Electrocardiographic biomarker based on machine learning for detecting overt hyperthyroidism

Figures

• 

  Figure 1. Study flow chart. To develop the deep learning model and conduct internal performance tests, we obtained patient data from a tertiary teaching hospital (Ajou University Medical Center, Hospital A) in Republic of Korea, from 1 January 1994 to 31 December 2020. We also conducted an external performance test using a dataset from a community-based secondary hospital (Incheon Sejong Hospital, Hospital B) from 1 March 2017 to 31 May 2021. The study population included patients who visited the hospital and underwent at least one standard 12-lead electrocardiogram and at least one thyroid function test within 24 h before and after the index electrocardiogram. In both hospitals, we obtained thyroid function test values, age, sex, and digitally stored electrocardiogram data. We excluded 21 and 7 patients, respectively, whose sex or age information was missing. As a result, 113 194 and 33 478 patients in Hospitals A and B, respectively, were included. For development, tuning, and internal validation, we divided the data from Hospital A by the patient (7:1:2).

• 

  Figure 2. Architecture of deep learning model for detecting hyperthyroidism. We used only 12-lead electrocardiogram waveform data as a predictor variable in the deep learning model. We removed baseline wander noise <0.5 Hz for preprocessing using a Butterworth pass filter and normalized the data by standard normalization.12 We trimmed both sides by 0.9 s to remove noise at the start and end of the electrocardiogram. Finally, we used a 4096-size vector as the input for the deep learning model. We utilized six residual blocks of the ResNet.13 Each residual block consists of a convolutional neural network, batch normalization, ReLU activation function, and a dropout layer. The filter size was set to 21. Six residual blocks were used, and the length of the input was reduced by half every time the three residual blocks passed. Each electrocardiogram lead had a different ResNet model. At the end of the residual block, the outputs were channel-wise average pooled and concatenated with all outputs from each lead. Using the concatenated output, dense layers finally predict the probability (0–1) of overt hyperthyroidism. The above hyperparameters were selected through GridSearch.

• 

  Figure 3. Performance of deep learning-based model for detecting hyperthyroidism. To maximize the utility of the deep learning model in various settings, we modified the 12-lead deep learning model to partially lead deep learning model, using only six limb leads (lead I, II, III, aVL, aVR, and aVF), six precordial leads (V1, V2, V3, V4, V5, and V6), or a single-lead (lead I) electrocardiogram. We validated the developed models using internal validation and external validation sets. The performance metrics were the area under the receiver operating characteristic curve, sensitivity, specificity, negative predictive value, and positive predictive value. To calculate the sensitivity, specificity, positive predictive value, and negative predictive value, we calculated the cut-off point using Youden’s J statistics in the development dataset and applied the cut-off point to internal and external datasets.15 The 95% confidence intervals were determined using the Sun-Su optimization of the De-long method.16

• 

  Figure 4. Cumulative hazard of developing hyperthyroidism in patients with an initially normal. We hypothesized that, even in normal patients, a high probability of the model could be a biomarker for predicting overt hyperthyroidism. To prove this hypothesis, we conducted a sub-analysis of 6800 patients with initial normal thyroid function test and follow-up thyroid function test, at least 4 weeks apart from the initial thyroid function test, from among the internal validation datasets. We used Youden’s J statistics with the development dataset to determine cut-off points,15 defined a group with a probability of deep learning model greater than or equal to the cut-off as a high-risk group and defined others as a low-risk group. Using Kaplan–Meier estimation, we analysed whether patients in the high-risk (n = 6323) and low-risk groups (n = 568) developed new overt hyperthyroidism within 36 months and compared the statistical differences between the two survival curves using a log-rank test. Thirty people in the high-risk group show overt hyperthyroidism, and 145 people in the low-risk group show overt hyperthyroidism. The incidence rate of overt hyperthyroidism is significantly higher in…