PHILADELPHIA— Across the United States, no hospital is the same. Equipment, staffing, technical capabilities, and patient populations can all differ. So, while the profiles developed for people with common conditions may seem universal, the reality is that there are nuances that require individual attention, both in the make-up of the patients being seen and the situations of the hospitals providing their care.
New research shows that artificial intelligence may improve care overall by combing through different hospitals' data to create more refined groups of patients similar to the local populations that hospitals are actually seeing. AI can help pinpoint typical care needs, such as what specific departments and care teams are required to meet patient needs. Led by researchers at the Perelman School of Medicine at the University of Pennsylvania, the project—whose findings were published in Cell Patterns—analyzed electronic health records of long-COVID patients, revealing a collection of four patient sub-populations—such as those with asthma or mental health conditions—and their specific needs.
"Existing studies pool data from multiple hospitals but fail to consider differences in patient populations, and that limits the ability to apply findings to local decision-making," said Yong Chen, PhD, a professor of Biostatistics and the senior author of the study. "Our work offers the benefit of more generalized knowledge, with the precision of hospital-specific application."
The study team used a machine learning artificial intelligence technique called "latent transfer learning", to examine de-identified data on long-COVID patients pulled from eight different pediatric hospitals. Through this, they were able to call out four sub-populations of patients who had additional conditions on top of long COVID, termed "comorbidities." These four included:
- Mental health conditions, including anxiety, depression, neurodevelopmental disorders, and attention deficit hyperactivity disorder
- Atopic/allergic chronic conditions, such as asthma or allergies
- Non-complex chronic conditions, like vision issues or insomnia
- Complex chronic conditions, including those with heart or neuromuscular disorders
With those sub-populations identified, the system was also able to track what care patients required across the hospital, pointing toward updated profiles that would allow hospitals to better address increases in different patient types.
"Without identifying these distinct subpopulations, clinicians and hospitals would likely provide a one-size-fits-all approach to follow-up care and treatment," said the study's lead author, Qiong Wu, PhD, a former post-doctoral researcher in Chen's lab who now is an assistant professor of biostatistics at the University of Pittsburgh School of Public Health. "While this unified approach might work for some patients, it may be insufficient for high-risk subgroups who require more specialized care. For example, our study found that patients with complex chronic conditions experience the most significant increases in inpatient and emergency visits."
The latent transfer learning system directly pulled out the effects these populations had on hospitals, pointing to exactly where resources should be allocated.
If the machine learning system had been in place around March 2020, Wu believes that it might have provided some key insight to mitigate some of the effects of the pandemic, including focusing resources and care on the groups most likely in need.
"This would have allowed each hospital to better anticipate needs for ICU beds, ventilators, or specialized staff—helping to balance resources between COVID-19 care and other essential services," Wu said. "Furthermore, in the early stages of the pandemic, collaborative learning across hospitals would have been particularly valuable, addressing data scarcity issues while tailoring insights to each hospital's unique needs."
Looking past crises such as the COVID-19 pandemic and its aftermath, the AI system developed by Wu, Chen, and their team could help hospitals manage much more common conditions.
"Chronic conditions like diabetes, heart disease, and asthma often exhibit significant variation across hospitals because of the differences in available resources, patient demographics, and regional health burdens," Wu said.
The researchers believe the system they developed could be implemented at many hospitals and health systems, only requiring "relatively straightforward" data-sharing infrastructure, according to Wu. Even hospitals not able to actively incorporate machine learning could benefit, through shared information.
"By utilizing the shared findings from networked hospitals, it would allow them to gain valuable insights," Wu said.
This study was supported in part by grants from the National Institutes of Health (U01TR003709, U24MH136069, RF1AG077820, 1R01LM014344, 1R01AG077820, R01LM012607, R01AI130460, R01AG073435, R56AG074604, R01LM013519, R56AG069880, R21AI167418, R21EY034179) and the Patient-Centered Outcomes Research Institute (PCORI) Project Program Awards (ME-2019C3-18315 and ME-2018C3-14899).
