There’s nothing more frustrating than patients who don’t comply with follow-up imaging recommendations. But a new study in JACR not only identifies the factors that can lead to patient non-compliance, it also points the way toward IT tools that could predict who will fall short – and help direct targeted outreach efforts.
The new study focuses specifically on incidental pulmonary nodules, a particularly thorny problem in radiology, especially as CT lung cancer screening ramps up around the world.
- Prevalence of these nodules can range from 24-51% based on different populations, and while most are benign, a missed nodule could develop into a late-stage lung cancer with poor patient survival.
Researchers from the University of Pennsylvania wanted to test a set of 13 clinical and socioeconomic factors that could predict lack of follow-up in a group of 1.6k patients who got CT scans from 2016 to 2019.
- Next, they evaluated how well these factors worked when fed into several different types of homegrown machine learning models – precursors of a tool that could be implemented clinically – finding …
- Clinical setting had the strongest association in predicting non-adherence, with patients seen in the inpatient or emergency setting far more likely skip follow-up compared to outpatients (OR=7.3 and 8.6)
- Patients on Medicaid were more likely to skip follow-up compared to those on Medicare (OR=2)
- On the other hand, patients with high-risk nodules were less likely to skip follow-up compared to those at low risk (OR=0.25)
- Comorbidity was the only one of the 13 factors that was not predictive of follow-up
The authors hypothesized that the strong association between clinical setting and follow-up was due to the different socio-demographic characteristics of patients typically seen in each environment.
- Patients in the outpatient setting often have access to more resources like health insurance, transportation, and health literacy, while those without such resources often have to resort to the emergency department or hospital wards when they become sick enough to require care.
In the next step of the study, the data were fed into four types of machine learning algorithms; all turned in good performance for predicting follow-up adherence, with AUCs ranging from 0.76-0.80.
The Takeaway
It’s not hard to see the findings from this study ultimately making their way into clinical use as part of some sort of commercial machine-learning algorithm that helps clinicians manage incidental findings. Stay tuned.
