254 Background: Clinicians in oncology are often challenged to identify when patients with cancer are at high risk for adverse outcomes and would benefit from more intensive clinical care. Preemptive identification of these patients may improve efficiency and improve patient care. The objective of this quality improvement pilot was to prospectively validate a machine learning (ML)-based clinical tool designed to identify patients with cancer who are at high risk for an emergency department (ED) visit, and whether they met eligibility criteria for clinical services at home, such as Huntsman at Home (H@H). Methods: Patients with cancer who received care at Huntsman Cancer Institute (HCI) between Jan. 4 and Feb. 7, 2020 were included in the analysis. For patients with HCI contact in a given week, the ML-based tool predicted the probability of an ED visit in the next 60 days to identify “high risk” patients using real-time structured EHR data (e.g. demographic characteristics, vital signs, and laboratory values). Risk of an ED visit was used as a proxy for eligibility for H@H. Patients were randomized to two cohorts to assess eligibility precision and outcome forecast precision. Eligibility precision was defined as the percentage of ML-classified high risk patients who were confirmed by a nurse practitioner to be eligible for admission to H@H. Outcome forecast precision was defined as the percentage of ML-classified high risk patients who experienced a future ED event within 60 days and was compared to the baseline prevalence of ED visits within 60 days during the same time period. The IRB determined this to be a quality improvement project. Results: This quality improvement pilot included 1,236 patients; 53% were women, median age was 65 years, and 84% were Caucasian. The most common cancers, excluding non-melanoma, were breast, prostate, lung, and myeloma. The observed prevalence of an ED visit within 60 days was 7%. The tool classified 9% of patients as high risk. The eligibility precision was 0.76 (95% CI: 0.62-0.89), demonstrating concordance with clinician assessment among patients classified as high risk. Compared to a baseline of 0.07, the outcome forecast precision was 0.22 (95% CI: 0.12-0.34) for future ED events. Conclusions: This quality improvement pilot demonstrates the potential application of an ML-based tool to identify patients with cancer who may benefit from further support through the H@H program. The approach creates a framework for ML-based tools to enhance clinical services at home.
P154 Improving the haemoglobinopathy screening programme for high-risk patients in a tertiary maternity hospital in dublin, a pilot quality improvement initiative