A machine learning-assisted system to predict thyrotoxicosis using patients' heart rate monitoring data: a retrospective cohort study

Abstract

Previous studies have shown a correlation between resting heart rate (HR) measured by wearable devices and serum free thyroxine concentration in patients with thyroid dysfunction. We have developed a machine learning (ML)-assisted system that uses HR data collected from wearable devices to predict the occurrence of thyrotoxicosis in patients. HR monitoring data were collected using a wearable device for a period of 4 months in 175 patients with thyroid dysfunction. During this period, 3 or 4 thyroid function tests (TFTs) were performed on each patient at intervals of at least one month. The HR data collected during the 10 days prior to each TFT were paired with the corresponding TFT results, resulting in a total of 662 pairs of data. Our ML-assisted system predicted thyrotoxicosis of a patient at a given time point based on HR data and their HR-TFT data pair at another time point. Our ML-assisted system divided the 662 cases into either thyrotoxicosis and non-thyrotoxicosis and the performance was calculated based on the TFT results. The sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of our system for predicting thyrotoxicosis were 86.14%, 85.92%, 52.41%, and 97.18%, respectively. When subclinical thyrotoxicosis was excluded from the analysis, the sensitivity, specificity, PPV, and NPV of our system for predicting thyrotoxicosis were 86.14%, 98.28%, 94.57%, and 95.32%, respectively. Our ML-assisted system used the change in mean, relative standard deviation, skewness, and kurtosis of HR while sleeping, and the Jensen-Shannon divergence of sleep HR and TFT distribution as major parameters for predicting thyrotoxicosis. Our ML-assisted system has demonstrated reasonably accurate predictions of thyrotoxicosis in patients with thyroid dysfunction, and the accuracy could be further improved by gathering more data. This predictive system has the potential to monitor the thyroid function status of patients with thyroid dysfunction by collecting heart rate data, and to determine the optimal timing for blood tests and treatment intervention.

Figure 1

Distribution of free T4 and TSH levels of data pairs according to the decision of the ML-assisted system. T4 thyroxine, TSH thyroid stimulating hormone, ML machine learning.

Figure 2

Sensitivity and Specificity for varying the number of days of collecting HR (A) for both training and test data and (B) for the test data when the duration of collecting HR data is fixed at 10 days for the training data. HR, heart rate.

Figure 3

Rank for each feature in the experiment. Feature importance considering (A) gain and split, (B) gain, and (C) split. Gain, relative contribution of a feature in the model; Split, relative count of times a feature occurs in the model.

Figure 4

Design of a ML-assisted system to predict the occurrence of thyrotoxicosis using HR data collected from wearable devices. ML machine learning, HR heart rate, T4 thyroxine, TSH thyroid stimulating hormone, RSD relative standard deviation, JS Div Jensen–Shannon divergence, TFT thyroid function test.