# Prediction hospital mortality for critical illness lung cancer patients with pneumonia

## TL;DR

This study creates a prediction tool to assess hospital mortality risk in lung cancer patients with pneumonia, helping doctors make better care decisions.

## Contribution

The study introduces a novel, validated prediction model specifically for lung cancer patients with pneumonia, which is not available in existing literature.

## Key findings

- The prediction tool achieved high accuracy (C-index: 0.763) in predicting hospital mortality.
- The model outperformed several machine learning algorithms in performance.
- Survival analysis showed clear risk stratification between high- and low-risk groups.

## Abstract

Pneumonia is a common and severe complication in patients with lung cancer, often resulting in prolonged intensive care stays and increased risk of death. Despite this, no predictive models have been specifically developed for this high-risk population to aid clinical decision-making and early risk identification.

This study retrospectively analyzed patient data from two large critical care databases: one used for model development and the other for external validation. Adult patients with a diagnosis of lung cancer and pneumonia were included. Clinical features associated with in-hospital death were first screened using single-variable regression, and those with statistical significance were further refined using a variable selection method based on penalized regression. A visual prediction tool was then developed using multivariable regression analysis. Performance was evaluated using standard metrics of discrimination and calibration. Additional machine learning algorithms, including tree-based models, were used to compare performance. Survival analysis was conducted to assess risk grouping capability.

A total of 1046 patients were included in the final analysis. The visual prediction tool incorporated clinical features such as severity scores, mental status assessments, white blood cell count, blood gas indicators, and use of life-support measures. It demonstrated high predictive accuracy (C-index: 0.763) in the external test cohort. The tool outperformed several commonly used machine learning models. Survival curves showed a clear distinction between high-risk and low-risk groups. Calibration and decision analysis confirmed the tool’s clinical usefulness.

This study developed and validated a practical, interpretable prediction model for hospital mortality in patients with lung cancer complicated by pneumonia. The tool enables risk stratification and supports personalized clinical management in intensive care settings.

Not applicable.

The online version contains supplementary material available at 10.1186/s12879-025-12484-z.