Clin-STAR Journey Story

Improving Medical Care Using Prediction Models: A Geriatrician’s Approach to Algorithms

W. James Deardorff, MD, is an Assistant Professor in the Division of Geriatrics at the University of California, San Francisco. He was introduced to geriatrics during medical school where he was impressed by how geriatricians primarily focused on improving the quality of life for this complex population of older adults. This patient-centered approach immediately resonated with him. “The default setting in medicine is to do more—try another treatment, add this new medication—because it might help reduce risk,” he said. “A geriatrician’s mindset is that, sure, we can start a medication or a new treatment, but are we really helping our patients? Or are we just increasing pill burden and making it harder for a caregiver to take care of them? What other options can we explore to improve the things they care about?”

Based on his experience working at a skilled nursing facility, Dr. Deardorff sought to measure these patient-centered outcomes using prediction models. In 2023, he received an NIA-funded grant and became a GEMSSTAR Scholar to develop a comprehensive prognostic model for older adults discharged to skilled nursing facilities after hospitalization. “Patients oftentimes feel blindsided when they come to a skilled nursing facility,” he said, because most expect to go back home and that “things are going to be fine”. Unfortunately, about 20% of these patients get readmitted to the hospital and another 15% transition to a nursing home long term. “For older adults who have been living at home independently, the last thing on their mind is that they’ll end up in a nursing home.”

His research aims to provide clinicians with an easy-to-use web calculator to obtain the probability of various outcomes after admission to a skilled nursing facility, such as hospital readmission, successful discharge home, long-term care transition, and 6-month mortality. With useful prognostic calculations, Dr. Deardorff explains, clinicians can “help frame conversations with patients and families. 'Well, based on other people like you, you’re someone who has a high chance of transitioning to a nursing home long term. So, here’s what we can start doing now to maybe get you back home, whether that be more tailored caregiver training or making home modifications.’ It’s primarily about managing expectations and establishing appropriate care plans.”

To design a robust model, certain logistical hurdles must be factored in. For example, older adults represent a quite heterogenous population. “In pediatrics,” Dr. Deardorff explained, “you have well defined stages—infant, toddler, childhood, and adolescence. Older adults get lumped into a single category when the reality is quite different. We have 95-year-olds living at home alone, independent, still driving, and very functional,” he said, “but we also have 65-year-olds with multiple medical conditions, on dialysis and with heart failure, and needing a lot of support from family. This is a heterogeneous group both in terms of their risk of adverse outcomes and their values and preferences.” Dr. Deardorff addressed some of these methodological concerns as first author on an article featured in the Around the EQUATOR with Clin-STAR series in JAGS. As a physician-scientist who became interested in clinical prediction models while on rotations in medical school, he cites three main issues to consider when developing prediction models for older patients. First is the competing risk of death. Dr. Deardorff explained that older adults have a high risk of death, and this needs to be factored in statistically to correctly estimate their risk of other conditions. Second, certain groups of older adults are underrepresented in clinical trials, which primarily enroll healthy patients. “People with frailty or multi-morbidity are frequently excluded from longitudinal cohorts and clinical trials,” he explained, “so when you start developing prediction models, if the patients aren’t well represented, then it’s hard to apply these models to those groups.” The third consideration extends beyond prediction models for older patients but can directly affect older patients, namely that prediction models can exacerbate inequities in health care. “The concept of algorithmic bias has been increasingly recognized in the literature,” he concluded, “so we need to make sure that we’re not using these models in a way that perpetuates harm.”