EHR Predictor Rules Out Systemic Steroids in Hospitalized Asthma

• Random-forest model predicted escalation from inhaled corticosteroids (ICS) to systemic corticosteroids (SCS) with moderate discrimination (AUC 0.75 internal; 0.69 temporal) and high rule-out performance (negative predictive value 87–89%) in 5,463 hospitalized patients (19.9% escalated).

• Operational use case for pharmacists: triage medication reviews and steward SCS—prioritize likely escalations for review and de-prioritize low-risk orders. A matched analysis (62 pairs) found no clinical outcome benefits from escalation but did find higher hospitalization costs associated with SCS escalation.

• In hospitalized asthma, deciding whether to escalate from ICS to systemic corticosteroids is high-stakes: individual responses to ICS vary, pulmonary function tests (PFTs) are often unavailable or inadvisable during acute exacerbations, and SCS carry well-documented harms (e.g., osteoporosis, gastrointestinal bleeding, neuropsychiatric effects).

• Without validated decision rules, escalation practices are inconsistent, which places pharmacy teams in the position of managing preventable systemic steroid exposure and added inpatient cost.

• A validated prediction tool enables steroid stewardship and pragmatic clinical decision support (CDS) prioritization so limited pharmacist effort targets patients most likely to need escalation while avoiding unnecessary systemic exposure and spend.

• Retrospective cohort of 5,463 hospitalized asthma patients (age ≥12) at Xinjiang Uygur Autonomous Region People’s Hospital, admitted 2006–2022; all received inhaled corticosteroids at admission.

• Developed and compared models (multivariable logistic regression, LASSO, XGBoost, random forest) using 41 routinely available admission variables; predictors with ≥25% missingness were excluded and analyses used complete-case records for the remaining predictors.

• Derivation cohort (2006–2018) was split into training and a 30% internal hold-out set; temporal validation used 2019–2022 admissions. Performance was evaluated by receiver operating characteristic (ROC) area under the curve and predictive values.

• Assessed potential SCS overuse with a non-experimental causal-inference style matched analysis: false-negative predicted patients who received SCS were matched to true-negative predicted patients to compare outcomes and costs in the temporal validation set.

• Random forest predicted ICS→SCS escalation with AUC 0.7483 (internal) and 0.6941 (temporal) across 5,463 hospitalized patients (19.9% escalated).

• At a pre-specified cutoff of 0.2, sensitivity/specificity were 68.5%/66.2% (internal) and 63.4%/61.5% (temporal); positive predictive value 34.5%/27.8%; negative predictive value 89.0%/87.8%.

• Top predictors were routine admission laboratory values and demographics: eosinophil count, neutrophil count, age, and lymphocyte count; the high NPV supports using the model as a rule-out tool so pharmacists can deprioritize low-risk SCS orders.

• In a matched analysis (62 pairs) of temporal-validation patients predicted as low-risk, those who nonetheless received SCS (false negatives) showed no improvement in mortality, ICU transfer, or length of stay but incurred higher total hospitalization costs—suggesting possible SCS overuse.

• Open question for implementation: which specific predictors or clinical workflows drive the model’s practical improvement in stewardship and outcomes requires further prospective evaluation.

• Implement an EHR (Epic) work queue that scores patients at ICS initiation using admission labs and demographics; show top drivers (eosinophils, neutrophils, lymphocytes, age) and provide brief clinician/pharmacist training with case reviews.

• Use the model as a rule-out screening tool starting at the study’s 0.2 cutoff: fast-track high-risk patients for pharmacist–pulmonology review and add a soft prompt to reconsider SCS orders for low-risk predictions.

• Governance and monitoring: require pharmacist override rationale in the EHR/Pyxis when initiating SCS on low-risk patients; monitor SCS initiation rate, total cost, ICU transfer, and length of stay in this group.

• Practical framing: treat the model as an 'expedite low-risk' lane—it speeds low-risk patients past default SCS escalation so clinicians can focus scrutiny on the remaining patients, while preserving pharmacist clinical judgment and override authority.

Strengths:

• Robust design with a large inpatient cohort spanning 2006–2022; events-per-variable >20 and adherence to TRIPOD reporting with both hold-out and temporal validation.

• Used widely available admission labs and demographics to support clinical interpretability; added a matched outcome analysis to probe potential SCS overuse and operational impact.

Limitations:

• Single-center, retrospective, complete-case approach with a ≥25% missingness exclusion reduced the analytical cohort and limits external generalizability; multicenter prospective validation is required.

• Key covariates were absent (pulmonary function indices, outpatient adherence, asthma phenotypes such as T2-high vs T2-low), and outcomes were limited to in-hospital metrics; prospective, real-time implementation and external-site calibration remain untested.

The random-forest model is ready for pilot implementation as a high–negative predictive value rule-out tool to triage systemic corticosteroid decisions, but it requires prospective, multicenter validation and workflow evaluation before widespread adoption.