Key result
AI-enabled 12-lead ECGs detect multiple blood abnormalities, showing strongest performance for hypokalemia with ~0.83 ROC-AUC.
Why the study?
Bedside ECG telemetry rarely exploits the full informational content of the signal, and emerging evidence suggests that continuous ECG could serve as a non-invasive surrogate for laboratory monitoring by inferring systemic biological variables from raw ECG waveforms.
Does a machine-learning framework applied to standard 12-lead and single-lead ECG signals accurately classify abnormalities in routine blood chemistry parameters in hospitalized patients?
Cohort (n=109,956)
Does a machine-learning framework applied to standard 12-lead and single-lead ECG signals accurately classify abnormalities in routine blood chemistry parameters in hospitalized patients?
Answers are grounded in this paper and cite their sources.
Standard ECG waveforms contain exploitable signatures of systemic biochemical alterations that allow machine-learning models to non-invasively screen for multiple blood laboratory abnormalities, particularly hypokalemia, with moderate-to-good discrimination.
Electrocardiographic (ECG) monitoring is a cornerstone of inpatient surveillance, yet bedside telemetry rarely exploits the full informational content of the signal. Emerging evidence that systemic biological variables can be inferred from raw ECG waveforms suggests that continuous ECG could serve as a non-invasive surrogate for laboratory monitoring. We developed and validated an end-to-end machine-learning framework to infer abnormalities in key blood parameters directly from standard ECG recordings. Using MIMIC-IV ECG traces paired with laboratory measurements acquired within 30 minutes, we trained two convolutional neural network architectures to classify 55 parameter/threshold combinations across 29 routinely measured laboratory parameters into normal and abnormal ranges, using 12-lead and single-lead (II) inputs combined with age and gender. Data were temporally split by subject to mimic prospective validation. The best models were further evaluated on long-term ICU ECG traces from MIMIC-III-WDB. Performance was assessed via ROC-AUC, average precision, sensitivity, specificity, PPV, NPV, and F1-score with bootstrap 95% CIs. The final datasets comprised 266,896 samples from 104,664 subjects (MIMIC-IV) and 79,855 samples from 5,292 subjects (MIMIC-III). In the 12-lead setting, 20 of 55 combinations achieved a lower ROC-AUC 95% CI bound above 0.7; hypokalemia (potassium < 3.3 mEq/L) performed best (ROC-AUC 0.834 [0.817–0.849]). Single-lead performance decreased only slightly, whereas generalization to ICU single-lead recordings declined. Saliency analysis revealed physiologically plausible focus, notably on the T-wave for hypokalemia. Standard ECGs contain exploitable signatures of systemic biochemical alteration, enabling classification of multiple laboratory abnormalities with moderate-to-good discrimination. Precision limitations caution against interruptive alerting for rare endpoints, but ECG-based estimation holds promise as a non-invasive screening and triage tool in acute care.
No takes yet. Share an insight, caveat, or question.
Maletzky et al. (2026) conducted a cohort in Inpatient and ICU patients with paired ECG traces and laboratory measurements (n=109,956). Convolutional neural network analysis of 12-lead and single-lead (II) ECG combined with age and gender vs. Laboratory measurements acquired within 30 minutes of ECG was evaluated on Classification of 55 parameter/threshold combinations across 29 routinely measured laboratory parameters (best performing: hypokalemia < 3.3 mEq/L) (ROC-AUC 0.834, 95% CI 0.817-0.849). Convolutional neural networks applied to 12-lead ECGs classified 20 of 55 blood abnormalities with lower.7, best detecting hypokalemia.
Synapse has enriched 10 related papers. Consider them for comparative context: