Explainable AI Predicts ED Hospital Admissions with High Accuracy, Enhancing Triage and Resource Allocation
What the Researchers Tested
This retrospective cohort study utilized a diverse set of routinely collected structured data from 526,145 adult ED visits across two academic hospitals in the McGill University Health Centre. The dataset spanned from January 2015 to December 2022, focusing on binary outcomes of admission or discharge. Data were meticulously split into training (70%), validation (10%), and testing (20%) sets using stratified random sampling to maintain outcome proportions. The input features for the machine learning models included comprehensive demographics, vital signs, chief complaints, documented comorbidities, and historical hospitalization frequency, all readily available at triage. The study rigorously evaluated several machine learning classifiers, including logistic regression (LR), decision trees, random forests, artificial neural networks, and gradient boosting algorithms like extreme gradient boosting (XGBoost), light gradient boosting machine, and categorical boosting. The XGBoost model consistently demonstrated superior performance. A critical aspect of this study was its focus on explainability, employing Shapley Additive Explanations (SHAP) values to elucidate the contribution of each feature to the model's predictions. Sensitivity analyses were conducted to test the robustness of the model against different exclusion criteria and temporal data shifts, while subgroup analyses confirmed consistent performance across varying age groups, sexes, and hospital sites.
What the Study Found
The final cohort comprised 526,145 adult ED visits, with 91,363 (17.4%) resulting in hospital admission.
The XGBoost model achieved the highest predictive performance with an AUROC of 0.86 (95% CI 0.858-0.864), demonstrating superior discrimination compared to the logistic regression baseline (AUROC 0.84).
At an optimized decision threshold, the XGBoost model yielded a sensitivity of 0.67 and a specificity of 0.86 for predicting hospital admission.
Key predictors identified through SHAP values included patient age, Canadian Triage and Acuity Scale (CTAS) scores, mode of arrival to the ED, and frequency of prior hospitalizations.
Subgroup analyses confirmed the model's strong and consistent performance across diverse demographics, including age, sex, and different hospital sites, indicating broad applicability.
Why This Matters to Clinical Practice
The persistent challenge of ED overcrowding significantly compromises the quality and efficiency of urgent care, leading to extended wait times, delayed diagnoses, and increased patient morbidity. Early and accurate prediction of hospital admission at triage offers a crucial opportunity to optimize resource allocation within strained healthcare systems. By identifying patients highly likely to require inpatient beds earlier, hospitals can initiate pre-emptive bed management, streamline transfer processes, and enhance coordination between ED and inpatient services. For practicing clinicians, this translates to improved patient flow, potentially reducing the burden on ED staff and enabling more focused care for critical cases. Department heads and hospital leadership can leverage such predictive tools for strategic operational planning, allowing for more dynamic staffing adjustments and better utilization of finite inpatient resources. Ultimately, reducing diagnostic and admission delays can lead to better patient outcomes, decreased length of stay, and enhanced overall patient satisfaction.
Clinical & Workflow Takeaways
Integrating an explainable AI model like XGBoost into ED triage workflows could provide clinicians with real-time, data-driven insights to anticipate hospital admissions. Physicians and nurses at triage could use this tool to prioritize initial workups for high-risk patients, alert inpatient teams sooner, and facilitate more efficient bed requests. Department leaders should consider pilot programs for prospective validation and evaluate the impact on key operational metrics such as ED length of stay for admitted patients, time to inpatient bed, and overall ED throughput. Training for clinical staff on interpreting AI predictions and SHAP explanations would be essential to foster trust and ensure effective utilization, moving beyond simple black-box predictions to informed decision-making.
Limitations and What Remains Uncertain
Despite its strengths, this study is a retrospective cohort analysis, which inherently limits its ability to establish causality and is susceptible to biases from historical data collection. The data originated from two academic urban tertiary care centers within a single health system, potentially limiting the generalizability of findings to different healthcare settings, patient populations, or rural EDs. While explainability was addressed using SHAP values, real-world clinical interpretability and acceptance by frontline staff would require further prospective validation and qualitative assessment. The study's exclusion of visits with dispositions other than admission or discharge simplifies the outcome but may not fully reflect the complexity of ED patient pathways, such as transfers to other facilities or deceased patients. Future work is imperative for prospective validation in diverse clinical environments and robust integration into real-time clinical workflows to truly assess its operational impact and clinical utility.
Explore the Work Behind Clinical AI
MedXchange Research Fellowship
Publish Clinical AI Research With Us
Our 10-week fellowship mentors physicians and researchers to build foundation models, analyze multimodal health datasets, and co-author peer-reviewed clinical studies.
Related Clinical AI Stories
Multimodal Foundation Models in Clinical Triage: Validating EHR and Imaging Fusion
A landmark multi-center prospective cohort investigation published in Nature Medicine evaluated the diagnostic accuracy of multimodal foundation models combining longitudinal electronic health records (EHR) with acute computed tomography (CT) scans in emergency department triage. Across 42,850 consecutive emergency admissions across four tertiary academic centers, the multimodal architecture achieved an AUROC of 0.93 for identifying life-threatening acute decompensations within 12 hours of presentation, significantly outperforming traditional single-modality clinical risk scores (NEWS2 AUROC 0.74, single-slice CT deep learning AUROC 0.81).
AI's Role in Diagnosing Rare and Undiagnosed Diseases: A Critical Perspective
A recent perspective from the NEJM Catalyst Insights Council highlights the significant potential of artificial intelligence (AI) in accelerating the diagnosis of rare and undiagnosed diseases. These conditions, often individually uncommon, collectively affect a substantial patient population, leading to prolonged diagnostic odysseys, increased healthcare costs, and significant patient morbidity. The Council emphasizes that while AI offers unprecedented capabilities in pattern recognition across vast datasets, it represents only one component of a multifaceted solution required to address this complex clinical challenge. The editorial underscores that the inherent complexity of rare diseases—characterized by phenotypic heterogeneity, genetic variability, and limited epidemiological data—makes them particularly amenable to AI's strengths in identifying subtle correlations and anomalies. However, achieving accurate and timely diagnoses will necessitate more than just technological advancements; it demands integrated strategies encompassing robust data infrastructure, multidisciplinary clinical collaboration, and careful consideration of ethical implications.
LLM Performance in PMOS/PCOS Information: DeepSeek-V4 Outperforms GPT-5.5 and Gemini 3.5 in Clinician-Rated Quality and Empathy
A recent study evaluated the information quality, readability, transparency, and perceived empathy of responses from three widely accessible large language models (LLMs)—DeepSeek-V4, GPT-5.5, and Gemini 3.5—to clinician-formulated questions regarding polyendocrine metabolic ovarian syndrome (PMOS), formerly known as polycystic ovary syndrome (PCOS). The study found significant differences in performance among the LLMs, with DeepSeek-V4 consistently receiving higher clinician ratings across multiple quality and empathy metrics. Thirty-three distinct English questions, crafted by clinicians and using established PCOS terminology, were submitted to each LLM. Two independent clinicians then rigorously assessed the 99 archived responses using standardized tools such as the DISCERN instrument for information quality, EQIP for transparency, and JAMA benchmarks. Readability was quantitatively analyzed using six different indices, while perceived empathy was subjectively rated. The methodological approach involved matched comparisons using Friedman tests, paired signed-rank permutation comparisons, and Holm correction to account for multiple comparisons.