Sophisticated Machine-Learning Approach Uses Patient EHRs to Predict Pneumonia Outcomes

By HospiMedica International staff writers
Posted on 30 Oct 2024

Image: The machine-learning approach using health records identifies five pneumonia ‘states’ with predictable prognoses (Photo courtesy of 123RF)

Pneumonia, an infection that results in difficulty breathing due to fluid accumulation in the lungs, is one of the leading causes of death worldwide. This condition is particularly challenging to treat because of the various ways it can manifest and be contracted, along with the risk of antibiotic overuse. Two patients suffering from pneumonia can present very differently and may experience contrasting outcomes. Traditionally, physicians have classified pneumonia patients in intensive care units based on the cause of the infection into three categories: community-acquired (which may follow a prior bacterial or viral infection), hospital-acquired, and ventilator-associated (developing after mechanical ventilation). However, this classification often provides minimal insight into a patient’s likelihood of recovery, making it difficult for doctors to accurately predict prognoses and determine the best treatment strategies. Now, a novel approach could assist clinicians in making more informed treatment decisions for critically ill patients and may have broader applications.

Researchers at Northwestern University (Evanston, IL, USA) have employed a sophisticated machine-learning method on electronic health records (EHRs) from pneumonia patients to identify five distinct clinical states. Three of these states are closely linked to patient outcomes, while the other two aid physicians in determining the cause of the disease. One identified state correlates with a 7.5% chance of mortality within 24 hours. Understanding individual survival probabilities can help prepare family members for the potential loss and guide physicians in avoiding unnecessary treatments. The research team faced multiple challenges while developing a suite of machine-learning tools to cluster patient conditions from two EHR data sources: one from Northwestern’s SCRIPT project and another from a standard clinical dataset.

First, they had to integrate various data types that were collected at different frequencies. Additionally, they needed to devise a new test to evaluate the reliability of their approach. Finally, they had to assess whether the information from these physiological variables could be condensed into fewer combinations. This analysis allowed the researchers to identify five distinct clusters—equating to different clinical states—that significantly outperformed current methods in predicting patient mortality. These five states incorporate a range of data (such as body temperature, respiratory rate, glucose levels, and oxygenation) to reveal relationships between different measures.

The study, published in the journal Proceedings of the National Academy of Sciences (PNAS), shows that linear combinations of variables reflecting motor response, renal function, heart rate, systolic blood pressure, and respiratory rate provided the most insight into patient status. Notably, one of the identified clusters primarily consisted of patients whose pneumonia was linked to COVID-19 infections. The technical advancements achieved during this research may have applications in other areas. In fact, the team is currently applying these methodologies to experimental data from a mouse model of sepsis. They have yet to explore why certain patients transition between states, a topic they are now investigating. Future research on pneumonia and other diseases may ultimately lead to more effective and predictable treatment options.