Artificial Intelligence-Assisted Electrocardiography for Early Diagnosis of Thyrotoxic Periodic Paralysis

Abstract

Context: Thyrotoxic periodic paralysis (TPP) characterized by acute weakness, hypokalemia, and hyperthyroidism is a medical emergency with a great challenge in early diagnosis since most TPP patients do not have overt symptoms.

Objective: This work aims to assess artificial intelligence (AI)-assisted electrocardiography (ECG) combined with routine laboratory data in the early diagnosis of TPP.

Methods: A deep learning model (DLM) based on ECG12Net, an 82-layer convolutional neural network, was constructed to detect hypokalemia and hyperthyroidism. The development cohort consisted of 39 ECGs from patients with TPP and 502 ECGs of hypokalemic controls; the validation cohort consisted of 11 ECGs of TPP patients and 36 ECGs of non-TPP individuals with weakness. The AI-ECG-based TPP diagnostic process was then consecutively evaluated in 22 male patients with TTP-like features.

Results: In the validation cohort, the DLM-based ECG system detected all cases of hypokalemia in TPP patients with a mean absolute error of 0.26 mEq/L and diagnosed TPP with an area under curve (AUC) of approximately 80%, surpassing the best standard ECG parameter (AUC = 0.7285 for the QR interval). Combining the AI predictions with the estimated glomerular filtration rate and serum chloride boosted the diagnostic accuracy of the algorithm to AUC 0.986. In the prospective study, the integrated AI and routine laboratory diagnostic system had a PPV of 100% and F-measure of 87.5%.

Conclusion: An AI-ECG system reliably identifies hypokalemia in patients with paralysis, and integration with routine blood chemistries provides valuable decision support for the early diagnosis of TPP.

Keywords: artificial intelligence; deep learning; electrocardiogram; hypokalemia; thyrotoxic periodic paralysis.

Figure 1.

Comparison between electrocardiography (ECG)-based potassium (K⁺) prediction and laboratory (LAB) K⁺ in the validation cohort. Each line represents a hypokalemia case. The patients with absolute error (AE) greater than 0.3 are colored red, and the others are colored green. The t test shows the mean AE (MAE) differences are not significantly different for thyrotoxic periodic paralysis (TPP) vs non-TPP (P = .409).

Figure 2.

Performance comparisons of electrocardiography (ECG) morphologies and deep learning models trained using 3 different weighting strategies in the validation cohort. The receiver operating characteristic curves were made by the predictions of the deep learning model (DLM) or each ECG morphology. The ECG morphology curves were generated from logistic regression using the development cohort. The DLM score 1 was trained using the raw data set; score 2 was trained using an age-matched strategy; and score 3 was trained using an age- and K⁺-matched strategy.

Figure 3.

Receiver operating characteristic (ROC) curves for combining patient characteristics with deep learning models in the validation cohort. The ROC curves for clinical characteristics were made by logistic regression. The combination models were generated for each score with the listed clinical characteristics. Score 1 was trained using the raw data set; score 2 was trained using an age-matched strategy; and score 3 was trained using an age- and K⁺-matched strategy.

Figure 4.

The prospective integrated artificial intelligence–electrocardiography (AI-ECG) diagnostic algorithm for actively identifying potential thyrotoxic periodic paralysis (TPP) cases. Male patients with metabolic paralysis after physical examination were included. The boxes at each step denotes the patients who progressed toward a diagnosis of TPP.