Appendix: Content Summaries of Selected Best Papers for the 2024 IMIA Yearbook, Section
Cancer Informatics
Manz, Christopher R., Yichen Zhang, Kan Chen, Qi Long, Dylan S. Small, Chalanda N.
Evans, Corey Chivers et al.
Long-term effect of machine learning-triggered behavioral nudges on serious illness
conversations and end-of-life outcomes among patients with cancer: A randomized clinical
trial.
JAMA oncology 9, no. 3 (2023): 414-418.
doi: 10.1001/jamaoncol.2022.6303
In this study the investigators evaluate the utility of a mortality prediction algorithm
combined with behavioral nudges on end-of-life care among cancer patients. Using a
randomized control trial including 20,506 patients with cancer across 9 clinical practices,
participates who were identified as having an 10% or higher risk for death in 6 months
were randomized to either receive behavioral nudges or standard of care. Behavioral
nudges consisted of weekly emails, lists, and texts to clinicians prompting them to
have serious illness conversations with high-risk patients. The authors found that
their machine learning based behavioral intervention led to 10% increase in serious
illness conversations and 3% decreasing use of end-of-life systemic therapy. This
study is among the few randomized control trials of machine learning interventions
within oncology. Additionally, the study represents a potential model for how to successfully
implement medical informatics solutions through behavioral interventions.
Placido, D., Yuan, B., Hjaltelin, J.X., Zheng, C., Haue, A.D., Chmura, P.J., Yuan,
C., Kim, J., Umeton, R., Antell, G. and Chowdhury, A.
A deep learning algorithm to predict risk of pancreatic cancer from disease trajectories.
Nature Medicine, 29(5), (2023), pp.1113-1122.
doi:10.1038/s41591-023-02332-5
The authors present a framework for predicting the risk of a rare cancer type by applying
deep learning to a real-world longitudinal dataset of disease trajectories. Using
longitudinal disease codes the investigators trained a deep learning model on 8,110,706
Danish patients (23,985 pancreatic cancers) and predicted the risk of developing cancer
in time intervals ranging from 6-36 months. The model was validated on 2,962,383 US
based patients (3,869 pancreatic cancers). The authors argue their model could be
helpful in identifying high risk cohorts of patients who potentially could be screened
more aggressively for rare cancer types. The authors model improved in performance
when looking at larger intervals prior to cancer diagnosis suggesting longer disease
histories do have value in predicting the risk of developing rare cancer types. Given
the model input is administrative disease codes, it is likely able to be implemented
at scale in a variety of clinical settings.
