Background: Shortage of human resources, increasing educational costs, and the need to keep social distances in response to the COVID-19 worldwide outbreak have prompted the necessity of clinical training methods designed for distance learning. Virtual patient simulators (VPSs) may partially meet these needs. Natural language processing (NLP) and intelligent tutoring systems (ITSs) may further enhance the educational impact of these simulators. Objective: The goal of this study was to develop a VPS for clinical diagnostic reasoning that integrates interaction in natural language and an ITS. We also aimed to provide preliminary results of a short-term learning test administered on undergraduate students after use of the simulator. Methods: We trained a Siamese long short-term memory network for anamnesis and NLP algorithms combined with Systematized Nomenclature of Medicine (SNOMED) ontology for diagnostic hypothesis generation. The ITS was structured on the concepts of knowledge, assessment, and learner models. To assess short-term learning changes, 15 undergraduate medical students underwent two identical tests, composed of multiple-choice questions, before and after performing a simulation by the virtual simulator. The test was made up of 22 questions; 11 of these were core questions that were specifically designed to evaluate clinical knowledge related to the simulated case. Results: We developed a VPS called Hepius that allows students to gather clinical information from the patient’s medical history, physical exam, and investigations and allows them to formulate a differential diagnosis by using natural language. Hepius is also an ITS that provides real-time step-by-step feedback to the student and suggests specific topics the student has to review to fill in potential knowledge gaps. Results from the short-term learning test showed an increase in both mean test score (P<.001) and mean score for core questions (P<.001) when comparing presimulation and postsimulation performance. Conclusions: By combining ITS and NLP technologies, Hepius may provide medical undergraduate students with a learning tool for training them in diagnostic reasoning. This may be particularly useful in a setting where students have restricted access to clinical wards, as is happening during the COVID-19 pandemic in many countries worldwide.

Differentiation of syncope from transient loss of consciousness can be challenging in the emergency department (ED). Natural Language Processing (NLP) enables the analysis of free text in the electronic medical records (EMR). The present paper aimed to develop a large language models (LLM) for syncope recognition in the ED and proposed a framework for model integration within the clinical workflow. Two models, based on both the Italian and Multilingual Bidirectional Encoder Representations from Transformers (BERT) language model, were developed using consecutive EMRs. The "triage" model was only based on notes contained in the "triage" section of the EMR. The "anamnesis" model added data contained in the "medical history" section. Interpretation and calibration plots were generated. The Italian and Multi BERT models were developed and tested on both 15,098 and 15,222 EMRs, respectively. The triage model had an AUC of 0·95 for the Italian BERT and 0·94 for the Multi BERT. The anamnesis model had an AUC of 0·98 for the Italian BERT and 0·97 for Multi BERT. The LLM identified syncope when not explicitly mentioned in the EMR and also recognized common prodromal symptoms preceding syncope. Both models identified syncope patients in the ED with a high discriminative capability from nurses and doctors' notes, thus potentially acting as a tool helping physicians to differentiate syncope from others transient loss of consciousness.

Background: Enrollment of large cohorts of syncope patients from administrative data is crucial for proper risk stratification but is limited by the enormous amount of time required for manual revision of medical records. Aim: To develop a Natural Language Processing (NLP) algorithm to automatically identify syncope from Emergency Department (ED) electronic medical records (EMRs). Methods: De-identified EMRs of all consecutive patients evaluated at Humanitas Research Hospital ED from 1 December 2013 to 31 March 2014 and from 1 December 2015 to 31 March 2016 were manually annotated to identify syncope. Records were combined in a single dataset and classified. The performance of combined multiple NLP feature selectors and classifiers was tested. Primary Outcomes: NLP algorithms’ accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F3 score. Results: 15,098 and 15,222 records from 2013 and 2015 datasets were analyzed. Syncope was present in 571 records. Normalized Gini Index feature selector combined with Support Vector Machines classifier obtained the best F3 value (84.0%), with 92.2% sensitivity and 47.4% positive predictive value. A 96% analysis time reduction was computed, compared with EMRs manual review. Conclusions: This artificial intelligence algorithm enabled the automatic identification of a large population of syncope patients using EMRs.

Background: Virtual patient simulators (VPSs) log all users’ actions, thereby enabling the creation of a multidimensional representation of students’ medical knowledge. This representation can be used to create metrics providing teachers with valuable learning information. Objective: The aim of this study is to describe the metrics we developed to analyze the clinical diagnostic reasoning of medical students, provide examples of their application, and preliminarily validate these metrics on a class of undergraduate medical students. The metrics are computed from the data obtained through a novel VPS embedding natural language processing techniques. Methods: A total of 2 clinical case simulations (tests) were created to test our metrics. During each simulation, the students’ step-by-step actions were logged into the program database for offline analysis. The students’ performance was divided into seven dimensions: the identification of relevant information in the given clinical scenario, history taking, physical examination, medical test ordering, diagnostic hypothesis setting, binary analysis fulfillment, and final diagnosis setting. Sensitivity (percentage of relevant information found) and precision (percentage of correct actions performed) metrics were computed for each issue and combined into a harmonic mean (F1), thereby obtaining a single score evaluating the students’ performance. The 7 metrics were further grouped to reflect the students’ capability to collect and to analyze information to obtain an overall performance score. A methodological score was computed based on the discordance between the diagnostic pathway followed by students and the reference one previously defined by the teacher. In total, 25 students attending the fifth year of the School of Medicine at Humanitas University underwent test 1, which simulated a patient with dyspnea. Test 2 dealt with abdominal pain and was attended by 36 students on a different day. For validation, we assessed the Spearman rank correlation between the performance on these scores and the score obtained by each student in the hematology curricular examination. Results: The mean overall scores were consistent between test 1 (mean 0.59, SD 0.05) and test 2 (mean 0.54, SD 0.12). For each student, the overall performance was achieved through a different contribution in collecting and analyzing information. Methodological scores highlighted discordances between the reference diagnostic pattern previously set by the teacher and the one pursued by the student. No significant correlation was found between the VPS scores and hematology examination scores. Conclusions: Different components of the students’ diagnostic process may be disentangled and quantified by appropriate metrics applied to students’ actions recorded while addressing a virtual case. Such an approach may help teachers provide students with individualized feedback aimed at filling competence drawbacks and methodological inconsistencies. There was no correlation between the hematology curricular examination score and any of the proposed scores as these scores address different aspects of students’ medical knowledge.

In recent years, machine learning (ML) has been promisingly applied in many fields of clinical medicine, both for diagnosis and prognosis prediction. Aims of this narrative review were to summarize the basic concepts of ML applied to clinical medicine and explore its main applications in the emergency department (ED) setting, with a particular focus on syncope management. Through an extensive literature search in PubMed and Embase, we found increasing evidence suggesting that the use of ML algorithms can improve ED triage, diagnosis, and risk stratification of many diseases. However, the lacks of external validation and reliable diagnostic standards currently limit their implementation in clinical practice. Syncope represents a challenging problem for the emergency physician both because its diagnosis is not supported by specific tests and the available prognostic tools proved to be inefficient. ML algorithms have the potential to overcome these limitations and, in the future, they could support the clinician in managing syncope patients more efficiently. However, at present only few studies have addressed this issue, albeit with encouraging results.

Syncope is a challenging problem in the emergency department (ED) as the available risk prediction tools have suboptimal predictive performances. Predictive models based on machine learning (ML) are promising tools whose application in the context of syncope remains underexplored. The aim of the present study was to develop and compare the performance of ML-based models in predicting the risk of clinically significant outcomes in patients presenting to the ED for syncope. We enrolled 266 consecutive patients (age 73, IQR 58–83; 52% males) admitted for syncope at three tertiary centers. We collected demographic and clinical information as well as the occurrence of clinically significant outcomes at a 30-day telephone follow-up. We implemented an XGBoost model based on the best-performing candidate predictors. Subsequently, we integrated the XGboost predictors with knowledge-based rules. The obtained hybrid model outperformed the XGboost model (AUC = 0.81 vs. 0.73, p < 0.001) with acceptable calibration. In conclusion, we developed an ML-based model characterized by a commendable capability to predict adverse events within 30 days post-syncope evaluation in the ED. This model relies solely on clinical data routinely collected during a patient’s initial syncope evaluation, thus obviating the need for laboratory tests or syncope experienced clinical judgment.