Background: Syncope, a common problem encountered in the emergency department (ED), has a multitude of causes ranging from benign to life-threatening. Hospitalization may be required, but the management can vary substantially depending on specific clinical characteristics. Models predicting admission and hospitalization length of stay (LoS) are lacking. The purpose of this study was to design an effective, exploratory model using machine learning (ML) technology to predict LoS for patients presenting with syncope. Methods: This was a retrospective analysis using over 4 million patients from the National Emergency Department Sample (NEDS) database presenting to the ED with syncope between 2016–2019. A multilayer perceptron neural network with one hidden layer was trained and validated on this data set. Results: Receiver Operator Characteristics (ROC) were determined for each of the five ANN models with varying cutoffs for LoS. A fair area under the curve (AUC of 0.78) to good (AUC of 0.88) prediction performance was achieved based on sequential analysis at different cutoff points, starting from the same day discharge and ending at the longest analyzed cutoff LoS ≤7 days versus >7 days, accordingly. The ML algorithm showed significant sensitivity and specificity in predicting short (≤48 h) versus long (>48 h) LoS, with an AUC of 0.81. Conclusions: Using variables available to triaging ED clinicians, ML shows promise in predicting hospital LoS with fair to good performance for patients presenting with syncope.
26 June 2025 in European Heart Journal - Digital Health
Aims: Syncope remains a diagnostic challenge despite advancements in testing and treatment. Cardiac syncope is an independent predictor of mortality and can be difficult to distinguish from other causes of transient loss of consciousness (TLOC). This paper explores whether artificial intelligence (AI) can improve the evaluation and management of patients with syncope. Methods and results: We conducted a literature review and incorporated the opinions of experts in the fields of syncope and AI. The cause of TLOC is often unclear, hospitalization criteria are ambiguous, diagnostic tests are frequently non-informative, and assessments are costly. Patients are left with unanswered questions and limited guidance. Artificial intelligence (AI) has the potential to optimize syncope evaluation by processing large data sets, detecting imperceptible patterns, and assisting clinicians. However, AI has limitations, including errors, lack of human empathy, and uncertain clinical utility. Liability issues further complicate its integration. We present three viewpoints: (i) AI is crucial for advancing syncope management; (ii) AI can enhance the patient experience; and (iii) AI in syncope care is inevitable. Conclusion: Artificial intelligence may improve syncope diagnosis and management, particularly through machine learning–based test interpretation and wearable device data. However, it has yet to surpass human clinical judgment in complex decision-making. Current challenges include gaps in understanding syncope mechanisms, AI interpretability, generalizability, and clinical integration. Standardized diagnostic approaches, real-world validation, and curated data sets are essential for progress. Artificial intelligence may enhance efficiency and communication but raises concerns regarding confidentiality, bias, inequities, and legal implications.
Objective To incorporate continuous vital sign and ventilator data to improve neonatal extubation failure prediction in high-risk neonates. Study Design: In a single-center retrospective cohort, neonates < 28 weeks' gestation with an extubation event and physiologic data were included. Extubation failure was defined as re-intubation within 7 days. The cohort was divided 75/25 into train-test sets with K-fold cross-validation, evaluating multiple machine learning algorithms. Model interpretability was assessed using SHapley Additive exPlanations (SHAP) values. Results Of 157 infants, 31 experienced extubation failure (mean time to failure, 58 hours). Using physiologic data and clinician-engineered features, logistic regression performed best (AUC 0.88). Top predictors included post-menstrual age at extubation, median FiO2, median SpO2, maximum heart rate, and minimum SpO2. Conclusion Incorporating physiologic data significantly improved extubation prediction over historical models. We propose optimizing negative predictive value to objectively guide earlier ventilator liberation, with future multicenter studies required to enhance performance and transferability.