📄 medicine

Development and preliminary external validation of a machine learning-based model for predicting 28-day mortality in non-HIV patients with Pneumocystis jirovecii pneumonia: a two-centre retrospective study

This two-centre retrospective study developed and preliminarily validated a LASSO logistic regression model using 13 routinely available clinical and laboratory variables that demonstrated promising performance in predicting 28-day mortality for non-HIV patients with Pneumocystis jirovecii pneumonia, though the findings require confirmation through larger prospective studies.

Original authors: fei yu, min shao, yousheng ye

Published 2026-08-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: fei yu, min shao, yousheng ye

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the complex landscape of human health, some infections strike with a particular ferocity in people whose immune systems have been weakened. One such infection is caused by a microscopic organism called Pneumocystis jirovecii, which lives quietly in the air but can turn into a severe pneumonia when a person's defenses are down. While this condition was once thought to be almost exclusively a threat to people living with HIV, doctors have observed a troubling shift: it is increasingly affecting patients with other serious health challenges, such as cancer, organ transplants, or autoimmune diseases. For these non-HIV patients, the illness often moves with startling speed, causing rapid breathing difficulties and a sudden drop in the body's ability to function. The central challenge for medical teams is knowing which patients are likely to survive and which are in immediate danger of dying within weeks, so they can act quickly. Yet, predicting this outcome is difficult because the body's reaction involves a tangled web of factors, from how well the lungs are working to how the blood is clotting and how the organs are handling stress.

To address this uncertainty, a team of researchers in China set out to build a digital tool capable of spotting the warning signs of a fatal outcome before it is too late. They focused on a specific group of adults who had been confirmed to have this pneumonia but did not have HIV. By looking back at the medical records of 165 patients treated at two major hospitals between 2020 and 2025, the team gathered a vast amount of information about each person's condition at the time of diagnosis. This data included basic details like age and sex, the severity of their illness as measured by standard scoring systems, the treatments they received, and a wide array of blood test results. The goal was to see if a computer program could learn to recognize the specific patterns in this data that distinguished the patients who survived for 28 days from those who did not.

The researchers tested several different mathematical approaches, known as machine learning algorithms, to see which one could best sort through the information. They compared methods that mimic human decision-making, those that build complex trees of choices, and others that look for subtle connections between variables. After running these different models against the data from the first hospital, they found that one specific approach, which uses a technique to filter out less important information, performed the best. This model identified thirteen key factors that were most closely linked to the risk of death. These factors included the patient's overall severity score, whether they were using drugs to support their blood pressure, the levels of certain proteins in their blood that indicate inflammation or clotting, and how well their lungs were exchanging oxygen.

When the team took this best-performing model and tested it on a separate group of 23 patients from a different hospital, the results were striking. The model correctly identified the patients who would survive with a high degree of accuracy, and while it achieved perfect specificity in identifying those who would not survive, its ability to detect all non-survivors was limited, correctly identifying only about two-thirds of them; however, the small size of this second group means the results should be viewed as a promising start rather than a final proof. The analysis also revealed which pieces of information mattered most to the computer's decision. It turned out that the level of a substance called lactate in the blood, which signals that the body's tissues are not getting enough oxygen, the level of a clotting marker called D-dimer, and the use of drugs to support blood pressure were the three most influential factors. These variables essentially told the model that the patient's body was under severe strain, struggling to circulate blood and maintain basic functions.

The study concludes that such a tool, built from routine medical tests and observations, holds real promise for helping doctors assess risk early in the course of the disease. However, the authors are careful to note that this work is preliminary. Because the study looked back at past records and involved a relatively small number of patients, the findings are best seen as a hypothesis that needs further testing. The model has not yet been proven to change how doctors treat patients or to improve survival rates in a real-world setting. Before this tool can be used to guide life-or-death decisions, it will need to be validated in much larger groups of people across different regions and healthcare systems. For now, it stands as a clear demonstration that the complex, chaotic signals of a severe infection can be organized into a coherent prediction, offering a potential new way to see the future of a patient's recovery.

Technical Summary: Machine Learning-Based Prediction of 28-Day Mortality in Non-HIV PJP

Problem Statement
Pneumocystis jirovecii pneumonia (PJP) is a severe opportunistic infection that, while historically linked to HIV, increasingly affects non-HIV patients with immunocompromising conditions (e.g., malignancies, organ transplantation, autoimmune diseases). Unlike HIV-associated PJP, non-HIV PJP is characterized by rapid clinical deterioration, severe hypoxemia, and high short-term mortality. Current prognostic assessments rely on clinical judgment or general severity scores, which may fail to capture the multidimensional nature of mortality risk involving host immunity, respiratory dysfunction, systemic inflammation, and treatment exposure. There is a lack of validated prediction models specifically designed for 28-day mortality in non-HIV PJP that incorporate external validation and interpretability.

Methodology
This study employed a two-centre retrospective design involving adult patients with clinically and microbiologically confirmed PJP diagnosed between October 2020 and August 2025 at two tertiary hospitals in Anhui Province, China.

Cohorts: Data from the First Affiliated Hospital of the University of Science and Technology of China served as the development cohort ( $n=142$ ), while data from the First Affiliated Hospital of Anhui Medical University served as the preliminary external validation cohort ( $n=23$ ).
Outcome: The primary endpoint was all-cause mortality within 28 days of diagnosis.
Feature Selection: Least Absolute Shrinkage and Selection Operator (LASSO) regression with 10-fold cross-validation was used to select candidate predictors from a broad set of demographic, clinical, treatment, and laboratory variables.
Model Development: Six machine learning algorithms were trained and compared in the development cohort: Random Forest, Logistic Regression, Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Decision Tree, and LASSO Logistic Regression.
Evaluation: Model performance was assessed using discrimination (Area Under the Receiver Operating Characteristic Curve, AUC), calibration, Decision Curve Analysis (DCA), and SHapley Additive exPlanations (SHAP) for interpretability. The final model was applied to the external validation cohort without refitting.

Key Contributions and Results

Predictor Selection: LASSO identified 13 candidate predictors: sex, APACHE II score, antiviral therapy, glucocorticoid therapy, vasoactive drug use, absolute monocyte count, blood urea nitrogen (BUN), alanine aminotransferase (ALT), D-dimer, lactate, procalcitonin (PCT), arterial partial pressure of oxygen ( $PaO_2$ ), and the $PaO_2$ /fraction of inspired oxygen ( $FiO_2$ ) ratio.
Model Performance (Development Cohort): Among the six algorithms, the LASSO Logistic Regression model demonstrated the highest discrimination with an AUC of 0.833 (95% CI: 0.744–0.922), outperforming standard Logistic Regression (0.828), SVM (0.825), Random Forest (0.815), XGBoost (0.698), and Decision Tree (0.697).
External Validation: In the independent external validation cohort, the LASSO Logistic Regression model achieved an AUC of 0.950 (95% CI: 0.773–1.000). It demonstrated a sensitivity of 0.667, specificity of 1.000, and accuracy of 0.957. The authors note that the wide confidence intervals and small sample size ( $n=23$ , 3 deaths) necessitate cautious interpretation of these high metrics.
Interpretability: SHAP analysis indicated that lactate, D-dimer, and vasoactive drug use were the most influential predictors, reflecting tissue hypoperfusion, coagulation activation, and hemodynamic instability, respectively.

Significance and Claims
The authors position this work as an exploratory, hypothesis-generating study rather than a definitive clinical implementation tool.

Clinical Utility: The study demonstrates that a model based on routinely available clinical and laboratory variables can effectively discriminate between survivors and non-survivors in non-HIV PJP. The integration of multiple clinical domains (severity, inflammation, coagulation, organ function) offers a more comprehensive risk assessment than single biomarkers.
Limitations and Scope: The paper explicitly acknowledges significant limitations, including the retrospective design, limited sample size, small external validation cohort, and geographic restriction to a single province in China. The authors state that the findings should be interpreted as preliminary due to the risk of overfitting and the instability of performance estimates in small cohorts.
Future Directions: The study concludes that larger, prospective, multicentre studies are required to confirm model robustness, assess generalizability across different healthcare systems, and evaluate the model's impact on clinical decision-making before routine implementation. The authors emphasize that the selected variables are predictors, not causal determinants, particularly regarding treatment-related factors.

Technical Summary: Machine Learning-Based Prediction of 28-Day Mortality in Non-HIV PJP

More like this