AI-driven ECG analysis via a smartphone app provided a rapid, accessible, and accurate screening tool for LV dysfunction in sepsis and outperformed NT-proBNP, with an ROC-AUC of 0.901 and a negative predictive value of 97.5%.
Key Findings
Results
The AI-based smartphone ECG application demonstrated high diagnostic performance for detecting LV dysfunction in sepsis patients, with an ROC-AUC of 0.901 across the multicenter cohort.
Study included 425 adult sepsis patients from three university hospitals.
All patients underwent both 12-lead ECG and transthoracic echocardiography within 72 hours of emergency department admission.
LV dysfunction was defined as left ventricular ejection fraction ≤ 40%.
ECG images displayed on a monitor were photographed and analyzed using an AI-based smartphone application (ECG Buddy™), which generated a quantitative digital biomarker called QCG-LVDys.
Results
Using an optimized threshold of 37.35 for QCG-LVDys, the application achieved a sensitivity of 89.6%, specificity of 75.1%, and a negative predictive value of 97.5%.
The high NPV of 97.5% suggests potential utility for ruling out LV dysfunction in sepsis patients.
The optimized threshold was 37.35 for the QCG-LVDys score.
These metrics were derived from the overall multicenter cohort of 425 patients.
The authors noted the high NPV suggests potential utility for early risk stratification and informing cautious hemodynamic management.
Results
QCG-LVDys significantly outperformed NT-proBNP in diagnosing LV dysfunction in the subset where both measures were available.
ROC-AUC for QCG-LVDys was 0.911 versus 0.637 for NT-proBNP (P < 0.001).
The comparison was performed in the single center where NT-proBNP was routinely measured.
NT-proBNP has been noted to have limited diagnostic accuracy in the sepsis setting.
This comparison was a predefined analysis in the study protocol.
Results
Higher QCG-LVDys values were associated with increased lactate levels and greater 28-day in-hospital mortality.
Association with increased lactate levels: P = 0.002.
Association with 28-day in-hospital mortality: P < 0.001.
These associations suggest the biomarker may reflect broader hemodynamic compromise beyond ejection fraction alone.
The findings indicate potential prognostic as well as diagnostic utility of QCG-LVDys in sepsis.
Background
Sepsis-induced LV dysfunction complicates fluid resuscitation by increasing the risk of pulmonary edema, and echocardiographic assessment is often delayed in clinical practice.
Natriuretic peptides such as NT-proBNP have limited diagnostic accuracy in the sepsis setting.
The study was designed as a multicenter retrospective observational study across three university hospitals.
The authors identified a clinical gap: standard echocardiography is often not immediately available, motivating the need for an accessible screening tool.
The smartphone-based approach was designed to be rapid and accessible without requiring dedicated echocardiography equipment.
Conclusions
The study was a multicenter retrospective observational design, and the authors noted that prospective studies are needed to define the application's role in sepsis treatment protocols.
Data came from three university hospitals in a retrospective design.
Total cohort: 425 adult sepsis patients.
The authors explicitly stated: 'prospective studies are needed to define its role in sepsis treatment protocols.'
The retrospective design limits causal inference and generalizability to real-time clinical decision-making workflows.
What This Means
This research suggests that a smartphone application using artificial intelligence to analyze ECG images can accurately detect weak heart pumping function (left ventricular dysfunction) in patients with sepsis — a life-threatening infection. The study tested 425 sepsis patients across three hospitals and found that the app's AI-generated score correctly identified heart dysfunction with about 90% accuracy (as measured by a standard diagnostic metric called AUC). Importantly, when the score was low, it was almost always correct that the patient did NOT have heart dysfunction (97.5% negative predictive value), meaning it could reliably reassure clinicians when heart problems are not present.
The smartphone app also substantially outperformed a commonly used blood test called NT-proBNP, which is supposed to indicate heart stress but performed poorly in sepsis patients (AUC 0.637 vs. 0.911 for the app). Additionally, higher app scores were linked to worse outcomes, including higher blood lactate levels (a marker of tissue stress) and greater risk of death within 28 days. This suggests the tool may have both diagnostic and prognostic value.
This research matters because sepsis patients often need careful fluid management — giving too much fluid to someone with a weak heart can cause dangerous fluid buildup in the lungs, but giving too little can worsen organ failure. Currently, assessing heart function requires an echocardiogram (ultrasound of the heart), which can be delayed or unavailable. A smartphone app that analyzes a standard ECG image could offer a faster, more accessible way to screen for heart problems early in sepsis care. However, the study was retrospective, and the authors emphasized that prospective trials are still needed before this tool can be formally integrated into treatment guidelines.
Park Y, Kim J, Jo Y, Kwon W, Song K, Lee H, et al.. (2026). Screening for Left Ventricular Dysfunction in Sepsis Using a Smartphone ECG Analysis Application: A Multicenter Validation Study.. Journal of Korean medical science. https://doi.org/10.3346/jkms.2026.41.e217