Doctors often find it challenging to accurately forecast the outcomes of pneumonia patients and identify the best treatments. However, recent research utilizing advanced machine-learning techniques on patients’ electronic health records (EHRs) has revealed five unique clinical states related to pneumonia.
Patients being treated for pneumonia, which causes breathing difficulties due to fluid in the lungs, can exhibit very different characteristics and outcomes. Still, physicians frequently find it hard to predict patients’ prognoses accurately and decide the optimal treatments.
Researchers from Northwestern University have recently applied a cutting-edge machine-learning methodology to analyze EHRs of pneumonia patients, identifying five distinct clinical states. Among these, three states are closely tied to patient outcomes, while two assist doctors in deciphering the underlying causes of the illness. Notably, one state indicated a 7.5% risk of mortality within 24 hours.
The findings and the methods used to develop this approach are detailed in a paper published in the journal Proceedings of the National Academy of Sciences (PNAS). The researchers believe that this method could enable clinicians to make more informed decisions about treating severely ill patients and have wider applications.
Pneumonia is a leading global cause of mortality and is notoriously challenging to treat due to its varied presentations and how it is contracted, alongside the risks of antibiotic overuse. Traditionally, doctors classified pneumonia patients in intensive care based on its origin into three groups: community-acquired (which could stem from a prior bacterial or viral infection), hospital-acquired, and ventilator-associated (contracted after mechanical ventilation).
However, according to LuÃs Amaral, the lead author of the study, this classification provides limited insight into a patient’s recovery chances.
“Alternative methods for categorizing pneumonia patients lack precision,” Amaral remarked. “They perform poorly in predicting the progression of the disease and outcomes, which is critical for end-of-life decision-making. Our research is the first to reveal clearly identifiable and distinct clinical states.”
Amaral, who specializes in complex systems and data science, holds the Erastus Otis Haven Professorship in Engineering Sciences and Applied Mathematics at Northwestern’s McCormick School of Engineering.
Amaral emphasized that understanding each patient’s likelihood of survival can help prepare their families for possible loss and assist doctors in avoiding unnecessary treatments.
The five identified states are based on a variety of data (including body temperature, breathing rate, glucose and oxygen levels) to understand the interrelationships among various measurements. The researchers discovered that specific combinations of indicators like motor response, kidney function, heart rate, blood pressure, and respiratory rate yielded the most significant insights into a patient’s condition.
The research team faced several obstacles while developing a set of machine-learning tools to analyze patient data from two EHR sources: one from Northwestern’s SCRIPT initiative and the other from a standard clinical database. They had to integrate various data types collected at different frequencies, create a new metric to assess the method’s reliability, and determine whether the physiological data could be simplified into a smaller number of effective combinations.
This analysis led to the identification of five different clusters—each representing a unique clinical state—showing a markedly improved ability to predict patient mortality compared to existing methods. Interestingly, one of these clusters included a significant number of patients whose pneumonia was linked to a COVID-19 infection.
Furthermore, the technical advancements achieved during this study may have broader implications. Feihong Xu, the lead author and a graduate student in Amaral’s lab, mentioned that the research team is “now applying these techniques to experimental data from a mouse model of sepsis.”
Currently, their research does not yet address the reasons why some patients transition between states, which is an area they are actively exploring. Ongoing investigations into pneumonia and other illnesses could ultimately pave the way for more effective and predictable treatment strategies.
