Artificial intelligence-powered chest computed tomography analysis unveils prognostic insights for COVID-19 mortality among prevalent hemodialysis patients

Ethics statement

This study was conducted in accordance with the ethical standards of the responsible committee on human experimentation (institutional and regional) and with the Helsinki Declaration of 1975, as most recently amended. The Institutional Review Board of Hallym University Kangnam Sacred Heart Hospital granted ethical approval for this study on July 28, 2021 (No. HKS 2021-07-013). The board waived the requirement for informed consent for data collected as part of routine care.

Study design and participants

This retrospective cohort study included 326 patients on prevalent hemodialysis admitted with COVID-19 from December 1, 2020 to November 30, 2021. The data were collected from the electronic health records of Good Samaritan Bagae Hospital. Initially, 338 patients were screened with 12 excluded for not undergoing CT scans. Chest CT scans were performed immediately after hospital admission along with COVID-19 confirmation by real-time reverse transcription-polymerase chain reaction (PCR). AI-assisted CT analysis categorized lung involvement to inform prognosis. Patients were also monitored for secondary infections through cultures and additional PCR tests, which were conducted when clinically indicated.

Data collection

Demographic and clinical data were collected at admission, such as age, sex, body mass index (BMI), comorbidities, and initial symptoms. Notably, the use of renin-angiotensin system (RAAS) inhibitors was included among the baseline characteristics due to its clinical relevance. This inclusion was based on speculation that COVID-19 patients taking these agents may be at increased risk for adverse events, as angiotensin-converting enzyme 2 acts as a receptor for SARS-CoV-2 [21–23]. Oxygen saturation (SpO₂) and various inflammatory markers were also obtained as baseline laboratory data. Treatment-related issues such as the use of high-flow nasal cannula (HFNC), intensive care unit (ICU) admission, and mechanical ventilator placement were also recorded. Crucially, the presence of pleural effusion and B-type natriuretic peptide (BNP) levels were evaluated when available. These additional factors are indicative of volume overload and are considered significant in predicting the prognosis of dialysis patients [24–26]. Follow-up data, including clinical outcomes, were identified and extracted from the hospital’s medical records. The primary endpoint was the in-hospital mortality.

Computed tomography scan assessment

Radiological data was obtained by extracting information from CT scan reports of the study population. Initial CT scans were interpreted by a thoracic imaging specialist with over a decade of experience, further supplemented by teleradiology training and COVID-19 scoring, ensuring proficient interpretation of relevant CT scan parameters. The professional radiologist evaluated the presence and the extent of lung features such as nodular consolidation, ground-glass opacities, patchy infiltration, and pleural effusion. Particular attention was paid to the nature and distribution of nodular consolidations, patterns such as centrilobular, perilymphatic, and random patterns; the extent and distribution of ground-glass opacities and patchy infiltrations; and the quantification of pleural effusions by volume and location, adhering to standard radiological guidelines [27].

This comprehensive approach enabled detailed segmentation and quantification of each radiological feature. Nodular consolidations were identified and quantified to recognize areas of denser lung tissue, typically round or irregular in shape, associated with localized infection or inflammation. Ground-glass opacities were detected and measured to show areas of opacity that did not obscure underlying structures indicative of interstitial or alveolar edema. Patchy infiltrations were also delineated and quantified by marking areas of increased opacity suggestive of inflammatory or infectious processes.

The degree of lung involvement, considering total lung volume, was categorized using commercially available AI segmentation software (MEDIP PRO, version 2.0.0; MEDICAL IP Co., Ltd.). AI was only used to calculate lung involvement. AI analysis was not applied to nodular enhancement, ground-glass opacities, or patchy infiltrates, as these aspects were interpreted exclusively by a radiologist. The AI utilized an existing model specifically designed for lung pathology [28], and 3DnnU-Net [29], which is known for its accuracy in medical image segmentation, was recently updated and improved for the specific context of COVID-19 pneumonia [30].

We adopted a classification scheme from previous studies that categorized lung involvement as absent or minimal (<10%), moderate (10%–25%), widespread (25%–50%), severe (50%–75%), and critical (>75%) [31]. In our cohort, minimal lung involvement was observed in 267 patients, 10% to 25% in 46 patients, 25% to 50% in 12 patients, and 50% to 75% in one patient. Therefore, we set the threshold for extensive lung involvement at >10%. We also determined a more granular cutoff value for lesion volume through receiver operating characteristic (ROC) curve analysis to improve the precision of lung involvement classification.

Computed tomography scan protocol

The CT scan was performed using a helical full scan method. The specific scanning parameters were as follows: tube voltage, 120 kV; tube current, 260 mA; exposure time, 3.61 seconds; rotation time, 0.5 seconds; detector coverage, 40 mm; helical thickness (collimation), 1.25 mm; and slice interval, 1.25 mm. The CT scanners used in this study were the WCT-600-140/optima CT 600 and WCT-440-140/LightSpeed 16, both manufactured by GE Healthcare. When high-resolution CT was utilized, scans were performed without contrast media administration. Patients were placed in a supine position and asked to hold their breath at end-inspiration. This was done to minimize motion artifacts. Specifically, high spatial resolution settings were chosen to optimize the imaging of the lung parenchyma and small airways. The image reconstruction method used was a deep learning reconstruction type. In accordance with the protocol of the participating medical facilities, all patients and medical staff wore face masks and donned personal protective equipment to minimize the risk of cross-infection. After each patient examination, a strict decontamination process was performed in the scan area. Regular calibrations and quality assurance checks were conducted on the CT scanner to ensure data accuracy and reproducibility.

Statistical analysis

Statistical comparisons of clinical and CT features based on the degree of lung involvement were conducted using chi-square tests for categorical variables and independent t tests for continuous variables. The impact of CT findings on clinical outcomes and mortality was assessed using logistic regression analysis. Survival differences were analyzed using Kaplan-Meier curves, with significance tested via the log-rank test. Univariate and multivariate logistic regression analyses were performed to evaluate the odds ratios (ORs) for mortality specifically associated with CT findings. Lung involvement was analyzed using a cutoff value for lesion volume determined by ROC curve analysis. Model 1 included adjustments for age, sex, and BMI. Building on Model 1, Model 2 additionally adjusted for medical comorbidities such as diabetes mellitus, coronary artery obstructive disease, heart failure, cerebrovascular accident, arrhythmia, and malignancy. Model 3 was further enriched by including clinical parameters like plasma hemoglobin, serum albumin, C-reactive protein, and BNP, along with all factors from Model 2. All statistical analyses were performed using SPSS software (ver. 23.0; IBM Corp.). A p-value of less than 0.05 was considered indicative of statistical significance in the analysis.

Baseline characteristics of the patients

Our study included 326 hospitalized patients on prevalent hemodialysis who were diagnosed with COVID-19. Using AI-assisted CT analysis, we identified 267 patients with non-extensive lung involvement, and 59 patients with extensive lung involvement (>10%) (Fig. 1). The baseline characteristics of the study population are presented in Table 1. The overall mean age of the patients was 66.7 ± 12.6 years, 61.0% were male, and 58.6% had diabetes. Symptoms included fever in 42.9% of patients, followed by cough (26.1%), dyspnea (20.9%), sputum (14.7%), sore throat (10.1%), and runny nose (6.4%). At admission, 30 patients (9.2%) had an SpO₂ level of <95%. RAAS inhibitors were used by 43.3% of the patients.

Figure 1.

Flow chart of the study population.

Of the 338 hospitalized coronavirus disease 2019 (COVID-19) patients undergoing renal replacement therapy, 12 were excluded because they did not have a chest computed tomography (CT) scan. For the rest, artificial intelligence (AI)-assisted CT analysis identified 267 without extensive lung involvement and 59 with extensive involvement.

Table 1.

Baseline characteristics by lung involvement

Baseline characteristics between the groups with and without extensive lung involvement were compared using AI-enhanced CT analysis. No significant differences were observed in demographics or comorbidities such as diabetes, coronary obstructive disease, heart failure, cerebrovascular accident, arrhythmia, and malignancy. However, patients with extensive lung involvement exhibited a higher rate of sputum production (25.4% vs. 12.4%, p = 0.04) and a higher rate of with an SpO₂ level below 95% (30.5% vs. 4.5%, p = 0.02). Laboratory tests also showed that the group with extensive lung involvement had significantly higher mean white blood cell counts, C-reactive protein, and BNP levels, and lower serum albumin levels.

Chest computed tomography findings with lung involvement

Using AI-assisted CT analysis, we evaluated image findings based on the degree of lung involvement detailed in Table 2. Only 35.6% of patients with extensive lung involvement were effusion-free, compared with 74.5% in the group without extensive lung involvement (p < 0.001). Unilateral effusion was slightly more prevalent in the extensive group (13.6% vs. 8.6%, p < 0.001), and bilateral effusion was also significantly more frequent in this group (50.8% vs. 16.9%, p < 0.001). For nodular consolidation, a significantly higher incidence was observed in the group with extensive lung involvement compared to the non-extensive group (93.2% vs. 63.3%, p < 0.001). Similarly, there was a significant difference in the incidence of patchy infiltrates between extensive and non-extensive groups (89.8% vs. 49.8%, p < 0.001). In addition, ground-glass opacities were identified in 58.0% of the total population, with a significantly higher frequency observed in the extensive group (78.0% vs. 53.6%, p = 0.001).

Table 2.

Chest computed tomography findings with comparison of lung lesions by lung involvement

Correlation of chest computed tomography findings with critical care interventions

Logistic regression analysis showed that patients with extensive lung involvement were significantly more likely to require HFNC (OR, 3.01; 95% confidence interval [CI], 1.55–5.86), mechanical ventilation (OR, 5.15; 95% CI, 2.44–10.87) and ICU care (OR, 4.53; 95% CI, 2.48–8.24) (Fig. 2). Similarly, pleural effusion and nodular consolidation were significantly associated with an increased need for HFNC, mechanical ventilation, and ICU care. In contrast, patchy infiltration and ground-glass opacities were not significantly associated with the need for HFNC, mechanical ventilation, and ICU treatment.

Figure 2.

Logistic regression analysis for clinical outcomes.

The adjusted odds ratio for the association of lung findings in COVID-19 patients with intensive care unit (ICU) treatment, mechanical ventilation (MV), and high-flow nasal cannula (HFNC).

CI, confidence interval.

Mortality based on chest computed tomography findings

During the study period, 70 of 326 patients (21.5%) died. Survival analysis determined the impact of CT findings on patient survival, as shown in Fig. 3. Kaplan-Meier survival curves indicated a higher mortality risk in the group with extensive lung involvement (>10%) compared to those with non-extensive lung involvement. Similarly, patients with pleural effusion, nodular consolidation, and patch infiltration also had higher mortality rates. However, no significant difference in mortality rates was observed in patients with or without ground-glass opacities.

Figure 3.

Kaplan-Meier curves for mortality.

Survival curves of patients hospitalized with coronavirus disease 2019 (COVID-19) according to extensive lung involvement (A, p < 0.001), pleural effusion (B, p = 0.001), nodular consolidation (C, p = 0.001), patchy infiltrates (D, p = 0.031), and ground-glass opacities (E, p = 0.176).

Univariate and multivariate logistic regression analyses were performed to further evaluate the impact of chest CT findings on mortality. Initial analyses used a threshold of 10% or greater pulmonary involvement, but more refined thresholds were also explored. ROC curve analysis identified a cutoff value of 2.7% of lesion volume as the most accurate predictor of mortality with a sensitivity of 0.800 and specificity of 0.659 (Fig. 4).

Figure 4.

ROC curve for mortality based on lung involvement volume.

At a 2.7% lesion volume cutoff, sensitivity is 0.800 and specificity is 0.659. At a 10% lesion volume cutoff, sensitivity is 0.429 and specificity is 0.890

AUC, area under the curve; ROC, receiver operating characteristic.

Multivariate logistic regression analyses were performed by adjusting the following variables (Table 3): Model 1 adjusted for age, sex, and BMI; Model 2 additionally adjusted for comorbidities such as diabetes mellitus, coronary artery obstructive disease, heart failure, cerebrovascular accident, arrhythmia, and malignancy; and Model 3 included laboratory parameters such as plasma hemoglobin, serum albumin, C-reactive protein, and BNP. Lung involvement greater than 2.7% consistently demonstrated a significant impact on mortality across all models, with an OR of 16.70 (95% CI, 4.35–65.63) in Model 3. Pleural effusion was also significantly associated with increased mortality in Model 3, with an OR of 3.28 (95% CI, 1.15–9.35). Likewise, findings of nodular consolidation and patchy infiltrates were also linked with increased mortality, with ORs in Model 3 being 4.08 (95% CI, 1.12–14.82) and 3.75 (95% CI, 1.17–12.03), respectively. However, ground-glass opacities did not demonstrate a significant effect in any model.

Table 3.

Multivariable logistic regression analysis for mortality