Towards Remote Monitoring of Sickle Cell Disease for Heart Failure: A Single Lead ECG-AI Model

Background

Cardiopulmonary dysfunction, including heart failure (HF), is a leading cause of death in patients with sickle cell disease (SCD). Early detection of HF in SCD patients is critical for timely intervention and improved outcomes. Conventional diagnostic modalities such as echocardiograms and cardiac MRI are often difficult to obtain in resource-limited environments or remote settings. There is a need for low cost and widely accessible tools to assist with early recognition of HF among SCD patient.

Hypothesis

We hypothesize that artificial intelligence (AI) models using single lead ECG as an input can accurately detect SCD patients with HF.

Methods/Approach

An ECG-AI model based on convolutional neural networks was previously developed to classify HF patients, utilizing a large ECG repository from Wake Forest Baptist Health (WFBH). This model was trained on 1,078,198 digital Lead I ECG of standard 12 lead 10 seconds clinical ECG records from 165,243 patients, with a demographic distribution of 73% White, 19% Black, and 52% female individuals, and a mean age (SD) of 58 (15) years. An AUC of 0.84 was achieved in identifying HF patients from controls in a hold-out validation – unseen subset of WFBH. In the current study, the ECG-AI model was externally validated on data from SCD patients at the University of Tennessee Health Science Center (UTHSC). Additionally, a logistic regression (LR) model was constructed for the UTHSC cohort, incorporating basic demographic variables along with the outcomes of the ECG-AI model.

Results/Data

The external validation cohort from UTHSC comprised data from 2,107 SCD patients, including 188 with HF and 1,919 without HF. The cohort was predominantly Black (98%), with 72% female individuals and a mean age of 39 (SD 14) years. Despite notable demographic differences—such as a higher proportion of Black patients and a younger average age compared to the derivation cohort—our single lead ECG-AI model yielded an AUC of 0.80 (0.77-0.82) in detecting HF within the UTHSC SCD population. When the ECG-AI outcome, representing an ECG-based risk score between 0 and 1, was combined with age and sex in a logistic regression (LR) model, a moderate improvement in AUC was observed (DeLong Test, p<0.01), reaching 0.82 (0.79-0.84) with a sensitivity of 0.73 (0.70-0.76) and a specificity of 0.73 (0.71-0.76).

Conclusion

The ECG-AI model has demonstrated moderately high accuracy in detecting HF in SCD patients using single Lead I of clinical 12-lead ECGs. Knowing that this is the lead mimicked by smartwatches and other wearables with ECG functionality, this study shows feasibility of remote monitoring of SCD patients for HF risk. Future studies will focus on calibration and testing of our ECG-AI models specifically for SCD patients and smartwatch ECGs.