Unplanned hospitalizations may be common among patients with cancer, but they diminish quality of care while racking up high costs for patients and the health-care system alike. According to data presented at the 2019 ASCO Quality Care Symposium, however, a new predictive model may help providers to identify high-risk patients and optimize interventions, improving quality and value of care.1
In patients with gastrointestinal cancer undergoing radiation therapy, the machine-learning model successfully identified those at low vs high risk of 30-day unplanned hospitalization. The study, which analyzed more than 1,300 courses of radiation therapy, found that high-risk patients faced upward of a 30% risk of unplanned hospitalizations and trended toward worse outcomes compared with their low-risk counterparts.
“If we can identify high-risk patients upfront before they start radiotherapy, we might be able to implement strategies to minimize the risks of unplanned hospitalizations.”— Kaitlin M. Christopherson, MD
Tweet this quote
“This predictive model for hospitalizations in patients undergoing radiation therapy for gastrointestinal tumors may be a useful tool for improving quality and value of care,” said Kaitlin M. Christopherson, MD, a resident physician at The University of Texas MD Anderson Cancer Center, Houston. “If we can identify high-risk patients upfront before they start radiotherapy, we might be able to implement strategies to minimize the risks of unplanned hospitalizations.”
As Dr. Christopherson explained, up to 20% of patients with gastrointestinal malignancies may experience unplanned hospitalizations after radiation therapy—a substantial cost for both patients and the health-care system to bear. For this study, Dr. Christopherson and colleagues sought to apply a machine-learning approach to project 30-day unplanned hospitalizations for patients receiving radiation to gastrointestinal subsites. The researchers separated the study population into abdominal and pelvic cohorts to account for different patient characteristics, treatment paradigms, and toxicities. They then evaluated the impact of unplanned hospitalizations on medical cost. Finally, the researchers evaluated the rates of hospitalization and overall survival based on low-risk and high-risk subgroups.
Kaitlin M. Christopherson, MD
The study authors noted that training and validation cohorts were used for machine-learning development. The former was used to test multiple models for unplanned hospitalizations, including random forest, gradient-boosted decision trees, and logistic models. A prespecified threshold of AUC greater than 0.7 was deemed clinically valid. The researchers then incorporated over 700 clinical and treatment variables into the models from multiple electronic medical records (eg, MOSAIQ and EPIC) and institutional databases.
Rate of Unplanned Hospitalizations Among High-Risk Subgroup
As Dr. Christopherson reported, researchers analyzed 836 abdominal (gastric, pancreatic, biliary, hepatic) and 514 pelvic (rectal, anal) courses of radiotherapy for gastrointestinal cancers treated at MD Anderson from 2016 to 2019. The majority of patients had locally advanced, recurrent, or stage IV disease, and most patients were treated with curative or consolidative intent in both cohorts.
Conformal radiation therapies such as intensity-modulated radiation and stereotactic body radiation were used more commonly in the abdominal group (75% of patients), whereas just 48% of patients in the pelvic group received these advanced treatment techniques. Approximately 60% of patients in the abdominal group received concurrent chemotherapy, whereas 81% of patients in the pelvic group had concurrent chemotherapy.
Among 1,350 cases, data showed that the incidence of unplanned hospitalization within 30 days of radiation therapy was 12.3% (13.3% in the abdominal cohort and 10.7% in the pelvic cohort). The average length of stay was the same between the two cohorts (5 days); however, 25% of patients had an unplanned hospital stay of 10 days or longer, Dr. Christopherson reported.
A secondary analysis of global costs in hospitalized vs nonhospitalized patients showed an increase of nearly $50,000 in median health-care costs among those hospitalized within 30 days of starting radiation therapy ($69,108 in nonhospitalized patients and $119,844 in hospitalized patients). Researchers also defined a low- vs high-risk patient subset within each cohort, with 80% defined as high-risk patients. The authors noted that the actual rate of hospitalizations for abdominal patients deemed to be at high risk was nearly 40% vs 9% in the low-risk group. Finally, Dr. Christopherson and colleagues analyzed overall survival as stratified by risk grouping and found a clear separation of curves between low- and high-risk patients.
“Given this separation, predictive analytics may be a key tool to help providers identify high-risk patients and optimize interventions before these hospitalizations occur,” Dr. Christopherson concluded.
Incorporating Risk Into Patient Management
“We need to figure out how to take this risk information and pair it with an intervention or a set of interventions to actually help our patients by improving outcomes, quality, and care.”— Gabriel A. Brooks, MD, MPH
Tweet this quote
Although risk stratification and prediction may pose challenges, according to the discussant of the abstract, Gabriel A. Brooks, MD, MPH, the more difficult problem is figuring out how to incorporate risk information into patient management to prevent unplanned hospitalizations. “Risk stratification identifies patients who may benefit from more intensive management, but I say ‘may’ because we don’t know that until we show that,” said Dr. Brooks, Assistant Professor of Medicine at the Geisel School of Medicine at Dartmouth, Hanover, New Hampshire. “We don’t yet know if we can modify the risk through some type of intervention.”
Regarding Dr. Christopherson’s data, Dr. Brooks noted that a 70% prediction rate of higher vs lower risk (AUC = 0.7) suggests that the machine-learning model works. What’s more, Dr. Brooks added, knowing that high-risk patients have a fourfold greater risk for hospitalization compared with low-risk patients “feels like actionable information.”
Gabriel A. Brooks, MD, MPH
According to Dr. Brooks, however, given differences among electronic medical records, performance of the prediction model—“validated in one cohort, one situation, one healthy system”— might suffer if moved from one system to another. “Although a complex model offers the possible advantage of better predictions, there is a trade-off in that simpler models are easier to implement and interpret,” said Dr. Brooks. He noted that a model incorporating just two variables—sodium and albumin—was able to yield good separation of high- and low-risk groups using a predefined cut point.2
In patients receiving first-line chemotherapy for advanced solid tumors, Dr. Brooks and colleagues identified a 24% risk of 30-day hospitalizations among high-risk patients, compared with 8% for low-risk patients.
“The key question, however, is which interventions are effective for preventing hospitalizations,” Dr. Brooks emphasized. “This is where research is most needed moving forward. We need to figure out how to take this risk information and pair it with an intervention or a set of interventions to actually help our patients by improving outcomes, quality, and care.”
DISCLOSURE: Drs. Christopherson and Brooks reported no conflicts of interest.
1. Christopherson KM, Berlind CG, Ahern CA, et al: Improving quality through AI: Applying machine learning to predict unplanned hospitalizations after radiation. 2019 ASCO Quality Care Symposium. Abstract 271. Presented September 6, 2019.
2. Brooks GA, Uno H, Aiello Bowles EJ, et al: Hospitalization risk during chemotherapy for advanced cancer: Development and validation of risk stratification models using real-world data. JCO Clin Cancer Inform 3:1-10, 2019.