytes (4.4 vs. 3.9 /pl) and CRP (4.6 vs. 7.3 mg/dL) showed significant differences between survivors and non-survivors; leucocytes and D-dimer did not differ significantly.

In our study, the predictive power of the Pan score with regard to all-cause mortality was only moderately pronounced. This contradicts the findings of Francone et al., who observed that the risk of death significantly increased at a CT score of ≥ 18, and to the observations of Szabó et al., who demonstrated that death was significantly associated with higher CT scores [6,7].

By contrast, the Pan score showed good discriminatory ability in predicting the need for mechanical ventilation, and furthermore, the score was strongly correlated with the duration of mechanical ventilation. The proportion of patients requiring mechanical ventilation was significantly higher among non-survivors than among survivors (56.0 vs. 32.1%), and patients who required invasive ventilation exhibited significantly higher Pan scores. Each additional point in the Pan score was associated with a 31% increase in the odds of requiring mechanical ventilation, and with regard to ventilation duration, the score showed a significant positive trend, with a stepwise increase in the duration of mechanical ventilation as scores rose. For ventilation lasting ≥ 24 hours, each additional point in the Pan score corresponded to a 33% increase in the odds of requiring ventilation.

The Pan score also showed good predictive performance in assessing the need for intensive care admission in patients with COVID-19 pneumonia, especially for ICU stays of ≥ 7 days. Both the rate of ICU admissions (42.7 vs. 25.4%) and the length of ICU stays (10.5 vs. 6.0 days) were significantly higher among non-survivors than among survivors. At a threshold of 11.9 points, the Pan score demonstrated an accuracy of 0.79 in identifying patients requiring prolonged intensive care. After adjustment for the cohort’s median age of 69 years, a Pan score of 15 was associated with a predicted probability of > 50% or an ICU stay of ≥ 7 days.

Among the laboratory parameters examined, only weak overall correlations were observed with the endpoints of mortality, mechanical ventilation, and intensive care admission: D-dimer showed a weak correlation with the need for mechanical ventilation (r = 0.20), while CRP and leucocyte count were likewise weakly correlated with patient mortality (r = 0.16 and 0.11, respectively). Pulmonary function parameters showed no meaningful correlation with the above-mentioned endpoints. Furthermore, pulmonary function tests were performed in only a small number of patients in our cohort, precluding any reliable statistical analysis. Patient age also demonstrated only a weak correlation with mortality (r = 0.23).

The main limitation of our study is its retrospective design. As a result, we were confronted, on the one hand, with different CT protocols, as patients with a confirmed COVID-19 infection predominantly underwent non-contrast chest CT to assess the extent of pulmonary involvement [25,26], whereas others received contrast-enhanced CT scans (e. g. for oncological staging), where COVID-19 was not the primary clinical concern. We had already confirmed in a previous study the hypothesis that patients undergoing non-contrast CT for symptomatic COVID-19 infection exhibit higher Pan scores than, for example, patients in oncological follow-up [9].

On the other hand, we had to deal with the fact that the laboratory parameters of interest were partially incompletely documented, and pulmonary function testing was performed in only a small number of patients. The latter is explained by the fact that, during the COVID-19 pandemic, pulmonary function testing was generally recommended only in emergency situations for hygiene reasons [27].

In recent years, the number of COVID-19 infections has declined markedly due to established vaccination programmes and the population immunity acquired through prior infection, meaning that COVID-19-related pulmonary changes are now observed only rarely. Nevertheless, further waves of transmission are still considered possible, and severe disease courses may continue to occur, particularly among older individuals and those with certain pre-existing conditions. Furthermore, antigenic shifts remain possible, which could temporarily lead to increased viral circulation and a higher disease burden once again [2].

Therefore, the assessment of semi-quantitative CT scores and their predictive value remains relevant, as these scoring systems may be adapted in the future for evaluating other pulmonary infectious diseases. Moreover, the observation that higher Pan scores are associated with increased healthcare resource utilisation may also be relevant when further evaluating therapeutic strategies and their effectiveness.

5. Conclusions

The semi-quantitative CT score investigated is a simple and reliable diagnostic tool for assessing the extent of COVID-19 pneumonia, which can be determined both subjectively by radiologists with varying levels of experience and fully automatically by AI-based software. The score showed only moderate predictive performance for all-cause mortality but performed well in predicting the need for and duration of mechanical ventilation, as well as intensive care requirement in affected patients.