# Revolutionizing oncology care: pioneering AI models to foresee pneumonia-related mortality

## TL;DR

This study uses AI to predict pneumonia-related deaths in cancer patients, identifying key risk factors and showing how early intervention could improve outcomes.

## Contribution

A novel AI model (CatBoost) is developed and validated for predicting pneumonia-related mortality in cancer patients with high accuracy.

## Key findings

- The CatBoost model outperformed other models in predicting survival rates with AUC scores ranging from 0.689 to 0.8384.
- Surgery was the most significant factor affecting prognosis for both short- and long-term survival.
- Digestive system cancers were most commonly associated with pneumonia-related deaths.

## Abstract

Pneumonia is a leading cause of morbidity and mortality among patients with cancer, and survival time is a primary concern. Despite their importance, there is a dearth of accurate predictive models in clinical settings. This study aimed to determine the incidence of pneumonia as a cause of death in patients with cancer, analyze trends and risk factors associated with mortality, and develop corresponding predictive models.

We included 26,938 cancer patients in the United States who died from pneumonia between 1973 and 2020, as identified through the Surveillance, Epidemiology, and End Results (SEER) program. Cox regression analysis was used to ascertain the prognostic factors for patients with cancer. The CatBoost model was constructed to predict survival rates via a cross-validation method. Additionally, our model was validated using a cohort of cancer patients from our institution and deployed via a free-access software interface.

The most common cancers resulting in pneumonia-related deaths were prostate (n=7300) and breast (n=5107) cancers, followed by lung and bronchus (n=2839) cancers. The top four cancer systems were digestive (n=5882), endocrine (n=5242), urologic (n=5198), and hematologic (n=3104) systems. The majority of patients were over 70 years old (57.7%), and 54.4% were male. Our CatBoost model demonstrated high precision and accuracy, outperforming other models in predicting the survival of cancer patients with pneumonia (6-month AUC=0.8384,1-year AUC=0.8255,2-year AUC=0.8039, and 3-year AUC=0.7939). The models also revealed robust performance in an external independent dataset (6-month AUC=0.689; 1-year AUC=0.838; 2-year AUC=0.834; and 3-year AUC=0.828). According to the SHAP explanation analysis, the top five factors affecting prognosis were surgery, stage, age, site, and sex; surgery was the most significant factor in both the short-term (6 months and 1 year) and long-term (2 years and 3 years) prognostic models; surgery improved patient prognosis for digestive and endocrine tumor sites with respect to both short- and long-term outcomes but decreased the prognosis of urological and hematologic tumors.

Pneumonia remains a major cause of illness and death in patients with cancer, particularly those with digestive system cancers. The early identification of risk factors and timely intervention may help mitigate the negative impact on patients’ quality of life and prognosis, improve outcomes, and prevent early deaths caused by infections, which are often preventable.