Patel, Shrinath

ABSTRACT Background Optimizing access to cost‐effective care in capacity‐constrained health systems is a contemporary imperative. We evaluated whether machine learning (ML) models using patient‐reported data could optimize access for patients requiring surgical care in rhinology clinics, without need for CT imaging. The outcome was defined as any rhinologic surgery within 90 days of initial evaluation. Methods A de‐identified electronic dataset from patients seen at five distinct sites within an integrated healthcare system between 2018 and 2025 was used to train models. Demographic data and 22‐item Sinonasal Outcome Test (SNOT‐22) responses were studied. Models were trained using stratified 5‐fold cross‐validation with an 80% development cohort and validated in a 20% held‐out validation cohort. Hyperparameters were optimized with Optuna. The primary outcome was performance of any rhinologic surgical intervention within 90 days of initial SNOT‐22. Results Data from 35,170 patients were evaluated. Item‐level responses outperformed use of total SNOT‐22 score in the models. Among models, XGBoost demonstrated best discrimination with AUC of 0.70 (95% CI, 0.69–0.71), outperforming logistic regression (0.66), random forest (0.63), and TabNet (0.66) (all p < 0.001). At optimal threshold, XGBoost achieved 66% sensitivity, 64% specificity, 28% PPV, and 90% NPV. Top predictors for surgery were age, nasal blockage, facial pain or pressure, and decreased sense of smell or taste. Validation on the held‐out cohort remained stable (AUC 0.70), with strong discrimination across sites despite surgical rates ranging from 11.5% to 27.4%. Conclusions Demographics and item‐level SNOT‐22 responses were successful in developing an ML model that demonstrated moderate discrimination and high NPV (90%) for performance of surgery within next 90 days. ML models balanced with human oversight may accelerate triage and optimize surgical yield for rhinology clinics.

Objectives: Electronic health records (EHR)-based triggers (eTriggers) have been used to study diagnostic errors in the emergency department (ED), often with suboptimal performance. Our objective was to investigate incremental value of multi-factor machine learning (ML) approaches to improve eTrigger performance. Methods: Patients presenting to an academic ED were categorized into trigger-positive and trigger-negative using standard trigger (T) definitions: (T1) ED return visits resulting in admission within 10 days; (T2) care escalation from the inpatient unit to the ICU within 24 hours; and (T3) deaths within 24 hours of admission. We trained and evaluated 6 supervised ML models. Results: A total of 124,053 consecutive encounters (5791 T-positive and 118,262 T-negative) were included. Among the T-positive, 4159 (72%) were associated with T1, 1415 (24%) with T2, and 217 (4%) with T3. The T-based positive predictive values (PPV) were 5.2% for T1, 8.2% for T2, and 6.5% for T3. ML models trained and evaluated on balanced training dataset and imbalanced test set had low classification performances (accuracy: 0.72–0.95; PPV: 0.00–0.16; F1-score: 0.00–0.23). Higher performances were observed in balanced test sets (accuracy: 0.80–0.97; PPV: 0.82–1.00; F1-score: 0.79–0.97). Comparing models trained on clinically annotated data with models trained on T-based labels identified other important factors. Conclusions: Utilizing machine learning to refine e-triggers slightly improves the identification of diagnostic errors, as evidenced by an increase in PPV values. We identified new potential factors contributing to ED diagnostic errors. These findings open new avenues to construct or modify more accurate e-triggers for diagnostic error identification.

Importance Diagnostic errors represent a major patient safety concern, with the potential to significantly impact patient outcomes. To address this, various trigger-based strategies have been developed to identify diagnostic errors, aiming to enhance clinical decision-making and improve patient safety. Objective To evaluate the performance of three pre-established triggers (T) in the emergency department (ED) setting and assess their effectiveness in detecting diagnostic errors. Design Consecutive cohort, retrospective observational design. Setting Academic ED with 80 000 annual visits. Participants Adults and children presenting to a single ED in the USA between 1 May 2018 and 1 January 2020. Intervention/outcomes Electronic health records (EHRs) were retrieved and categorised into trigger-positive and trigger-negative cases using the following criteria: T1—unscheduled returnvisits to the ED with admission within 7–10 days of theinitial visit; T2—care escalation from the inpatient unitto the intensive care unit (ICU) within 6, 12 or 24 hoursof ED admission; and T3—all deaths in the ED or within24 hours of ED admission, excluding palliative care. A random sample of trigger-positive cases was reviewed using the SaferDx tool to determine the presence or absence of a diagnostic error. Results A total of 5791 trigger-positive and 118262 trigger-negative cases were identified. Among trigger-positive cases, 4159 (72%) were associated with T1, 1415 (24%) with T2, and 217 (4%) with T3. A preliminary chart review of 462 trigger-positive and 251 trigger-negative cases showed most were error-negative (279 and 217, respectively). Detailed reviews found 32 diagnostic errors among 183 trigger-positive cases, yielding PPVs of 5.4% (T1), 8.9% (T2), and 6.9% (T3). No errors were found in 34 reviewed trigger-negative cases, resulting in a 100% NPV. Sepsis was the most common diagnosis among error-positive cases (n=11, 34.4%). Those with non-specific chief complaints like altered mental status or shortness of breath had higher diagnostic error risk. Conclusion and relevance While previously proposed EHR-based triggers can identify some diagnostic errors, they are insufficient for detecting all cases. To improve error detection performance, we recommend exploring data-driven strategies, such as machine learning techniques, to more effectively identify underlying contributing factors to diagnostic errors and enhance detection accuracy in the ED.