Age and Sex Estimation Using Artificial Intelligence From Standard 12-Lead ECGs

We identified 774 783 adult patients (18 years or older) with at least one digital, standard 10 second 12-lead ECG acquired in the supine position between January 1994 and February 2017. Data was collected from the Mayo Clinic digital data vault, with institutional review board approval. For patients with multiple ECGs, only the earliest ECG was selected.

ECGs were acquired at a sampling rate of 500 Hz using a GE-Marquette ECG machine (Marquette, WI) and stored using the MUSE data management system.

We implemented a convolutional neural network (CNN)¹¹ using the Keras Framework with a Tensorflow (Google, Mountain View, CA) backend and Python. While CNNs are mainly applied to images, we adjusted the network architecture to have spatial and temporal feature extraction layers. The network operates by adjusting the weights of the convolutional filters during training to extract meaningful and relevant features in an unsupervised way. Both the sex and age network was built using stacked blocks of convolutional, max pooling, and batch normalization¹²; each block was followed by a nonlinear activation function. After the first group of blocks extracted temporal features, another spatial block was used to fuse data from all leads, and then the extracted features were used in a fully connected network. The difference between the age and sex networks was in the output layer. The sex detection network had 2 outputs (male and female) and was activated with a Softmax output, while the age network had a single output (age) as a continuous number, without a following nonlinear function (linear activation). Although we used the same data for training both the sex and age algorithms, the networks were trained separately, but with the same data split for training, validation, and testing sets. Of note, the only data inputted for training were the raw digital 12-lead ECG signal and the associated age and self-reported sex of each individual. Figure 1 summarizes the network architecture for each network. Hyperparameters (batch size, initial learning rate, number of neurons in the fully connected layers, and number of convolutional layers) were changed during training to get the optimal model based on the validation set. Initial learning rates tested were in the range 3e–3 to 1e–5, and batch size was in the range of 16 to 128. Selected hyperparameters were learning rate of 3e–4 and batch size of 64 for the age network and 16 for the sex network.

ECGs have some biometric features that might be learnt in the process of training the network. To avoid cross-contamination of the training and testing data sets, only a single ECG per patient was used. After an initial split into the development (64% of the population) and holdout (36% of the population) data sets, the development data set was further divided into training (75% of the development set) and internal validation data (25%). The training and validation sets were mutually exclusive with respect to patient identification. For training both networks, ECGs were fed to the network, and the network weights were optimized using the Adam optimizer¹³ with categorical cross-entropy as the loss function for the sex network and mean squared error for the age network. The internal validation data set was used to select the optimal model and hyper parameters and to find the optimal threshold for sex classification.

Separate models for sex and age were tested using the same holdout data set. The sex network is a binary classifier (male versus female), so the model output is the probability that the ECG was obtained from a male–denoted as P (convolutional CNN-predicted sex). The output P is a continuous number in the range 0 to 1, and in cases of inconclusive ECG-sex discrimination, P will be closer to 0.5. Due to the nonlinear nature of the network and the uneven number of male and female ECGs used for training, a trivial threshold of 0.5, meaning P>0.5 implies male, may not be the optimal threshold to maximize classification accuracy for males and females. To address this, a receiver operating characteristic curve was created using P as the linear discriminator to estimate the area under the curve (AUC) and accuracy at various candidate thresholds. An optimal threshold was selected using the internal validation data set, and uncertainty in classification was evaluated using the output probability P to analyze if the model could evaluate its own classification results by rejecting samples with low certainty. To evaluate whether specific portions of the ECG were critical in arriving at a specific sex estimation, separate neural networks were trained on 35 000 independent individuals using discrete portions of the ECG (P wave, QRS complex, etc), and similarly a receiver operating characteristic curve was generated to determine accuracy in prediction of sex. For the age estimation network, the network output was an estimate for age based on the ECG as an input (CNN-predicted age). The R square metric and the mean average error metric were used to evaluate the network. In the training stage, the networks were fed the whole training set, and in each iteration (epoch) the model was evaluated on the validation set, a software callback that saved only the best model based on the validation performance was used. The training was stopped if, for 5 consecutive epochs, the validation performance did not improve (to avoid training in the overfitting stage). The age prediction optimal model was trained for 20 epochs while the sex detection model achieved optimal results after 12.

Within the data set, an algorithm to extract those patients with multiple ECGs recorded over a minimum 2 decades of life was created, and 100 patients were randomly selected. The choice of 100 patients was (1) to permit feasibility of deep chart review and (2) given this was meant to be an exploratory analysis of how this algorithm might perform in a random limited subset of patients. All ECGs at every chronologic year of life were evaluated using the trained network for age, and CNN-predicted age was plotted against the chronologic age at which that ECG was obtained. If multiple ECGs were obtained at any year of life, the median CNN-predicted age at that year of life was used. After this, clinical histories for patients were extracted via chart review. Clinical comorbidities, including prior history of myocardial infarction, low ejection fraction, coronary disease, hypertension, diabetes mellitus, atrial arrhythmias, and prior history of cardiac surgery were included. In addition, incident events over the follow-up period including hospitalization for hemodynamic shock, incident cancer, incident comorbidities that were first diagnosed over the follow-up period, new diagnoses of coronary disease, heart failure, or thromboembolic events, and cardiac transplantation were recorded. The R-squared correlation values between CNN-predicted and chronologic age were then calculated per patient and effect of prior comorbidities, and incident events on the R-squared or differences between CNN-predicted and chronologic age were determined. To compare impact of preexisting comorbidities on deviation of CNN-predicted age from chronologic age over time, the mean deviation of CNN-predicted age from chronologic age was calculated over the course of the follow-up period. The frequency of specific comorbidities among patients whose CNN-predicted exceeded the chronologic age by greater than 7 years, was within 7 years of the chronologic age, or was less than the chronologic age by 7 years was recorded. χ² analysis was used to determine whether frequency of comorbidities was significantly different between groups. P<0.01 was considered significant.

A total of 774 783 patients with ECG were evaluated. The first ECG was used for developing and testing the model. Patients’ mean age was 58.6±16.2, and 52% were male. In the development set, 399 750 unique patients’ ECGs were in the training set, and 99 977 in the internal validation set. The remaining 275 056 ECGs were in the holdout testing set.

All the ECGs in the internal validation set and in the holdout data set were tested using the network. The algorithm output for each ECG included a probability number (P) between 0 and 1. A lower number implied high probability of being female while a number closer to 1 implied high probability of being male. The AUC of the holdout data set was 0.968 versus an AUC of 0.973 on the internal validation data set (Figure 2A). Using the Youden index calculated on the internal validation set (Y= 0.517) the overall accuracy was 90.4% (95% CI, [90.3%–90.5%]) 90.6% (95% CI, [90.4%–90.7%]) for females and 90.3% (95% CI, [90.2%–90.5%]) for males.

After rejecting the 10% ECGs with intermediate P values (0.31<P<0.69) which suggested an inability to differentiate male versus female for the given ECG (seen in about 10% of all ECGs), the accuracy of the classification improved to 94.3 %. When applying the same technique wherein P values between 0.17 and 0.83 were excluded (accounting for 20% of all ECGs), accuracy improved even further to 96.8%. These findings suggest the classifier certainty can be derived from the output probability.

When testing the classifier in a younger population (<45 years of age), the overall accuracy for sex identification was 93% versus 89% in patients older than 55. In addition, a separate independent cohort of 35 000 individuals was used to test whether specific ECG features (P wave area in lead I, QRS duration, P wave time to peak in lead I, QTc interval, and T wave area in lead V4) could be used separately to determine sex. The receiver operating characteristic curves are summarized in Figure 2B for each portion of the P-QRS-T complex, with QRS duration carrying the highest AUC (0.683) but none individually exceeding 0.7.

As the output of the age estimation network was a continuous variable, the statistics of the absolute error were calculated together with the overall correlation and the explained variance (R squared). For the holdout data set, the mean absolute error was 6.9±5.6 years and R squared of 0.7 (r=0.837, P<0.0001). A scatter plot with chronologic age versus CNN-predicted age is presented in Figure 3A. For detection of age ≥40, the AUC was 0.94 with a sensitivity of 87.8%, specificity of 86.8%, and accuracy of 87%. For the multi-group classification to the age groups of 18 to 25, 25 to 50, 50 to 75, and 75 and above the overall accuracy was 71.6% (95% CI [71.5%–71.9%]; Figure 3B).

When supplying the network the results of the sex network and the reported sex of the patient, for the 10% of patient with an inaccurate sex estimation, the R squared for age estimation decreased to 0.62 (r=0.78, P<0.00001) even though the networks were trained separately.

For the 100 patients with longitudinal data, Figure 4 shows the numbers of patients with specific R-squared values between CNN-predicted age and chronologic age over the course of a minimum 20 year follow-up time span. The plurality of patients (N=24) had an R-squared value of 0.8 to 0.9 with 4 patients having an R-squared 0.9 to 1.0. Among patients with an R-squared value of 0.9 to 1.0, all had no significant medical comorbidities and were otherwise healthy, obtaining ECGs only for general physical examinations, executive physical examinations, or for routine noncardiac orthopedic or minor surgical procedures (Figure 5A presents one patient example of this). These patients all had minimal difference between ECG-estimated age and chronological age (±7 years). Several patients (N=17) had R-squared values of ≤0.3. Among these patients, multiple incident events including new-onset low ejection fraction, incident myocardial infarction, new-onset hypertension and diabetes mellitus, cardiac transplantation, and major events such as hemodynamic shock occurred over the follow-up period, resulting in major deviations from chronologic age during the year in which the events occurred, (Figure 5B presents one patient example of this). In most of these patients, with stabilization from the incident clinical insult, CNN-predicted age again approximated chronologic age. (Figure 5B)

In patients with CNN-predicted age having a deviation from chronologic age of more than 7 years, there was a higher incidence of preexisting comorbidities including prior diagnoses of myocardial infarction, low ejection fraction, coronary disease, hypertension, and atrial fibrillation. (P<0.01; Table). In one patient of interest, CNN-predicted age exceeded chronologic age for most of the follow-up period, but after subsequent cardiac transplant, CNN-predicted age decreased but was still greater than the age of the donor. After a period of weight loss with improvement in other comorbidities including hypertension and diabetes mellitus, the patient’s CNN-predicted age further decreased closer to the donor age and, in fact, to less than the patient’s chronologic age. (Figure 5C)

In a chart review of a random selection of patients, several clinical factors associated with mortality including myocardial infarction, low ejection fraction, and coronary artery disease, appeared to correlate with deviations between CNN-predicted age and chronologic age. This may account for some differences between CNN-predicted and chronologic age, and supports the hypothesis that CNN-predicted age might prove a useful marker for physiological age. The higher incidence of these diseases with older age might also explain the corresponding differences in sex estimation as sex identification was more accurate in younger than older patients.

In fact, the lower accuracy of sex in older individuals may be hypothesized to be due to changes in hormone levels that may result in greater difficulty for the neural network to differentiate men from women. It is well recognized that circulating sex hormones may impact ECG characteristics.^16–18 However, we could not account for hormonal status or gender in our population (only reported sex as male or female). Thus, there may have been factors in individual patients (eg, hormone therapy, gender reassignment, incident diseases, etc) that may have resulted in inaccuracy of sex or age assignment by the neural network. Further work is needed to better understand the nature of these outliers.

We have demonstrated that by applying advanced AI techniques to the ECG, it is feasible to accurately extract sex and age in most individuals. Furthermore, the system is able to provide the user with a certainty value about its categorization of an individual to a specific sex. The ability of such a system to provide a certainty value is critical in future applications to make it less of a black box. Limitations to such AI approaches include current inability to simply explain the features of the ECG that result in a given result (such as assignment of a specific age or sex). Figure 6A through 6D show example ECGs from patients with their actual age/sex, their CNN-predicted age and sex, and their comorbid conditions (if any). The reasons for the limitation of explainability includes that the AI is not relying on a singular measure or a simple scoring metric, but rather a complex set of factors that include varying levels of nonlinear interaction.

This issue of neural network explainability faces the additional issue of understanding the clinical features driving the decision making of the algorithm. In this study, we did not seek to specifically compare the benefits of a CNN approach for predicting age and sex from the ECG to other approaches to machine learning, such as principal component analysis, Bayesian approaches, or other high-level statistical methodologies. The neural network approach is subject to inherent limitations, including the reliance on a big data approach, which is characterized by its own flaws related to data verification, the as-of-yet inability to oversee the system’s reasoning given the limitations in achieving immediate explainability, and limited research on how best to apply it to medicine. However, neural networks may offer a higher level of accuracy than other statistical methods as it may operate independently of the biases of the investigator, only dependent on the data that is input. The reason for this is that a neural network learns through repeated exposure over time to associations between inputs. Of course, one key limitation is when considering the nature of the data inputted and resultant generalizability to real-world cohorts. For example, all of our cases were, effectively, patients, whether for general wellness evaluations or for evaluation of active cardiac conditions. Our population (both those used to train the algorithm and those in whom a deeper dive was performed) was randomly selected, but we cannot specifically speak to the percentage that had total absence of any cardiovascular disease. Thus, future work on (1) whether algorithms derived from such a cohort are relevant when studied in an ostensibly healthy population and (2) whether other demographic differences may exist in our population (eg, ethnicity, age distribution, etc) that further impact generalizability will need to be done.

One key limitation to these results is that all individuals included were patients, and thus an ECG was obtained for some clinical indication. Whether these results are similarly accurate among an ostensibly healthy population is unknown and revalidation in such a cohort will be critical. The population as a whole was randomly selected and thus not precharacterized according to presence or absence of specific diseases or other factors thus making assumptions regarding their disease status or presence or absence of comorbidities difficult to assume since they were not rigorously characterized for this specific purpose. Furthermore, we are unable to indicate which features of the ECG contributed to the final network output and how the final output was reached, due to the very nature of neural networks. Finally, we could not consider other clinical characteristics (eg, we could only consider patients’ self-reported sex) in the data set and thus understanding why certain outliers were in fact outliers is not possible with the current data set. This does reflect one critical limitation of our study as we could only consider sex as reported in the ECG system, where male/female is based on self-reported sex of the individual, and further details were not readily extractable from the health record without manual abstraction.

Applying AI to the ECG allows robust detection of patient sex and estimation of age. The ability of an AI algorithm to determine physiological age, with further validation, has the potential to serve as a rapid measure of overall health status, though will require further validation in well-characterized, prospectively followed cohorts.