Risk of Hospitalization and Emergency Room Visits Among Community Oncology Patients

Author(s)

Namasivayam G1, Rahman MM2, Mohammad N3, Chang B3, Karhade M3, Robert N3, Wu N4, Heller B5, Hoang S6, Alwardt S3, Neubauer M3, Staggs S4, Moore L7, Smith H3
1Ontada, Livermore, USA, 2Ontada, magnolia, USA, 3Ontada, The Woodlands, TX, USA, 4US Oncology Network, The Woodlands, TX, USA, 5Southern Cancer Center, Mobile, AL, USA, 6US Oncology, Austin, TX, USA, 7Ontada, Shaker Heights, OH, USA

Presentation Documents

OBJECTIVES : Machine Learning (ML) solutions can be used to bring insights to providers at the point of care. This research aimed to establish an explainable ML model to predict patient risk of emergency room (ER) visits and hospitalizations within 30 days after an oncology practice visit.

METHODS : A retrospective cohort of 98,686 patients within McKesson Specialty Health | The US Oncology Network’s iKnowMed electronic health records (EHR) system between 05/01/16 and 06/30/20 was identified for research purposes. Inclusion criteria included Oncology Care Model patients aged >65 years and on cancer treatment. Approximately 300 clinical metrics from the EHR system were considered including diagnosis, staging, labs, vitals, treatments, comorbidities, drugs, and performance. Data were split into training (90%) and testing (10%). Model was selected using performance metrics (AUC, sensitivity, specificity, and accuracy). Model interpretations were reviewed by a team of oncologists.

RESULTS : The extreme gradient boosting model (XGBoost) achieved the following testing results: AUC 72%, balanced accuracy 66%, sensitivity 65%, and specificity 67%. At threshold 0.5. The model correctly identified 65% of hospitalized and 67% of non-hospitalized patients. The top 5 attributes in order of SHapley Additive exPlanations (SHAP) impact values included albumin value (0.18), hemoglobin value (0.14), pain score (0.12), days since first chemotherapy (0.09) and lymphocytes (0.08). SHAP plots illustrate all associations between features and 30-day hospital admissions and ER risk.

CONCLUSIONS : Real-world data were harnessed and applied in an ML approach to establish a high-performing patient ER visit and hospitalization prediction model. The next phase will include model deployment to several US Oncology Network practices to validate effectiveness in the real world. Results will inform providers when a patient is at risk for an ER visit or hospitalization with the aim of improving the overall quality of oncology care and reducing admissions and ER visits.

Conference/Value in Health Info

2021-11, ISPOR Europe 2021, Copenhagen, Denmark

Value in Health, Volume 24, Issue 12, S2 (December 2021)

Acceptance Code

P73

Topic

Clinical Outcomes, Methodological & Statistical Research

Topic Subcategory

Artificial Intelligence, Machine Learning, Predictive Analytics, Clinical Outcomes Assessment, Clinician Reported Outcomes

Disease

Oncology

Your browser is out-of-date

ISPOR recommends that you update your browser for more security, speed and the best experience on ispor.org. Update my browser now

×