A recent study explored the use of a machine learning algorithm to predict short-term mortality in cancer patients. This knowledge could help patients have important discussions regarding treatment and end-of-life preferences with providers and family members.
Regarding the importance of their study, the authors explained, “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. 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 goal of the study, the results of which appeared in JAMA Network Open, was to create and compare three machine learning models designed to predict estimated six-month mortality in cancer patients, and to evaluate the technology for ease of use in a community oncology practice to identify patients who could benefit from a timely discussion.
This cohort study included adult patients with outpatient oncology or hematology/oncology encounters at a large academic center and 10 affiliated community practices; observation took place for up to 500 days following the encounter. The study exposures were logistic regression, gradient boosting, and random forest algorithms. The primary outcome measure was 180-day mortality from the index encounter, and the secondary outcome was 500-day mortality.
Of the 26,525 patients included in the study, 1,065 (4%) died during the 180 period following the index encounter; mean age of patients who died was 67.3 (95% CI, 66.5–68.0) years, and just under half (47%, n = 500) were women. The remaining patients had a mean age of 61.3 (95% CI, 61.1–61.5) years, and most (n = 15,922, 62.5%) were women.
These patients were randomized into either a training (n = 18,567, 70%) or validation (n = 7,958, 30%) group; an encounter was chosen at random to include in either of the sets. The positive predictive value was highest in the random forest algorithm (51.3%), followed by the gradient boosting (49.4%) and logistic regression (44.7%) algorithms. Discrimination did not largely differ among the three models: random forest area under the receiving operating characteristic curve (AUC), 0.88; 95% CI, 0.86-0.89; gradient boosting AUC, 0.87; 95% CI, 0.85-0.89; and logistic regression AUC, 0.86; 95% CI, 0.84-0.88 (P for comparison=0.02).
The researchers reported an observed 180-day mortality in the random forest model of 51.3% (95% CI, 43.6%-58.8%) in the high-risk and 3.4% (95% CI, 3.0%-3.8%) in the low-risk group.
Fifteen oncology clinicians were surveyed and said that of 171 patients deemed high risk by the gradient boosting algorithm, more than half (n = 100, 58.8%) were considered appropriate to take part in a treatment and end-of-life preferences discussion within the upcoming week.
The study authors concluded, “Our findings suggest that [machine learning] tools hold promise for integration into clinical workflows to ensure that patients with cancer have timely conversations about their goals and values.”