Assessing the true nature of unplanned cancer care.

183 Background: Unplanned care can result in poor outcomes that are potentially preventable. The design of effective interventions to improve outcomes for cancer patients requires a better understanding of the true nature of unplanned care. Although cancer care teams document each patient’s care trajectory in detailed free-text notes, care outcomes are typically measured from structured patient record data and do not contain key information necessary for quality improvement efforts, such as the etiology of emergent events, or events that occur at outside facilities. To inform clinical effectiveness work at Stanford’s Cancer Institute, we describe our application of text-mining to improve the assessment of post-diagnosis morbidity outcomes. Methods: We conducted a retrospective study of unplanned care among 3,318 patients with a new diagnosis of breast, gastrointestinal, or thoracic cancer during 2010-13. Using a validated framework for clinical text-mining, we analyzed 308,000 notes for two tasks. First, we extract information on external unplanned events that are documented by providers. Second, we profile symptom mentions in Emergency Department (ED) notes. Results: For all cancer patients, text-mining detected over 400 unplanned events (93% PPV) at outside facilities, resulting in patient rates of 5% in the first 30 days, and 11% up to one year post-diagnosis. Among breast cancer patients, the top three symptoms reported in ED notes are pain (89%), nausea (37%) and fever (18%). Pain is consistently the most prevalent symptom up to one year after diagnosis, other symptoms exhibit more dynamic trends; wound related disorders and nausea are more prevalent among ED admissions in the first three months, whereas fever, cognitive impairment and mental health issues become more prevalent among admissions after the first three months of cancer care. Conclusions: The application of text-mining methods can improve the quantification of morbidity outcomes by improving the estimation of unplanned care rates and by providing continued learning for symptom-driven interventions to mitigate preventable emergent care. Although additional information gaps in care trajectories may continue to exist, text-mining can aid in assessing the true nature of unplanned care.