When physicians have conversations with their patients regarding advanced care early during treatment, it results in patients receiving care that is in line with their goals and wishes. While guidelines recommend conversations with high-risk patients regarding advanced care, many patients with cancer die without ever having had this conversation. Because oncologists are only able to identify approximately 30% of patients with cancer who will die in the next year, it is difficult to determine which patients should have advanced care conversations. At the Supportive Care Oncology Symposium, Ravi Parikh, MD (University of Pennsylvania, Philadelphia, Pennsylvania, United States), presented results from a study using a machine learning based algorithm to identify patients with cancer who are most likely to die within the next 6 months. The results were simultaneously published in JAMA Network Open.
The study used electronic health records (EMR) from 26,525 patients with cancer to develop a machine learning algorithm that identified patients at high-risk of short-term mortality. Three different algorithms were tested in the study: a random forest alorightm, a gradient boosting algorithm, and a logistic regression model. Of the three, the random forest algorithm had the highest positive predictive value (51.3%, versus 49.4% for gradient boosting and 44.7% for logistic regression). Patients were sorted into high-risk and low-risk cohorts based on results from the random forest algorithm. The 180-day mortality was 51.3% for high-risk patients and 3.4% for low-risk patients. Similarly, the 500-day mortality was 64.4% for high-risk patients and 7.6% for low-risk patients. When 15 oncologists were asked to evaluate 171 patients who had been identified as high risk based on the gradient boosting algorithm, the oncologists identified 100 patients (58.8%) as appropriate for conversations about end-of-life preferences.
In his conclusion, Dr Parikh said that these data support machine learning algorithms as one method to accurately identify patients with cancer who are at high risk of short-term mortality. These models might be useful in identifying high-risk patients and facilitating timely conversations with high-risk patients regarding treatment goals and advanced care preferences.
Read more about this article on Medscape Medical News.