Machine Learning Tool Predicts 6-Month Mortality in Cancer Patients

Roxanne Nelson, RN, BSN

October 28, 2019

SAN FRANCISCO — Machine learning algorithms may be useful in identifying patients with cancer who are likely to die within the next 6 months, and this assessment was concordant with that made by oncologists, a new study reports.   

It also prompted physicians to have more "serious illness conversations."

The authors note that 5 months after the intervention was implemented, the average monthly rate of serious illness conversation rose by 23%.

"When integrated into routine clinical practice, a machine learning-based decision support tool can increase rates of documented serious illness conversations," said lead author Ravi B. Parikh, MD, MPP, an instructor in medical ethics and health policy at the University of Pennsylvania and staff physician at the Corporal Michael J. Crescenz VA Medical Center, Philadelphia.

Parikh presented findings from the study here at the Supportive Care in Oncology Symposium (SCOS) 2019, and the results were simultaneously published in JAMA Network Open. The study was supported by grants from the National Institutes of Health and Penn Center for Precision Medicine.

"These findings need to be tested more formally in a randomized controlled trial, which we are doing at Penn right now," he commented.

Most Patients Do Not Have a Conversation About Serious Illness

Conversations about early advance care planning will generally result in care that is in agreement with patient goals and wishes, especially at the end of life, the authors comment.

However, most patients with cancer die without ever having a conversation about their treatment goals and end-of-life preferences — or at least do not have such a conversation documented, they add. In addition, most patients with cancer do not have hospice care.

"It's well known that prognostic awareness and accuracy can inform timely conversations about treatment goals and care preferences for patients with cancer," Parikh said. "Nevertheless, using currently existing tools, oncologists are only able to identify about 30% of patients who will die in the next year."

"Prognostic aids and algorithms exist, but they are usually limited to disease- or context-specific settings," he continued. Additionally, existing prognostic aids do not apply to all cancer types, are unable to identify most patients who will die within 6 to 12 months, and generally require time-consuming data input.

Good Predictability

Machine learning algorithms based on electronic health record (EHR) data have been able to identify patients at high risk of short-term mortality in general medicine settings, the authors note. In addition, machine learning oncology-specific algorithms have been accurately used to predict short-term mortality when chemotherapy is initiated.

"But amidst all the hype about machine learning, there are a lot of questions that need to be answered," Parikh pointed. "How do different learning models compare to each other? How do they perform in unselected cohorts like all patients you see in your clinics? And importantly, how are they going to be used clinically to improve outcomes for seriously ill patients?"

Parikh and colleagues hypothesized that a machine learning algorithm could accurately identify all patients at risk of short-term mortality and that oncologists would believe most patients identified as high risk were appropriate for a serious illness conversation.

The authors then developed, validated, and implemented a machine learning algorithm that could predict mortality in a general oncology setting using EHR data prior to a clinic visit.

The study involved 26,525 adults treated at a large academic cancer center and 10 affiliated community practices. Patients were observed for up to 500 days after the hematology/oncology encounter.

Three modeling approaches were used: logistic regression, gradient boosting, and random forest algorithms. The study's primary outcome was 180-day mortality following the first encounter, and the secondary outcome was 500-day mortality.

The cohort was randomly divided into a training (70%) and validation (30%) group, and the three algorithms were trained to predict 180-day mortality.

Of the three algorithms, positive predictive values were higher for the random forest (51.3%) and gradient boosting (49.4%) algorithms, compared with the logistic regression algorithm (44.7%). In the random forest model, the observed 180-day mortality was 51.3% among high-risk patients vs 3.4% in low-risk group. At 500 days, mortality was 64.4% in the high-risk group and 7.6% in the low-risk group.

In the final segment of the study, 15 providers at community-based hematology/oncology practices assessed 171 potential high-risk encounters. Of this patient group, 100 unique patients (58.8%) were indicated as being appropriate for a conversation about goals and preferences in the upcoming week. The mean predicted 6-month mortality risk of all high-risk encounters was 65%.

Exciting Work

During a discussion of the findings, Jean Kutner, MD, MPH/MSPH, of the University of Colorado School of Medicine, Aurora, noted that cancer is a stressful event. "It has significant psychosocial implications, and not surprisingly, national guidelines indicate that we have to integrate a psychosocial domain into our care," she said.

Despite guideline recommendations emphasizing their importance, "serious illness conversations are not happening routinely," Kutner said. "Only about one third of patients in their last year of life report having serious illness conversations, even though we keep saying how important it is."

The emergence of new technology is important for many reasons. One example is the current workforce shortage, she pointed out, which emphasizes the need to target resources and introduce innovative care delivery models.

"There has been increasing data application in the context of serious illness and a growing body of work that is using big data to help identify prognostication and identify populations we should be targeting," Kutner added. "I think there is some exciting work going on in this sphere."

Supportive Care in Oncology Symposium (SCOS) 2019: Abstract #131. Presented October 25, 2019.

JAMA Netw Open. Published October 25, 2019. Full text

For more from Medscape Oncology, join us on Twitter and Facebook.


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.