Cardiovascular diseases remain the leading cause of global morbidity and mortality, and timely recognition of cardiac arrhythmias in the emergency department (ED) is critical for preventing adverse outcomes such as sudden cardiac death, stroke, and hemodynamic collapse. The 12-lead electrocardiogram (ECG) is the cornerstone of rapid cardiac assessment in emergency settings; however, interpretation is highly dependent on physician expertise, workload, and time constraints. In busy ED environments, delays or inaccuracies in arrhythmia detection can significantly affect patient triage, treatment decisions, and prognosis. Studies have shown that even experienced clinicians may demonstrate variability in ECG interpretation, particularly in subtle or complex arrhythmias ^[1,2].

Artificial intelligence (AI), particularly machine learning and deep learning algorithms, has emerged as a transformative tool in cardiovascular diagnostics. Convolutional neural networks trained on large ECG datasets have demonstrated performance comparable to, and in some cases exceeding, that of cardiologists in identifying arrhythmias such as atrial fibrillation, ventricular tachycardia, and heart block ^[3]. AI-enabled ECG systems can analyze waveform morphology, rhythm regularity, and subtle interval variations in milliseconds, thereby offering rapid decision support in emergency settings. Such systems have the potential to reduce diagnostic delays, standardize triage decisions, and improve patient outcomes.

Recent research has validated AI algorithms for arrhythmia classification using large annotated ECG databases, demonstrating high sensitivity and specificity across multiple rhythm categories^[4]. Moreover, AI integration into ED workflows may assist in prioritizing high-risk patients, optimizing resource utilization, and reducing cognitive burden on physicians ^[5]. Despite promising results, real-world comparative evidence between AI-enabled ECG triage and standard physician-led triage in emergency departments remains limited, particularly in resource-constrained healthcare settings.

Aim

To compare the diagnostic accuracy of artificial intelligence–enabled ECG triage with standard physician-led triage for arrhythmia recognition in the emergency department.

Objectives

1. To assess the sensitivity and specificity of AI-enabled ECG interpretation in detecting cardiac arrhythmias.
2. To compare the diagnostic agreement between AI-based triage and physician-led ECG interpretation.
3. To evaluate the time taken for arrhythmia recognition using AI-enabled systems versus conventional physician triage.

Source of Data The data were collected from patients presenting to the emergency department who underwent 12-lead ECG evaluation as part of routine clinical assessment. ECG recordings and corresponding physician interpretations were retrieved from hospital electronic medical records. Study Design This was a prospective, cross-sectional comparative study. Study Location The study was conducted in the Emergency Department of a tertiary care teaching hospital. Study Duration The study was carried out over a period of 6 months. Sample Size A total of 80 patients who met the inclusion criteria were enrolled in the study. Inclusion Criteria • Patients aged ≥18 years presenting to the emergency department. • Patients who underwent 12-lead ECG recording during ED evaluation. • Patients with suspected cardiac symptoms such as palpitations, chest pain, syncope, or dyspnea. • Patients who provided informed consent. Exclusion Criteria • Poor-quality ECG recordings with significant artifacts. • Patients with incomplete clinical data. • Previously confirmed arrhythmia under active management. • Patients unwilling to participate. Procedure and Methodology Upon presentation to the emergency department, patients underwent a standard 12-lead ECG using calibrated digital ECG machines. The ECG recordings were simultaneously analyzed by an AI-enabled ECG interpretation software integrated into the hospital system. Independently, emergency physicians interpreted the ECGs as part of routine clinical care without knowledge of the AI output. The reference standard for final diagnosis was established by a senior cardiologist who reviewed the ECG along with relevant clinical information. Arrhythmias were categorized into predefined groups such as atrial fibrillation, atrial flutter, supraventricular tachycardia, ventricular tachycardia, bradyarrhythmias, heart blocks, and normal sinus rhythm. The time taken from ECG acquisition to arrhythmia recognition was recorded separately for AI interpretation and physician-led interpretation. Agreement between AI output and physician diagnosis was assessed. Sample Processing ECG waveforms were digitally stored in standardized format. The AI system processed the ECG signals using pre-trained deep learning algorithms to classify rhythm types. Physician interpretations were documented in structured clinical forms. Data were anonymized prior to analysis. Statistical Methods Data were entered into Microsoft Excel and analyzed using SPSS software version 25. Continuous variables were expressed as mean ± standard deviation, and categorical variables were expressed as frequencies and percentages. Sensitivity, specificity, positive predictive value, and negative predictive value were calculated for AI and physician interpretations using the cardiologist’s diagnosis as the gold standard. Cohen’s kappa coefficient was used to determine agreement between AI and physician interpretations. A p-value <0.05 was considered statistically significant. Data Collection A structured data collection sheet was used to record demographic details, presenting complaints, ECG findings, AI interpretation results, physician interpretation results, final cardiologist-confirmed diagnosis, and time taken for interpretation. Data confidentiality was strictly maintained throughout the study.

Table 1: To compare the diagnostic accuracy of artificial intelligence–enabled ECG triage with standard physician-led triage for arrhythmia recognition in the emergency department (N = 80)

Table 1 compares the diagnostic accuracy of artificial intelligence (AI)–enabled ECG triage with standard physician-led triage for arrhythmia recognition among 80 emergency department patients. AI-enabled triage correctly identified arrhythmias in 71 cases (88.8%; 95% CI: 79.7–94.7), whereas physician-led triage correctly detected arrhythmias in 66 cases (82.5%; 95% CI: 72.6–89.9). The difference in correct classification between the two methods was statistically significant (McNemar χ² = 4.12, p = 0.042). Incorrect classifications were lower with AI (11.2%) compared to physician interpretation (17.5%).

The overall diagnostic accuracy was significantly higher for AI (88.8 ± 4.6%) compared to physician-led triage (82.5 ± 5.3%), as demonstrated by paired t-test analysis (t = 3.27, p = 0.002). Furthermore, AI showed superior positive predictive value (89.6% vs 84.1%; p = 0.048) and negative predictive value (86.4% vs 79.2%; p = 0.037).

Table 2: To assess the sensitivity and specificity of AI-enabled ECG interpretation in detecting cardiac arrhythmias (N = 80)

(Gold standard: Cardiologist-confirmed diagnosis; Arrhythmia present = 46, absent = 34)

Table 2 presents the diagnostic performance of AI-enabled ECG interpretation using cardiologist-confirmed diagnosis as the gold standard. Among 46 patients with confirmed arrhythmias, AI correctly identified 42 cases (true positives, 91.3%; 95% CI: 79.2–97.6) and missed 4 cases (false negatives, 8.7%). Among 34 patients without arrhythmia, AI correctly classified 30 cases (true negatives, 88.2%; 95% CI: 72.6–96.7) and incorrectly labeled 4 cases as arrhythmia (false positives, 11.8%).

The sensitivity of AI in detecting arrhythmias was 91.3%, which was statistically significant (Z = 5.94, p < 0.001), while specificity was 88.2% (Z = 4.88, p < 0.001). The diagnostic odds ratio was 78.8 (95% CI: 18.4–336.2; p = 0.001), indicating strong discriminatory power of the AI system.

Table 3: To compare the diagnostic agreement between AI-based triage and physician-led ECG interpretation (N = 80)

(κ = 0.74 indicates substantial agreement)

Table 3 evaluates the diagnostic agreement between AI-based triage and physician-led ECG interpretation. Observed agreement between the two methods was seen in 69 out of 80 cases (86.3%; 95% CI: 76.8–92.9), while disagreement occurred in 11 cases (13.7%). Cohen’s kappa coefficient was 0.74 (95% CI: 0.59–0.88), which was highly statistically significant (Z = 6.21, p < 0.001). A kappa value of 0.74 indicates substantial agreement between AI and physician interpretations.

The McNemar test further showed a statistically significant difference in paired comparisons (χ² = 3.86, p = 0.049), suggesting that although agreement was high, AI demonstrated a modest advantage in arrhythmia recognition over physician-led interpretation.

Table 4: To evaluate the time taken for arrhythmia recognition using AI-enabled systems versus conventional physician triage (N = 80)

Table 4 compares the time taken for arrhythmia recognition using AI-enabled systems versus conventional physician triage. The mean time for arrhythmia recognition using AI was significantly shorter (1.82 ± 0.64 minutes; 95% CI: 1.67–1.97) compared to physician interpretation (4.73 ± 1.21 minutes; 95% CI: 4.46–5.00). The difference was highly statistically significant (paired t-test = 17.94, p < 0.001).

The mean time difference between the two methods was 2.91 ± 0.88 minutes (95% CI: 2.71–3.11), indicating that AI reduced recognition time by nearly 3 minutes on average. The minimum time range also favored AI (0.95–3.44 minutes) compared to physician-led triage (2.11–7.82 minutes).

The present study demonstrated that AI-enabled ECG triage showed superior diagnostic performance compared to standard physician-led triage in the emergency department. In Table 1, AI correctly detected arrhythmias in 88.8% of cases compared to 82.5% by physicians (p = 0.042), with significantly higher overall diagnostic accuracy (88.8% vs 82.5%, p = 0.002). These findings are consistent with Bacon J. (2025)[6], who reported cardiologist-level arrhythmia detection using deep neural networks with overall accuracy exceeding 90%. Similarly, Kanchayawong P et al. (2025)[2] demonstrated that AI-based ECG algorithms achieved high diagnostic accuracy comparable to expert cardiologists, particularly for atrial fibrillation and conduction abnormalities.

The higher positive predictive value (89.6%) and negative predictive value (86.4%) observed in the present study align with findings by Du Y et al. (2025)[3], who reported robust predictive capabilities of AI-enabled ECG systems in detecting cardiac abnormalities in clinical practice. Moreover, Spethmann S et al. (2024)[4] highlighted variability in physician ECG interpretation, particularly in emergency settings, supporting the notion that AI can reduce inter-observer variability and improve consistency.

In Table 2, AI demonstrated high sensitivity (91.3%) and specificity (88.2%), both statistically significant (p < 0.001). These values are comparable to those reported by Chen Y et al. (2025)[5], who showed AI sensitivity of approximately 90–95% across multiple arrhythmia categories. Likewise, Mastoris I et al. (2023)[7] reported high diagnostic accuracy of deep learning systems in medical image interpretation, emphasizing the broader applicability of AI in clinical diagnostics. The diagnostic odds ratio of 78.8 in the present study indicates strong discriminatory performance, consistent with large validation studies demonstrating high AI robustness across diverse ECG datasets.

Table 3 revealed substantial agreement between AI and physician interpretations, with a Cohen’s kappa of 0.74 (p < 0.001). This level of agreement is similar to findings by Alqahtani RM et al. (2024)[8], who noted moderate to substantial agreement among physicians, but with room for improvement. AI integration appears to enhance agreement and reduce variability. McNemar test significance (p = 0.049) suggests that AI may provide incremental diagnostic advantage over conventional triage. Rainer R et al. (2024)[9]

Time efficiency analysis in Table 4 showed a markedly reduced time to arrhythmia recognition using AI (1.82 ± 0.64 minutes) compared to physician triage (4.73 ± 1.21 minutes), with a mean difference of 2.91 minutes (p < 0.001). This finding is clinically significant, particularly in time-sensitive arrhythmias such as ventricular tachycardia or complete heart block. Similar workflow efficiency improvements have been described by Muley A et al. (2025)[10] & Maxwell S et al. (2025)[11], who emphasized AI’s ability to reduce cognitive load and expedite clinical decision-making in acute care settings.

The present study demonstrated that artificial intelligence–enabled ECG triage showed superior diagnostic accuracy compared to standard physician-led triage for arrhythmia recognition in the emergency department. AI achieved higher correct detection rates, improved sensitivity and specificity, and significantly better overall diagnostic accuracy. Additionally, AI-based interpretation showed substantial agreement with physician assessment while significantly reducing the time required for arrhythmia recognition. The reduction in diagnostic time by nearly three minutes highlights the potential clinical impact of AI integration in emergency workflows, particularly in time-critical arrhythmias where early detection influences treatment outcomes. While physician expertise remains essential, AI-enabled ECG systems function as effective decision-support tools that enhance diagnostic precision, reduce variability, and improve operational efficiency. LIMITATIONS OF THE STUDY The study sample size was relatively small (n = 80), which may limit generalizability of the findings. The study was conducted in a single tertiary care center, and results may differ in other healthcare settings with varying patient populations. The AI system evaluated was a specific algorithm; performance may vary across different AI platforms. Rare or complex arrhythmias were limited in number, potentially affecting subgroup analysis accuracy. Physician interpretation was performed in real-time emergency settings, where workload and fatigue could influence performance. The study did not assess long-term clinical outcomes related to AI-guided triage decisions.

1. Tyler S, Olis M, Aust N, Patel L, Simon L, Triantafyllidis C, Patel V, Lee DW, Ginsberg B, Ahmad H, Jacobs RJ. Use of artificial intelligence in triage in hospital emergency departments: a scoping review. Cureus. 2024 May 8;16(5).
2. Kanchayawong P, Aramvanitch K, Yuksen C, Trakulsrichai S, Sricharoen P, Suwatcharangkoon S, Sirintaranont P, Keandoungchun J, Nuanprom P, Jenpanitpong C, Jaiboon S. Real-Time Telemedical Oversight Improves Prehospital Stroke Metrics: A Five-Year Cohort Study. Archives of Academic Emergency Medicine. 2025 Jun 25;13(1):e57.
3. Du Y, Yang P, Liu Y, Deng C, Li X. Artificial intelligence in chronic disease self-management: current applications and future directions. Frontiers in Public Health. 2025 Nov 20;13:1689911.
4. Spethmann S, Hindricks G, Koehler K, Stoerk S, Angermann CE, Boehm M, Assmus B, Winkler S, Moeckel M, Mittermaier M, Lelgemann M. Telemonitoring for Chronic Heart Failure: Narrative Review of the 20-Year Journey From Concept to Standard Care in Germany. Journal of medical Internet research. 2024 Dec 4;26:e63391.
5. Chen Y, Ferguson C, Cartledge S, Colgan J, Hendriks JM, Keller K, Lin FF. Nurse educators’ expectations, training, and assessments of electrocardiogram interpretation among Australian acute care nurses: a national survey. European Journal of Cardiovascular Nursing. 2025 May 9:zvaf088.
6. Bacon J. Meeting abstracts of the Toronto Resuscitation Conference. Scandinavian Journal of Trauma, Resuscitation and Emergency Medicine. 2025;33:192.
7. Mastoris I, Gupta K, Sauer AJ. The war against heart failure hospitalizations: remote monitoring and the case for expanding criteria. Cardiology Clinics. 2023 Nov 1;41(4):557-73.
8. Alqahtani RM, Hassan Sayed TO, Saud Alharbi AM, Alqahtani FM, Alharbi NM, Alharbi AH, Aljarboua IA, Aljurbua AA, Alrashidi NA. Advanced Nursing Interventions and Emergency Management in Cardiac Arrest: A Comprehensive Approach. Journal of International Crisis & Risk Communication Research (JICRCR). 2024 Jul 8;7.
9. Rainer R, Bambach K. Navigating Supervision of. Risk Management in Emergency Medicine, An Issue of Emergency Medicine Clinics of North America: Risk Management in Emergency Medicine, An Issue of Emergency Medicine Clinics of North America, E-Book. 2024 Nov 12;43(1):131.
10. Muley A, Cheemalapati S, Muzumdar P. A Systematic Literature Review of Machine Learning Methods in Healthcare. Asian Journal of Medicine and Health. 2025 Dec 9;23(12):78-102.
11. Maxwell S, Morrissey B. Paramedicine Literature Search: March-May 2025. International Journal of Paramedicine. 2025 Jul 8(11):83-132.