Deep learning using wearable single-lead electrocardiograms for early detection of post-thyroidectomy hypocalcemia

Timely recognition of post-thyroidectomy hypocalcemia remains challenging. We evaluated whether deep learning models using wearable single-lead electrocardiograms detect biochemical early warning hypocalcemia, defined as albumin-corrected calcium <8.5 mg/dL, compared with a heart rate-corrected QT interval benchmark.

In a prospective cohort (N = 51), at baseline and postoperative day 1, postoperative day 2, 2-3 weeks, and 2-3 months, we obtained synchronized laboratory measurements and four 30-second wearable single-lead electrocardiogram recordings within 60 minutes and recorded symptom severity. Data were split at the patient level (no overlap) into internal and external cohorts for training, validation, and testing. We evaluated DenseNet, EfficientNet, ResNet, ResNeXt, and RegNet architectures. The primary end point was the area under the receiver operating characteristic curve for corrected calcium <8.5 mg/dL. Corrected QT interval from contemporaneous 12-lead electrocardiograms served as the non-deep learning benchmark.

We analyzed 490 electrocardiogram images from 38 patients (internal: 31; external: 7). DenseNet achieved area under the receiver operating characteristic curves of 0.878 (internal) and 0.870 (external) for corrected calcium <8.5 mg/dL, whereas postoperative day 1 12-lead corrected QT interval showed limited discrimination (area under the receiver operating characteristic curve, 0.526). In the full cohort (N = 51), the prevalence of biochemical early warning hypocalcemia was 60.8% (postoperative day 1), 58.8% (postoperative day 2), 17.6% (2-3 weeks), and 29.4% (2-3 months); the prevalence of biochemically significant hypocalcemia with low parathyroid hormone was 27.5%, 27.5%, 2.0%, and 0%, respectively.

A wearable single-lead electrocardiogram-based deep learning model demonstrated strong discriminative performance for biochemical hypocalcemia after thyroidectomy. Compared with corrected QT interval-based assessment, the model achieved superior performance. Larger multi-institutional validation is warranted before clinical deployment.