ML Helps Stratify Adverse Outcome Risk in Older Cancer Patients

A study published last month in JAMA Network Open described a new machine learning (ML) tool that may assist clinicians in identifying older adults with advanced cancer who are at higher risk of adverse outcomes.

According to the researchers, older adults with advanced cancer often present with a variety of symptoms prior to cancer treatment, such as pain, fatigue, and insomnia. Patients with high pretreatment symptom severity frequently experience adverse events during cancer treatment, highlighting the need for a method to stratify these patients based on their risk for adverse outcomes.

To do this, the researchers developed an unsupervised ML model and evaluated it using a secondary analysis of the Geriatric Assessment Intervention for Reducing Toxicity in Older Patients With Advanced Cancer (GAP70+) Trial.

The researchers pulled a cohort of patients from the trial who completed the National Cancer Institute Patient-Reported Outcomes version of the Common Terminology Criteria for Adverse Events (PRO-CTCAE) prior to starting a new cancer treatment regimen. All participants also received care at community oncology sites in the US.

The final cohort included 706 patients with various cancer types. Using data from these patients, the ML model clustered patients based on similarities of baseline symptom severity. This yielded variables related to severity categories for 24 PRO-CTCAE symptoms.

The total severity score generated by this analysis was calculated as the sum of the 24 variables for each participant.

Then, the research team investigated potential associations between these variable clusters and unplanned hospitalization, death, and toxic effects. These adverse outcomes were quantified as unplanned hospitalization over 3 months, all-cause mortality over 1 year, and any clinician-rated grade 3 to 5 toxic effect over 3 months.

The model identified three patient clusters characterized by symptom severity, which the researchers labeled low, moderate, and high. These clusters were significantly associated with increased risk of certain adverse outcomes, namely unplanned hospitalization and death.

Using this framework, the ML tool classified 43.9 percent, 41.8 percent, and 14.3 percent of the cohort into low-, medium-, and high-severity clusters, respectively.

The research team also found that risk increased even when controls for sociodemographic variables, clinical factors, study group, and practice site were applied to the analysis. After controlling for these factors, patients in the moderate-severity cluster were more likely to experience hospitalization than those in the low-severity cluster.

The study also revealed that participants in the moderate- and high-severity clusters were associated with a higher risk of death, but not toxic effects.

The researchers concluded that these findings indicate ML’s potential to guide risk stratification tool development and serve as an assistive tool to help clinicians identify older adults with high risk of hospitalization and death prior to starting a new cancer treatment regimen.

As health systems work to personalize cancer care, predictive analytics have shown promise to bolster that effort.

In September, researchers shared that a prediction model could accurately forecast health-related quality of life (HRQOL) among adult survivors of childhood cancer using sociodemographic, lifestyle, and health state factors.

In their research, they argued that such predictions are key to developing interventions to improve outcomes for these patients, noting that previous studies had looked at HRQOL among childhood cancer survivors, but none had investigated predicting declines in HRQOL or poor HRQOL for these individuals.

The researchers found a strong association of chronic health conditions, emotional and neurocognitive impairment, and current smoking with declines in physical and mental HRQOL. To improve outcomes for adult survivors of childhood cancer, they stated that interventions must target those specific risk factors.