Chest computed tomography-based artificial intelligence-aided latent class analysis for diagnosis of severe pneumonia

Abstract

Background:

There is little literature describing the artificial intelligence (AI)-aided diagnosis of severe pneumonia (SP) subphenotypes and the association of the subphenotypes with the ventilatory treatment efficacy. The aim of our study is to illustrate whether clinical and biological heterogeneity, such as ventilation and gas-exchange, exists among patients with SP using chest computed tomography (CT)-based AI-aided latent class analysis (LCA).

Methods:

This retrospective study included 413 patients hospitalized at Xinhua Hospital diagnosed with SP from June 1, 2015 to May 30, 2020. AI quantification results of chest CT and their combination with additional clinical variables were used to develop LCA models in an SP population. The optimal subphenotypes were determined though evaluating statistical indicators of all the LCA models, and clinical implications of them such as guiding ventilation strategies were further explored by statistical methods.

Results:

The two-class LCA model based on AI quantification results of chest CT can describe the biological characteristics of the SP population well and hence yielded the two clinical subphenotypes. Patients with subphenotype-1 had milder infections (P <0.001) than patients with subphenotype-2 and had lower 30-day (P <0.001) and 90-day (P <0.001) mortality, and lower in-hospital (P = 0.001) and 2-year (P <0.001) mortality. Patients with subphenotype-1 showed a better match between the percentage of non-infected lung volume (used to quantify ventilation) and oxygen saturation (used to reflect gas exchange), compared with patients with subphenotype-2. There were significant differences in the matching degree of lung ventilation and gas exchange between the two subphenotypes (P <0.001). Compared with patients with subphenotype-2, those with subphenotype-1 showed a relatively better match between CT-based AI metrics of the non-infected region and oxygenation, and their clinical outcomes were effectively improved after receiving invasive ventilation treatment.

Conclusions:

A two-class LCA model based on AI quantification results of chest CT in the SP population particularly revealed clinical heterogeneity of lung function. Identifying the degree of match between ventilation and gas-exchange may help guide decisions about assisted ventilation.

Introduction

Since the global coronavirus disease (COVID)-19 outbreak in 2019, severe pneumonia (SP) has become a high-incidence public health issue.^[1] SP is a serious and even lethal systemic disease caused by local or diffuse lung parenchymal or/and interstitial infection and is clinically characterized by hypoxemia or hypotension.^[2,3] SP has a predilection for children aged <5 years and adults aged >70 years with comorbidities and immunodeficiency.^[4] Many patients with SP need intensive care and mechanical ventilation.^[5] However, the World Health Organization reports that lung protective mechanical ventilation was insufficient to support all SP patients with hypoxemia, especially patients with severe COVID-19.^[6] This may be why SP still accounts for 25–50% of in-hospital mortalities.^[7,8] According to empirical evidence, we propose that clinical and biological heterogeneity, such as ventilation and gas-exchange, exists among patients with SP, which can also be responsible for high mortality, even if given ventilatory therapy. Hence, we hypothesized that identifying the SP subphenotypes with different ventilatory and gas-exchange characteristics could be an efficient pathway to develop the optimal ventilatory strategy in such patients.

There is currently little literature describing the artificial intelligence (AI)-based diagnosis of SP subphenotypes, and the association of the subphenotypes with the ventilatory treatment efficacy and prognosis. Although there are pneumonia-scoring methods in clinics, they are not appropriate for identifying the clinical subphenotypes of SP in patients.^[9] Imaging modalities, such as computed tomography (CT) with higher spatial resolution, are useful for diagnosing pneumonia and quantifying lung lesion.^[2,10] Moreover, AI has greatly increased the value of medical imaging in clinical practice, such as diagnosing pulmonary diseases, assessing treatment efficacy, and predicting progress.^[11] Hence, images-based AI-aided analysis can further mine a deeper level of organismal information of patients, which is unrecognized by the naked eye.^[12,13] We assumed that chest CT-based AI-aided analysis could supply a large volume of quantitative and qualitative data rich in macro- and micro-pathological information.^[14] Thus, our first goal was to use chest CT-based AI-aided analysis to obtain the flood of data and information in patients with SP. Meanwhile, latent class analysis (LCA) is a data-driven, unbiased algorithm used to recognize acute respiratory distress syndrome (ARDS) subphenotypes in a large population by data characteristics.^[15,16] We hypothesized that LCA with unsupervised learning could analyze the subphenotypes of SP. Therefore, our second goal was to base on the above large data to proceed with LCA, thus identifying SP subphenotypes with clinical heterogeneity. Furthermore, we explored the clinical significance of the identification of the subphenotypes, including evaluation of short- and long-term survival. Chest CT is the most accurate and timely method in vivo for quantifying lung ventilation. Furthermore, arterial blood oxygen saturation (SaO₂) is an easily available indicator for reflecting the gas-exchange state of the lungs. Thus, we derived the lung function marker via them and further focused on assessing the association between subphenotypes and lung function.

Clinical heterogeneity in many diseases may be responsible for treatment failure. Identifying disease heterogeneity may pose significant opportunities for precision medicine.^[17,18,19] We postulated that identifying the clinical heterogeneity of SP would help achieve precise medicine. Hence, our final goal was to evaluate how to guide assisted ventilation for SP subpopulations with clinical heterogeneity with respect to ventilation and gas exchange.

Methods

Study design and patients

From June 1, 2015 to May 30, 2020, approximately 54,672 hospitalized patients aged >18 years hospitalized in the respiratory department or emergency department at Xinhua Hospital were diagnosed with pneumonia based on the presence of fever or cough, abnormal hemogram, and lung infiltration on chest radiographs, in which COVID-19 cases were not included. Of them, 582 patients progressed to SP, according to the Infectious Diseases Society of America/American Thoracic Society Consensus (IDSA/ATS) during their hospital stay. For SP diagnosis, patients met one of the major criteria, namely use of invasive mechanical ventilation and septic shock requiring vasopressors, or three of the minor criteria, namely respiratory rate ≥30 breaths/min; arterial partial pressure of oxygen (PaO₂)/fraction of inspiration oxygen ratio ≤250, and presence of multilobar infiltrates, confusion/disorientation, uremia (blood urea nitrogen level ≥20 mg/dL), or hypotension requiring aggressive fluid resuscitation, according to the simplified IDSA/ATS SP criteria.^[20] In total, 458 patients underwent chest CT within 48 h and routine blood tests within 24 h of SP diagnosis. Forty-five patients were excluded because of the deletion of the lung structure, bronchiectasis, or pulmonary tuberculosis. Finally, 413 patients were included in the study. The study flowchart is shown in Figure 1.

Figure 1.

Flow diagram for generation of the study population. CT: Computed tomography; SP: Severe pneumonia.

This retrospective study was approved by the Institutional Ethics Committee of Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine on December 2, 2022 (No. XHEC-D-2022-252). The need for informed consent was waived as all patient data were anonymously taken. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013).

Chest CT-image and AI analysis

CT scans were performed using Brilliance Ict (256-slice)(Philips Healthcare, The Netherlands), SOMATOM Definition (64-slice) (SIEMENS, Germany), and Optima CT680 (128-slice) (GE Healthcare, USA). The scan parameters are detailed in Supplementary Table 1, http://links.lww.com/CM9/C86. All CT images of patients were processed using a platform called uAI Research Portal (uRP, Shanghai United Imaging Intelligence Co., Ltd.) in DICOM format.^[21,22] This platform automatically outlined the lung-infected areas as regions of interests (ROIs) with a Dice coefficient of 91.6%. A radiologist with 6 years of experience reviewed each of the ROIs and made manual corrections, if necessary. AI quantification results were then generated, including the volume of the whole-lung, volume of the infected region, and the infected region percentages (infected region volume to whole-lung volume). The infected region percentages corresponding to a different range of CT values, categorized as <–750 hounsfield units [HU]; –750 HU to –300 HU; <–300 HU to 50 HU; and >50 HU. The non-infected lung region percentages were obtained, too. The calculations were performed using the following equation: Infected lung region percentages corresponding to different range of CT values = infected region lung volume corresponding to different range of CT values/the whole-lung volume. Non-infected lung region percentages = 1 – infected lung region percentages. Details are shown in Figure 2A. A radiologist with 10 years of experience used the AI quantification results to analyze the dominant infected areas in terms of vertical and lateral distribution and calculated the percentage of volume for non-infected lung region (P_{non-infected lung region}), which was defined as the proportion of the non-infected lung region of the whole-lung. Furthermore, the radiologist assessed traditional CT features, such as atelectasis and pleural effusion, and obtained CT quantitative data such as CT values of the intercostal muscles and thickness of subcutaneous adipose tissue, defined as the longitudinal diameter from the pectoralis muscle to the skin surface.^[14]

Figure 2.

The overall process of AI analysis in an instance. (A) The overall process of chest CT-based AI aided analysis in an instance, which were as following: importing chest CT images, automatic outlining of ROIs, quantitative analysis and outputting the results. (B) Data of analysis used, key statistical method, main statistical results and the potential clinical significance of results. Weights of the percentages of infected regional lung volumes corresponding to different CT values in subphenotypes analysis. (C) Weights of the percentages of infected regional lung volumes corresponding to different CT values in subphenotypes analysis. Subphenotype 1: Red spots; Subphenotype 2: Green spots; Both X-axis and Y-axis represented percentages of the infected lung volume in the whole lung volume; the cut-off value of the percentage of the infected lung region corresponding to CT values of the range from –750 HU to –300 HU was approximately 20% in the two subphenotypes, and the cut-off value of volume percentage of the infected lung region corresponding to CT values of the range from –300 HU to 50 HU was approximately 10% in the two subphenotypes. AI: Artificial intelligence; CT: Computed tomography; HU: Hounsfield unit; ROIs: Regions of interests.

Clinical data collection

Data on baseline information, laboratory characteristics, clinical information, comorbidity, treatment protocols, and clinical outcomes of all patients were collected from medical records. Baseline information included sex and age; laboratory examination results, such as those for procalcitonin (PCT), C-reactive protein (CRP), neutrophil percentage, lymphocyte percentage, and monocyte percentage; and arterial blood gas analysis, such as for PaO₂ and SaO₂. These data were collected on the same day when chest CT was performed and during the periods of without assisted ventilation. Additionally, we collected clinical information regarding pneumonia scores at admission, including Acute Physiology and Chronic Health Evaluation (APACHE) II, Sequential Organ Failure Assessment (SOFA), and Curb-65 scores, and comorbidities, including hypertension, diabetes, chronic obstructive pulmonary disease, heart disease, and malignancies. We also recorded later-stage clinical complications, such as ARDS and sepsis; treatment protocols used, such as admission to the intensive care unit; and modes of assisted ventilation, including only non-invasive ventilation, invasive ventilation, and early-stage non-invasive ventilation, but the abandonment of later-stage invasive ventilation. Clinical outcomes included 30-day, 90-day, in-hospital, and 2-year mortalities. Finally, with the help of AI quantification results and arterial blood gas analysis, we derived a lung function marker, which is the ratio of non-infected lung region percentages to SaO₂ (P_{non-infected lung region}/SaO₂ ratio).

Statistical analysis

Continuous variables with a normal distribution are presented as means (standard deviation), and those with a skewed distribution are presented as median (interquartile range). Categorical variables are presented as numbers (percentage, %). Missing data primarily included those for arterial blood gas analysis, which accounted for approximately 20% and was filled by the k-nearest neighbors algorithm.^[23] Among them, the original version was defined as the version excluding missing data, and the full version was defined as the version including missing data. Other missing data represented less than 3% and were excluded from the analyses.

Next, LCA was performed via Mplus v7.0 (https://www.statmodel.com).^[19] We only used baseline demographics (age and sex), laboratory variables (CRP and PCT), AI quantification results, traditional CT features and CT quantitative data as class-defining variables to perform LCA. We then added other clinical variables as class-defining variables, such as pneumonia scores and comorbidity, to perform LCA furtherly. We evaluated four types of models, including from one to four classes. We compared these four models to identify the optimal model using the Bayesian information criterion, entropy, and the Vuong–Lo–Mendell–Rubin likelihood ratio test.

When we determined the optimal classification model, the study participants were assigned to their most likely subphenotypes. We then compared their clinical characteristics, such as laboratory indicators, arterial blood gas analysis, comorbidities, later-stage clinical complications, and clinical outcomes using the t, chi-squared, or Mann–Whitney U test as appropriate. We further analyzed the performance of assessment indicators, including optimal classification and pneumonia scores, using Pearson’s χ² test. Finally, we used the Kaplan–Meier (K–M) method to analyze survival curves stratified by subphenotypes and treatment. Bar and line graphs were used to display the characteristics of distribution of the original and full versions of the P_{non-infected lung region}/SaO₂ ratio stratified by ventilation schemes and survival rate. Statistical significance was set at P values <0.05. Statistical analyses were performed using SPSS software (IBM Inc., Armonk, NY, USA).

Results

Determining SP subphenotypes

The overall idea of the study is presented in Figure 2B. The LCA model results are listed in Table 1. The AI quantification results-based LCA model demonstrated that the two-class model had the highest statistical significance among the four models (P <0.001). It also showed the highest entropy of 0.963 [Table 1]. We then combined AI quantification results with clinical variables to construct the LCA model further. The results showed that clinical variables addition did not improve the LCA model fit, with an entropy of 0.951–0.962. Hence, the two-class model using AI quantification results as the main class-defining variable was finalized as the ultimate classification model. Among 413 SP patients, 325 were assigned to subphenotype-1, with an average probability of 0.99, and 88, to subphenotype-2, with an average probability of 0.98, showing a strong trend in the subphenotype classification.

Table 1.

All the models of LCA based on different class-defining variables in the study.

Models	Number of classes	Bayesian information criteria	Entropy	P value	Number of patients assigned to each class
					1	2	3	4
AI quantification results-based LCA model	1	45,142.181	–	–	413
	2	43,943.997	0.963	<0.001	325	88
	3	43,500.764	0.951	<0.001	257	37	119
	4	43,177.548	0.961	<0.001	254	119	37	3
LCA model of combination AI quantification results with clinical variables	1	53,225.105	–	–	413
	2	52,063.782	0.962	<0.001	326	87
	3	51,649.682	0.951	<0.001	257	37	119
	4	51,358.765	0.961	<0.001	254	118	38	3

AI: Artificial intelligence; LCA: Latent class analysis.

All class-defining variables are listed in Supplementary Table 2, http://links.lww.com/CM9/C86. AI quantification results, such as those on infected lung volume, its percentage, and the percentage of infected lung volume corresponding to different CT values, played an important role in classification decisions. As seen in the Figure 2C, the two SP subphenotypes were clearly distinguished by the percentage of infected lung region corresponding to CT values of the range from –750 HU to <–300 HU and from –300 HU to 50 HU. Values of AI quantification results were higher in subphenotype-2 patients than in subphenotype-1 patients, indicating that patients with SP in subphenotype-2 experienced more severe infection.

Clinical characteristics and implication in SP subphenotypes

Clinical characteristics and outcomes comparison of the SP subphenotypes are shown in Table 2, which also listed the original and full versions of the SaO₂ and P_{non-infected lung region}/SaO₂ ratios. Subphenotype-2 patients had higher neutrophil percentages (P <0.001) and lower lymphocyte (P = 0.01) and monocyte percentages (P <0.001) than subphenotype-1 patients. The oxygenation level of subphenotype-2 patients was significantly lower than that of subphenotype-1 patients (P <0.001 or P = 0.001). Excluding patients with active abandoning invasive ventilation, only one-fifth of subphenotype-1 patients received invasive ventilation compared with one-third of subphenotype-2 patients.

Table 2.

Clinical characteristics and outcomes comparison based on subphenotypes classified by LCA results.

Clinical variables	Total population (N = 413)	Subphenotype-1 (N = 325)	Subphenotype-2 (N = 88)	P value
Laboratory indicators, median (Q1, Q3)
Neutrophil percentage	84.70 (75.93, 92.00)	82.80 (72.93, 89.10)	88.45 (84.35, 92.85)	<0.001
Lymphocyte percentage	8.15 (4.50, 13.90)	9.45 (5.03, 15.20)	6.25 (3.68, 9.03)	0.010
Monocyte percentage	5.70 (3.10, 8.33)	6.10 (3.60, 8.58)	3.90 (1.73, 5.88)	<0.001
Arterial blood gas analysis, median (Q1, Q3)
PaO2 (original version) (Kpa)	8.43 (6.67, 10.20)	8.74 (7.05, 10.50)	7.56 (5.98, 09.10)	0.002
PaO2 (full version) (Kpa)	9.85 (8.10, 11.14)	10.14 (8.53, 11.29)	8.43 (6.54, 9.87)	<0.001
SaO2 (original version) (%)	94.30 (88.00, 96.45)	95.00 (90.15, 97.00)	90.00 (84.00, 95.00)	0.001
SaO2 (full version) (%)	94.50 (91.33, 96.21)	95.00 (93.18, 96.40)	88.5 (83.35, 93.73)	<0.001
Derived lung function marker, median (Q1, Q3)
Pnon-infected lung region/SaO2 (original version)	0.91 (0.79, 0.98)	0.94 (0.89, 1.01)	0.70 (0.59, 0.78)	<0.001
Pnon-infected lung region/SaO2 (full version)	0.92 (0.82, 0.98)	0.94 (0.89, 0.99)	0.72 (0.60, 0.77)	<0.001
Clinical complications, n (%)
Sepsis	121 (29.30)	96 (29.50)	25 (28.40)	0.840
ARDS	316 (76.50)	245 (75.40)	71 (80.70)	0.300
Treatment protocols
Assisted ventilatory support				0.002
Non-invasive, n (%)	260 (63.00)	219 (67.40)	40 (45.50)
Invasive, n (%)	93 (22.50)	65 (20.00)	28 (31.80)
Early-stage non-invasive, but active abandonment of later-stage invasive, n (%)	61 (14.80)	41 (12.60)	20 (22.70)
ICU admission, n (%)	207 (50.10)	156 (48.00)	51 (58.00)	0.100
Clinical outcome, n (%)
30-day mortality	98 (23.70)	55 (16.90)	43 (48.90)	<0.001
90-day mortality	136 (32.90)	86 (26.50)	50 (56.80)	<0.001
Intra-hospital mortality	165 (40.00)	116 (35.70)	49 (55.70)	0.001
2-year mortality	189 (45.80)	134 (41.20)	55 (62.50)	<0.001

Original version was defined as the version with excluding missing data; Full version was defined as the version with filling missing data. Subphenotype-1 can be refered to as mild, with less inflammation within the lungs, and the patients have milder clinical symptoms; and subphenotype-2 can be refered to as severe, with more diffuse inflammation within the lungs , and patients are in a more serious condition. ARDS: Acute respiratory distress syndrome; IQR: Interquartile range; PaO₂: Arterial partial pressure of oxygen; P_{non-infected lung region}/SaO₂: The percentage of non-infected lung region volume/arterial blood oxygen saturation; SaO₂: Arterial blood oxygen saturation.

Supplementary Figure 1, http://links.lww.com/CM9/C86 shows patients’ K–M survival curve for 30-day, 90-day, intrahospital, and 2-year mortality rates. The 30-day (χ² = 39.036, P <0.001) and 2-year (χ² = 12.621, P <0.001) mortality rates of subphenotype-2 patients were significantly higher than those of subphenotype-1 patients. Therefore, subphenotype-2 patients had more severe disease and worse prognoses than subphenotype-1 patients. Meanwhile, we also found that the current pneumonia scores had equally good performance in evaluating patients’ short- and long-term mortality [Supplementary Table 3, http://links.lww.com/CM9/C86].

Furthermore, we found that there were significant differences in P_{non-infected lung region}/SaO₂ ratios (original version) between the two subphenotypes (P <0.001) [Figure 3A]. Namely the P_{non-infected lung region}/SaO₂ ratios (original version) were higher in subphenotype-1 patients than in subphenotype-2 patients. But there were no significant correlations between P_{non-infected lung region}/SaO₂ ratios (original version) and the APACHE II (R = 0.06, P = 0.51), SOFA (R = 0.12, P = 0.23), and Curb-65 pneumonia scores (R = 0.03, P = 0.69) [Figure 3B–D].

Figure 3.

Statistical analysis between P_{non-infected regional lung}/SaO₂ ratios and SP subphenotypes as well as pneumonia scores. The figure shows that there was an unexpectedly significant difference in the P_{non-infected lung region}/SaO₂ ratios (original version) between two subphenotypes (A), but P_{non-infected lung region}/SaO₂ ratios (original version) had no significant association of APACHE II (B), SOFA (C), and CURB-65(D). APACHE II: Acute Physiology and Chronic Health Evaluation II; CURB-65: Confusion, Urea, Respiratory rate, Blood pressure, and Age ≥65 years; P_{non-infected lung region}/SaO₂ ratios: Ratio of non-infected lung region percentages to SaO₂; SOFA: Sequential Organ Failure Assessment.

Distribution trend of derived lung function markers among different efficacies of ventilatory support

The 30-day (P_{non-invasive ventilation} = 0.01, P_{invasive ventilation} <0.001; Figure 4A) survival curves showed a significant difference between subpopulations stratified by classification and treatment. The 2-year survival curves showed a significant difference in subpopulations after receiving invasive ventilation (P_{invasive ventilation} <0.001; Figure 4B), but they did not show a significant difference after non-invasive ventilation (P_{non-invasive ventilation} = 0.14; Figure 4B). After receiving invasive ventilation, patients survived for 30 days or 2 years; their P_{non-infected lung region}/SaO₂ ratios of original (P = 0.035, P = 0.020) and full versions (P = 0.002, P = 0.088) was higher than those of non-survivors, with most of the results showing statistical significance [Figure 4C, D]. P_{non-infected lung region}/SaO₂ ratios of both versions in patients stratified by treatment and survival outcomes displayed similar distribution patterns. When the ratio of P_{non-infected lung region} and SaO₂ was approximately 1, patients obtained significant survival benefits with invasive ventilation.

Figure 4.

Kaplan-Meier survival curves analysis of SP population stratified by subphenotypes and treatment. (A) There were significant differences in the 30-day survival of the two subphenotypes populations not only after receiving non-invasive ventilatory support, but also after receiving invasive ventilatory support by Kaplan–Meier analysis. (B) Significant differences in the 2-year survival of the two subphenotypes populations after receiving invasive ventilatory support, but no significant differences after receiving non-invasive ventilatory support. Comparison of P_{non-infected lung region}/SaO2 ratios of 30-day and 2-year survivors and non-survivors. Considering results including the P_{non-infected lung region}/SaO₂ ratios of the original (blue) and full (green) versions between survivors and non-survivors presenting a normal distribution, with all the bar and line graphs represented as mean (C, D). (C) Distribution trend of P_{non-infected lung region}/SaO₂ ratios of the original (blue) and full (green) versions between 30-day survivors and non-survivors after receiving invasive ventilation, and there were statistical significances in P_{non-infected lung region}/SaO₂ ratios both of original (blue) version (P = 0.035) and of full (green) version (P = 0.002) between 30-day survivors and non-survivors after receiving invasive ventilation. (D) Distribution trend of P_{non-infected lung region}/SaO₂ ratio of the original (blue) version between 2-year survivors and non-survivors after receiving invasive ventilation; there were statistical significances in P_{non-infected lung region}/SaO₂ ratio of original (blue) version (P = 0.020) between them, but no statistical significances in one of the full (green) version (P = 0.088). SP: Severe pneumonia.

Discussion

Our study presented three important findings with clinical implications for patients with SP. First, LCA of patients with SP yielded two stable and distinct subphenotypes. Second, the derived pulmonary function markers had significant differences in two subphenotypes, revealing the heterogeneity of ventilation and gas-exchange among patients with SP. Third, with the aid of the derived pulmonary function marker, patients who had a better match degree for lung ventilation and gas-exchange were more suited to invasive ventilation management, thus improving their prognosis.

The uRP platform and LCA were key to the present study.^[13,19] Studies have proved that the AI software adopted in the study is characterized by high accuracy and can be generalized to clinical practice.^[21,24] As biological subphenotypes of ARDS may not reflect differences in alveolar inflammation,^[25] we innovatively combined chest CT-based AI-aided diagnosis with LCA to achieve successful classification. We found that using CT value of the range from –750 HU to 50 HU was the most effective in recognizing the two clinical subphenotypes, and there was a significant positive association between SP subphenotypes and the percentage of neutrophils representing disease severity.^[26] Thus, we obtained two subphenotypes in SP population using imaging metrics associated with clinical variables.

We found that SP subphenotypes could evaluate short- and long-term mortality rates. Radiomics studies have shown that high-throughput data extracted from images can be used to diagnose diseases and evaluate clinical course and prognosis.^[27] Considering that AI quantification results extracted from images harbor micro-pathological information, there has been an understandable result of LCA models based on AI quantification results to evaluate patients’ mortality. We further compared the efficacy of related pneumonia scores in evaluating them and found that the SP subphenotypes analysis was comparable to pneumonia scores in evaluating them.^[28,29]

Additionally, we found that derived (rather than directly obtained) pulmonary function markers showed significant differences in two subphenotypes. Percentages of non-infected lung region were quantitatively obtained by calculation of AI quantification results, which can reflect the ventilation status of normal lung tissue to some extent. Also, SaO₂ would reflect their gas-exchange status accordingly. Considering that SaO₂ was influenced by the collected time of blood gas and treatment, we rigidly selected eligible SaO₂, which was acquired by arterial blood gas analysis on the same day of CT examination and during periods of without assisted ventilation. Hence, the P_{non-infected lung region}/SaO₂ ratios can precisely reflect the degree of matching between patients’ ventilation and gas-exchange. Currently, no marker can directly reflect this matching. A similar marker, the ventilation/perfusion ratio, is a semi-quantitative marker initially obtained by ventilation/perfusion scintigraphy of nuclear medicine.^[30] The radiation exposure risk in patients who underwent nuclear medicine examination was markedly higher than in those who underwent CT examination. Although new CT/magnetic resonance imaging (MRI) technology can also provide some lung function biomarkers, its implementation in clinical practice has largely lagged due to relatively high costs and imaging complexity.^[31] In contrast, the P_{non-infected lung region}/SaO₂ ratio proposed in this study is a precise, safe, convenient, and easy-to-access biomarker for clinical practice, which was a pleasant surprise. Biomarkers are a cornerstone in the management of patients with SP, which was well supported in the literature.^[32] We also performed correlation analyses between pneumonia scores and the derived marker but found no statistical significance. Thus, SP subphenotypes were significantly superior to markers such as pneumonia scores for evaluating pulmonary function.

We tried to explore the possibility of providing precision medicine based on SP subphenotypes in such a population.^[33,34] Hypoxemia is a key indicator of needing ventilation support in clinical practice.^[35] But not all patients with hypoxemia improve well after assisted ventilation. Our results showed that the P_{non-infected lung region}/SaO₂ ratios of original and full versions for 30-day or 2-year survivors were higher than those for non-survivors after invasive ventilation. Namely, patients with SP whose lung ventilation and gas-exchange matched relatively well may benefit more from assisted invasive ventilation than those with poor matching of lung ventilation and gas-exchange.^[36] Hence, we concluded that patients with P_{non-infected lung region}/SaO₂ ratios closer to 1 may be more appropriate for invasive ventilation strategies.

We used a computer-assisted tool in this study, aiming to support the image analysis of clinicians. There may be questions on the security and service availability. In a recent study, an interesting work was proposed with the implementation of a service-oriented networks (SON) service within a clinical scenario.^[37] In particular, service continuity can be enhanced by designing self-organized systems, with a non-fixed structure, able to change their structure and organization as well as adaptively react to internal and external environment changes. In SONs, nodes communicate with each other by means of stimulation or suppression chains that give rise to emergent behaviors to defend against foreign invaders, attacks, and malfunctions. In the clinical environment, continuity of a service, such as the one used in this work, is an absolutely mandated requirement.

Our study has several strengths. First, LCA using AI quantification results of chest CT is an effective method whose model contributes to the stability of the results. We identified SP subphenotypes with the heterogeneity of ventilation and gas-exchange. Based on the SP subphenotypes, we searched for a novel marker to guide precise medical treatment. Nevertheless, this study also has a few limitations. This was a retrospective study. In clinical practice, prospective studies are more suitable for enrichment strategies of SP subphenotypes. Meanwhile, a larger sample size would yield more accurate results. Although our sample size was limited. The trend of our results was clear. Additionally, arterial blood gas analysis data were partly missing due to the requirements for their acquisition time. Given the limitations and the importance of the available data, we have made some efforts. We listed the original and full versions of SaO₂ and P_{non-infected lung region}/SaO₂ ratios, respectively, and showed the similar distribution trend of both compellingly in the results of subsequent analyses.

LCA based on AI quantification results of chest CT in the SP population identified two stable and reliable subphenotypes, one of which had a mild infection and good prognosis compared to the other. When the SP population showed a good degree of matching between ventilation and gas-exchange, their survival can be improved with invasive ventilation.

Funding

This study was supported by the National Natural Science Foundation of China (Nos. 82172138 and 81873947).

Conflicts of interest

None.

Supplementary Material

cm9-138-2316-s001.docx ^{(132.2KB, docx)}