Risk of acute care events: Designing prognostic models for real world use.

Abstract

e13562 Background: Electronic health records (EHRs) combined with advances in cloud processing and computing power have made prognostic modeling a growth area in biomedical research. Despite calls to deploy these models in practice, they are rarely used in part because of challenges integrating models within EHRs, clinical workflows or both. We address this translational bottleneck in the context of risk for emergency department visits and/or unplanned hospitalizations (i.e., acute care events, ACEs) among people receiving systemic cancer treatment by partnering with a Clinical Advisory Panel (CAP) and EHR specialists to consider clinical integration during model development. Here we present the resulting prognostic model, illustrating the influence of advisors on model development. Methods: We identified adults aged 21+ with cancer initiating systemic therapy in 2022 at an academic medical center or six affiliated community sites. We compiled clinical and socio-demographic characteristics from structured EHR data, including age, sex, cancer diagnosis, anti-cancer drug orders and infusions, supportive medication orders, and prior ACEs. We used geocoding to characterize patients' residential community. Data were divided into 50% training, 25% validation and 25% test sets. Prognostic models were fit iteratively alongside discussions with the CAP and other advisors to optimize model design, variable construction, and guidance for use with risk-stratified supportive interventions. After initial discussion, elastic net regression was used to guide initial variable selection. Final models were estimated with ridge regression. Results: 4697 study eligible patients were identified. Based on feedback, we identified a need for two prognostic models. First, a baseline model that used information from oncology treatment plans to predict ACE risk prior to therapy initiation, enabling proactive intervention. Second, a time-updated model that used infusion room dispensing information to continually re-calculate ACE risk in 30-day increments based on updated treatment and clinical events over patients’ course of therapy. In both models, ACE in the past month emerged as the strongest predictor, displaying the largest standardized coefficient magnitude following penalization (OR = 1.91 and 1.60, respectively). The initial model had a C-statistic of 0.70 and the time-updated model had a C-statistic of 0.72. Conclusions: We present prognostic models designed for clinical use with risk-stratified supportive interventions. By limiting predictors to structured EHR data fields, considering clinical workflows, and planning for integration during model development we have illustrated a way to address a translational bottleneck with prognostic modeling. Engaging diverse subject matter experts and end users prior to and during model development has the potential to accelerate the adoption and use of prognostic models in practice.