Machine learning algorithms can be used to identify cancer patients at risk of short-term mortality, according to a new study published in JAMA Network Open. This helps providers plan for necessary conversations about end-of-life preferences.
“Among patients with cancer, early advance care planning conversations lead to care that is concordant with patients’ goals and wishes, particularly at the end of life,” wrote lead author Ravi B. Parikh, MD, MPP, University of Pennsylvania in Philadelphia, and colleagues. “Nevertheless, most patients with cancer die without a documented conversation about their treatment goals and end-of-life preferences and without the support of hospice care. A key reason for the dearth of such conversations may be that oncology clinicians cannot accurately identify patients at risk of short-term mortality using existing tools.
The authors explored data from more than 26,000 adult patients who received care at one of 11 outpatient sites from the same health system from Feb. 1 to July 1, 2016. The 180-day mortality of each patient was noted using electronic health record (EHR) data. Four percent of the patients died by the end of the 180-day follow-up period. More patients who were alive at the end of that follow-up period were female, and they also had a younger mean age than the patients who died. There was “no significant difference in race.”
The team then trained three machine learning algorithms—one random forest, one gradient boosting and one logistic regression—with 70% of the patient cohort (more than 18,000 patients) and validated them with the other 30% (more than 7,000 patients). The accuracy of all three models was at least 95%, and the specificity was at least 98.9%. Using a prespecified alert rate, the positive predictive values (PPVs) for the random forest (51.3%) and gradient boosting (49.4%) AI models were higher than the logistic regression model (44.7%). No significant differences in the area under the ROC curves (AUCs) for the three models were noted after making certain adjustments.
“In this cohort study, machine learning models based on structured EHR data accurately predicted the short-term mortality risk of individuals with cancer from oncology practices affiliated with an academic cancer center,” the authors wrote. “The gradient boosting and random forest models had good PPV at manageable alert rates, and all machine learning models had adequate discrimination (ie, AUC, 0.86-0.88) in predicting six-month mortality.”
Parikh and colleagues explained that these findings could make a big impact on patient care, helping clinicians provide patients with the care they need and consider important discussions about end-of-life preferences.
“Machine learning algorithms can be relatively easily retrained to account for emerging cancer survival patterns,” the authors wrote. “As computational capacity and the availability of structured genetic and molecular information increase, we expect that predictive performance will increase and there may be a further impetus to implement similar tools in practice.”
The team did add that its algorithms were developed specifically with general medical oncology in mind, meaning they “may not be generalizable” to patients in radiation oncology, gynecologic oncology and other oncology settings.
“However, the features used in our models are all commonly available in structured data fields in most health systems EHRs,” the authors concluded.