Outcomes of Utilizing a Mortality Prediction Tool in the Community Oncology Setting

Ping Ye, PhD, The US Oncology Network, McKesson, The Woodlands, TX developed a machine learning model to predict 90-day mortality risk among patients with metastatic cancer, aiming to enable earlier Advance Care Planning (ACP) discussions and increase hospice enrollment. Dr Ye assessed the tool usage on ACP documentation in a community oncology setting.

Transcript

My name is Ping Ye. I'm a senior data scientist with the US Oncology Network. We're very pleased to present our abstract. The title of our abstract is the initial outcome of deploying a machine learning tool in a commuting oncology sighting. We're very pleased that our abstract was selected as a poster presentation, and we drew quite a crowd yesterday when we presented at the poster session.

I will start with giving a background of our study. So aligning care at end of life with a patient's values and goals helps reduce the likelihood of undesirable hospital stay, ER visit, and side effect of therapy. So our group has developed a machine learning model to predict the mortality risk in the next 90 days for patients with metastatic cancer.

The tool was designed to help physicians to proactively initiate end of life discussion with identified patients. Therefore, allowing patients have access to the care that align with their value as they near the life's end stages.

So the abstract we presented at this year's ASCO 2022 reports the initial outcome of deploying this mortality prediction tool in a real-world oncology setting. Specifically in a community-based oncology setting. The outcome we examined is called advanced care planning, abbreviated as ACP. So advanced care planning is a conversation for the physician to have with the patient to talk about the aggressive nature of the disease, to talk about identify someone who can help you make decisions when you're unable to do so. To talk about care available to manage symptoms and improve life quality. For example, hospice and palliative care.

So we deployed the two in a real-world oncology setting, and we looked at the utilization of ACP ways and without the mortality to deployment. So I can go into the specific methodology. In total, 12 practices within the US Oncology Network were included in our study. Five practices deployed are mortality prediction tool and incorporated in their clinical workflow.

The other seven practice didn't participate, and, therefore, served as a control for our study. Then over a period of 11 months of mortality to deployment, we made predictions for patients in the 12 practices on a biweekly basis to provide insight on their mortality risk in the next 90 days. And then, these results were delivered to the hands of the primary care oncologist to provide insight, help them with the decision-making.

And then for the patients who were predicted to at risk for mortality. We looked at their advanced care planning documentation as obtained from electronic health record and client data.

Over a period of 11 month, we made predictions for over 11000 patients. Among them, 1600 patients were predicted at high risk for mortality in the next 90 days. And then, among this 1600 patients, 800 were the patients in the five practices that incorporate mortality prediction tool in their clinical workflow. The other 800 patients were affiliated with the seven practice that didn't participate. So we looked at the advanced care planning utilization by comparing the two groups of practices to see whether there's a difference. And we are really glad to see there is a significant difference in the utilization of advanced care planning between the two groups of practices. To be more specific, the advanced care planning utilization increased two and a half fold when comparing the practice that incorporated mortality tools in their practice, as compared to the practice that didn't participate.

So we're really glad to see these encouraging results. Currently, in order to identify the significant association between mortality tool deployment versus increased advanced care planning utilization, we are developing propensity scores, regression models, to reduce the effect of potential confining factors. For example, difference between practices, difference between patient demographics and clinical information. We're also starting to look at other important outcomes besides advanced care planning. For example, hospice enrollment, palliative care referral, chemotherapy in the last 30 days of life. Those are all very important outcomes we want to look at, and hopefully we can present our new results in the next year ASCO meeting.

I Really appreciate the opportunity to interview with Journal of Clinical Pathways. Thank you for the opportunity to deliver our exciting results to your readers and audience.