Post-Treatment Partial Pressure of Carbon Dioxide for Prognostic Stratification and Non-Invasive Ventilation Guidance in Elderly Patients with Community-Acquired Pneumonia and Heart Failure: The "Subgroup with Impaired Cardiac Response”

Figures

• 

  Figure 1. PCO2_post-Guided Stratification in Elderly CAP with Heart Failure Mechanism: Machine learning identified 72h-post-admission PCO2 as a superior prognostic marker compared to PSI scores. Stratification: A threshold of 47 mmHg effectively splits patients into distinct prognostic groups. High Risk (\ge 47≥47 mmHg): Associated with catastrophic mortality outcomes (HR 23.55); NIV shows limited benefit. Low Risk (< 47<47 mmHg): Patients demonstrate favorable survival and derive significant clinical benefit from Non-Invasive Ventilation (NIV), including reduced heart failure recurrence and faster symptom improvement. Conclusion: PCO2_post monitoring enables precise, cost-effective resource allocation, identifying the specific subgroup (Low Risk) that benefits most from NIV.

• 

  Figure 2. Visualization of the feature selection process using the Boruta algorithm and recursive elimination with SHAP importance. (A) Boruta Algorithm Graphical Display of Feature Selection Result. The blue box indicates the minimum, average, and maximum Z-values for shadow attributes. The red, yellow, and green boxes represent the Z-scores for rejected features, features awaiting confirmation, and confirmed features, respectively. It is evident that the Z-scores for the most significant shadow attributes clearly distinguish important from non-important attributes. However, the yellow (pending confirmation) and blue (maximum) values are close, necessitating further evidence for confirmation. (B) Evolution of Z-Score for Attributes During Boruta Runs. Green lines correspond to confirmed features, red indicates rejected features, yellow denotes features awaiting further confirmation, while blue represents the importance of the minimum, average, and maximum shadow attributes respectively. (C) Feature Selection Process via Recursive Elimination with SHAP Importance. This figure illustrates the recursive feature elimination (RFE) process used to identify the optimal subset of predictive…

• 

  Figure 3. Collinearity analysis of predictor variables. Variance Inflation Factor (VIF) values for the four variables selected by SHapley Additive exPlanations-based Recursive Feature Elimination (SHAP-RFE): respiratory failure (RespFailure), post-treatment partial pressure of carbon dioxide (PCO2_post), Pneumonia Severity Index score (PSIScore), and symptom improvement time (SymptomImproveTime). All VIF values were below 5, indicating absence of significant multicollinearity among predictors.

• 

  Figure 4. Receiver operating characteristic curves comparing the predictive performance of post-treatment PCO2 versus PSI score for 6-month mortality. (A) Training cohort (n=70) and (B) validation cohort (n=30) predictive performance. PCO2_post demonstrated an area under the curve (AUC) of 0.968 (95% CI: 0.928–1.000) in the training set and 0.986 (95% CI: 0.963–1.000), while the PSI score had AUC values of 0.960 (95% CI: 0.912–1.000) and 0.932 (95% CI: 0.861–1.000) in the training and validation sets, respectively. No statistically significant differences were observed between the two models across both datasets (all comparison P-values >0.05). The diagonal represents the performance of a random classifier (AUC =0.50).

• 

  Figure 5. Calibration of Post-Treatment PCO2 Model for 6-Month Mortality Prediction. (A) Training cohort (n=70). (B) Validation cohort (n=30). The x-axis shows predicted probabilities, and the y-axis shows observed probabilities. Ideal calibration is marked by the 45° dashed line. The training set exhibited strong calibration (Brier score 0.041, intercept 0.000, slope 1.000). The validation set also showed good calibration (Brier score 0.064), albeit with slight underestimation at higher probabilities (intercept 0.155, slope 0.757). Nonparametric curves with 95% confidence intervals are displayed. Hosmer-Lemeshow tests were non-significant for both cohorts (P = 0.570 for training, P = 0.655 for validation), indicating proper model fit.

• 

  Figure 6. Decision Curve Analysis of PCO2_post and PSIScore Predictive Utility. (A) (Training Set): The PCO2post model (red curve) occupies the highest position across the vast majority of threshold probabilities (0–1.0), demonstrating a markedly higher net benefit than the PSI Score (blue curve). Both models show greater clinical utility than the “All” and “None” strategies at low to intermediate thresholds (approximately 0–0.7), indicating that both possess clinical value, though the superiority of the PCO2post model is more pronounced. Near the threshold of 1.0, the red curve shows a brief rebound followed by a decline, which may be associated with the model’s predictive behavior for very high-risk samples within the training set. (B) (Validation Set): The overall trend remains consistent with the training set, with the PCO2_post model still outperforming the PSI Score. However, in the high-threshold region (approximately 0.8–1.0), the red curve shows a distinct decline, followed by a sharp peak around 0.9. This suggests slight instability in the model’s sensitivity at very high-risk thresholds during external validation. The PSI Score curve is generally flat and yields lower net…

• 

  Figure 7. Distribution of PCO2_post with Outlier Detection (Statistical boundaries:[35, 51], Clinical boundaries: [30, 45]). This histogram shows PCO2_post values for elderly pneumonia and heart failure patients, with blue for survivors and red for non-survivors. The x-axis is PCO2_post (mmHg) and the y-axis is frequency. Statistical limits are [35, 51] mmHg (blue lines), and clinical limits are [30, 45] mmHg (red lines). Patients exceeding 51 mmHg (red line) all died, linking high PCO2_post to poor outcomes. Lower mortality was seen below 45 mmHg. This highlights PCO2_post as a key predictor, identifying high-risk patients needing intervention.

• 

  Figure 8. Performance Metrics for PCO2_post Thresholds. This graph evaluates the sensitivity and specificity across various PCO2_post thresholds. The red line indicates sensitivity, while the blue line represents specificity. The optimal thresholds are marked: 47 mmHg (green dashed line, ROC method) and 48 mmHg (blue dashed line, Youden index). An optimal range of 47–50 mmHg balances sensitivity and specificity effectively.

• 

  Figure 9. Kaplan-Meier Survival Analysis. (A) Survival probability by PCO2_post risk stratification. The red line represents the high-risk group (PCO2_post ≥47 mmHg), and the blue line indicates the low-risk group (PCO2_post <47 mmHg). A significant survival difference is observed (P <0.0001), with the high-risk group experiencing a rapid decline in survival probability. (B) Survival probability by non-invasive ventilation (NIV) application. The red line shows patients without NIV, and the blue line shows those with NIV. No significant survival difference is found (P = 0.37), suggesting NIV does not impact survival probability in this cohort. The number at risk for each stratum is provided below each time point, illustrating the decreasing sample size over time.

• 

  Figure 10. Clinical Outcomes by Risk Group and NIV Treatment Status. Acute heart failure (AHF) incidence (Ai/ii), days to symptom improvement (Bi/ii), and hospitalization costs (Ci/ii) are compared between patients who did not receive non-invasive ventilation (NIV) and those who did, stratified into low-risk (Ai, Bi, Ci) and high-risk (Aii, Bii, Cii) groups. Red dots denote median values.