Non-invasive hemoglobin estimation with outcome prediction via deep learning analysis of ECG-derived cardiac...

Anemia is a global health issue, especially in resource-limited areas, where traditional hemoglobin (Hb) testing is invasive and costly. This study aimed to develop an electrocardiogram-hemoglobin (ECG-Hb) deep learning model (DLM) for detecting anemia and assess its impact on all-cause mortality and new-onset heart failure.

This retrospective study analyzed ECGs and corresponding Hb levels from two hospitals. The DLM was trained on 388,166 ECGs from 187,202 patients and tested on 24,279 and 29,247 patients in internal and external sets, respectively. Anemia was defined as moderate (Hb ≤ 10 g/dL) or severe (Hb ≤ 8 g/dL). Diagnostic performance was evaluated using receiver operating characteristic (ROC) curve analysis, and an 8-year follow-up assessed mortality and heart failure risk with Cox regression.

The areas under the ROC curves (AUCs) for detecting moderate-to-severe anemia were 0.8545 (internal) and 0.8243 (external), with sensitivities of 65.9% and 71.0%, and specificities of 84.8% and 77.4%, respectively. ECG-Hb performed better in detecting severe anemia (AUC = 0.9038/0.8766) than in mild anemia. Pearson correlations between ECG-Hb and Hb were 0.56 (internal) and 0.53 (external). Key ECG features, including heart rate variability, significantly influenced ECG-Hb. Patients with severely low ECG-Hb had higher risks of mortality (hazard ratio [HR]: 1.71, 95% confidence interval [CI]: 1.42-2.06) and heart failure (HR: 2.47, 95% CI: 2.07-2.94) compared to those with standard ECG-Hb levels.

The ECG-Hb DLM offers strong diagnostic and prognostic potential for anemia and cardiovascular risks, making it a valuable, non-invasive screening tool in low-resource settings.