香京julia种子在线播放

Cardiovascular disease is underdiagnosed in women compared to that in men; the lack of sex-specific diagnostic criteria contributes to this issue (St. Pierre et al., 2022). With the currently used non-sex-specific criteria, the prevalence of dilated and hypertrophic cardiomyopathy is 3:1 and 3:2 for men-to-women, respectively, indicating that men are diagnosed more frequently than women for these cardiomyopathies (Olivotto et al., 2005; Cannatà et al., 2020). On average, women have a smaller wall thickness than men. For hypertrophic cardiomyopathy, the lack of sex-specific criteria implies that female hearts have to disproportionally increase more in thickness than male hearts to reach the diagnostic threshold of a wall thickness of 15 mm (van Driel et al., 2019). Strikingly, women are half as likely to be diagnosed during a routine examination for hypertrophic cardiomyopathy compared to men (Olivotto et al., 2005). Women are also diagnosed at an older age and with more symptoms than men for hypertrophic and dilated cardiomyopathies (Olivotto et al., 2005; Halliday et al., 2018; Cannatà et al., 2020).

The need for sex-specific diagnostic criteria is also visible in heart failure with the preserved left ventricle ejection fraction where the cut-off is an ejection fraction of ≥50% (Ponikowski et al., 2016). However, women have a higher baseline ejection fraction than men on average (Rutkowski et al., 2020). Studies have already shown that women benefit from therapies at a higher range of ejection fractions than men (McMurray et al., 2020; Solomon and McMurray, 2021). Clearly, there is an urgent need for sex-specific research to understand the different impacts of heart failure on men and women (Peirlinck et al., 2021b; Lala et al., 2022; Yang et al., 2023).

Clinically used risk prediction models for cardiovascular diseases typically include components of the Framingham Risk Score: age, sex, total cholesterol, high-density lipoprotein (HDL) cholesterol, systolic blood pressure, blood pressure treated through medicine, diabetes status, and smoking status (Wilson et al., 1998). The body mass index (BMI) is also common to include in risk models (Alaa et al., 2019). These risk models are easy to use; they only require a handful of easy-to-measure variables, and risk evaluation is a simple score based on discrete thresholds for each of these variables. Clinicians use these risk models to determine if an otherwise asymptomatic person would benefit from medical intervention (Alaa et al., 2019; Kremers et al., 2008; D’Agostino et al., 2008; Sjöström et al., 2004).

Tabular data consist of features that can be input into a spreadsheet, including continuous variables, like the age and binary variables and the smoking status, which are coded with zero for negative and one for positive. The state-of-the-art approaches for supervised learning on tabular data are gradient-boosted tree ensembles (Borisov et al., 2022), which are conventional machine learning methods. The top gradient-boosted models based on benchmark performance for five independent datasets (Borisov et al., 2022) are XGBoost (eXtreme Gradient Boosting) (Chen and Guestrin, 2016), LightGBM (Ke et al., 2017), and CatBoost (Prokhorenkova et al., 2018). The strengths of tree ensemble methods include robustness against outliers and noisy data (Aceña et al., 2022) and the fast exact extraction of feature importance via methods such as SHapley Additive exPlanations (SHAP) (Lundberg and Lee, 2017). A weakness of decision trees is that they are unstable and tend to overfit the training data (Aceña et al., 2022).

Although gradient boosting frameworks have been shown to be the best-performing approaches for tabular data in the past decade (Shwartz-Ziv and Armon, 2022), deep learning architectures are becoming increasingly prevalent (Gorishniy et al., 2021), sometimes outperforming the state-of-the-art approaches (Somepalli et al., 2021; Borisov et al., 2022). In particular, deep learning frameworks, such as TabTransformer (Huang et al., 2020), DeepFM (Guo et al., 2017), TabNet (Arik and Pfister, 2020), and SAINT (Self-Attention and Intersample Attention Transformer) (Somepalli et al., 2021), showed significant promise for effective tabular data modeling, with the SAINT model outperforming the gradient-boosted frameworks on some learning tasks using the power of representation learning (Somepalli et al., 2021; Borisov et al., 2022). Nonetheless, a well-known downside of deep learning models, as compared to tree-based frameworks, is that they are generally much slower to train (Grinsztajn et al., 2022).

Machine learning models can effectively utilize a large number of input features, which allows them to discover new, more accurate risk models to better identify people at risk (Madani et al., 2018; Alaa et al., 2019; Alber et al., 2019). XGBoost has been applied extensively for cardiovascular disease diagnoses (Rajliwall et al., 2018; Alaa et al., 2019; Athanasiou et al., 2020; Rajadevi et al., 2021; Papadopoulou et al., 2022). Prior statistical or deep learning models of cardiovascular diseases focused on the lifestyle factors (Sjöström et al., 2004; Alaa et al., 2019; Papadopoulou et al., 2022; Sharma et al., 2022), medical history (Alaa et al., 2019; Papadopoulou et al., 2022), sociodemographics (Alaa et al., 2019; Papadopoulou et al., 2022), dietary and nutritional information (Alaa et al., 2019; Papadopoulou et al., 2022), genetics (Papadopoulou et al., 2022; Sharma et al., 2022), and/or one of the four clinical tests: pulse wave analysis (Davies and Struthers, 2005; Said et al., 2018), electrocardiograms (Attia et al., 2019; Ramírez et al., 2021; Papadopoulou et al., 2022), carotid ultrasounds (Zhao et al., 2016; Sharma et al., 2022), or magnetic resonance imaging (Chung et al., 2006; Gilbert et al., 2019; Bai et al., 2020), but not all four.

First, we investigate whether women are underdiagnosed for cardiovascular diseases in the UK Biobank cohort. Then, we compare the performance of three models, a multilayer perceptron deep learning baseline, XGBoost, and the novel deep learning framework, SAINT, on their ability to predict whether a person can be diagnosed with a cardiovascular disease. Lastly, we identify the top sex- and disease-specific risk factors from four cardiovascular-related tests, pulse wave analysis, electrocardiograms, magnetic resonance imaging, and carotid ultrasounds, against the traditional Framingham Risk Score.

The UK Biobank comprises data from half a million individuals from the UK who were over the age of 40 (Sudlow et al., 2015). From these, we selected individuals who underwent ECG testing, magnetic resonance imaging, carotid ultrasounds, and pulse wave analysis, resulting in a population of 20,542 individuals. We also pulled features associated with the Framingham Risk Score, sex, age, total cholesterol, HDL cholesterol, smoking status, diabetes status, end-systolic blood pressure, the body mass index of all participants, and their medical diagnoses. Table 1 shows the demographic data on this population. We did not include treatment for blood pressure as a feature in our models as this directly reflects one of the diagnostic outcomes, hypertension, that we are trying to predict. All 57 features are shown in Table 2.

We created two feature groups: 1) eight features, including Framingham Risk Score features and the body mass index only, and 2) all 57 features. We labeled each person to be in the positive class if they were diagnosed with a given cardiovascular disease (Said et al., 2018). We split the datasets based on four disease categories, as shown in Table 3, namely, any disease, hypertension (ICD-10 codes I10–I15), ischemic (I20–I25), and conduction disorders (I44–I49) (World Health Organization, 2004), such that the detection of each disease can pose as a binary classification. We chose these disease categories because they had the largest number of participants who both had these diagnoses and data from all four imaging studies, ECG, heart MRI, pulse wave analysis, and carotid ultrasounds. To train the sex-specific classifiers, we further designated three input groups, i.e., both sexes, female only, and male only.

Using the three input groups (both sexes, female only, and male only) together with the four binary label sets (any disease, hypertensive disease, ischemic heart disease, and conduction disorders), we constructed 12 dataset variants to train the binary classifiers, as shown in Figure 1. We generated each of the datasets by the direct slicing of the randomly pre-shuffled data frame. Since the datasets are relatively small, we applied a 70–15–15 split to create the training, validation, and test sets for each of the variants. The positive class in all 12 dataset variants is significantly underrepresented relative to the negative class, so we applied oversampling to approximately equalize the number of negative and positive samples in the training sets of the corresponding 12 datasets.

Using the cardiovascular and Framingham Risk Score features, we implemented three distinct model types: 1) a multilayer perceptron (MLP); 2) an XGBoost ensemble model, which is a state-of-the art approach for tabular data learning (Chen and Guestrin, 2016); and 3) the SAINT model (Somepalli et al., 2021). We used the MLP as a baseline for deep learning performance and the XGBoost as a baseline for a state-of-the-art performance. For each model type, we trained and evaluated 12 individual classifiers, according to our 12 dataset variants. For evaluation purposes, we consider both untuned and tuned XGBoost ensemble models and introduce an additional set of tuned XGBoost ensembles trained only on the Framingham Risk Score features. The 12 cardiovascular disease datasets, three model types, and two additional XGBoost variants result in a total of 60 individually trainable classifiers.

An underdiagnosis is calculated as the ratio of the number of people who, at a single time point, met the criteria for a given disease diagnosis but were never diagnosed for that disease across the entire timespan of the medical record divided by the number of people who had been diagnosed for the same disease at any point in time across the entire span of their medical records. The UK Biobank ICD-10 medical records start between 1981 and 1988, depending on the country in the UK, and last until 2022. Only 4% of the UK Biobank population has ICD-9 records, so we have not included these. The blood pressure was measured between 2006 and 2010, and data from the first time point of imaging visits, ECG, MRI, carotid ultrasound, and pulse wave analysis were collected after 2014. The numerator represents the minimum possible value; it implies that at a single time point between 1981 and 2022, people met the disease criteria but never got a diagnosis. It is a minimum because there are likely people from the healthy category who, at any other time point, would have met the criteria but were not measured at that precise time and were also never diagnosed. The denominator represents the maximum possible value; it reflects whether people have ever been diagnosed across the entire timeline from 1981 to 2022. So, this metric of underdiagnosis is in itself an underestimation.

We first investigate whether women are underdiagnosed relative to men for cardiovascular diseases in the UK Biobank cohort. We chose diseases where the diagnosis is a simple, non-sex-specific cut-off.

Figure 2 shows the cut-off criteria in red. Plots (a–d) show individuals who have not been diagnosed with the disease in orange and those who have in purple. Plots (e,f) are divided into four categories. The truncated violin plots show the distribution of each sex for each category with the box plots showing the mean in white and the 25th–75th percentiles. Each dot represents a single person in the Biobank dataset. There may be comorbidities or alternate medical diagnoses that result in similar presentations, so the magnitude of an underdiagnosis, in the following examples, should be understood as a first approximation.

Essential primary hypertension is diagnosed by a systolic blood pressure of ≥140 mmHg and/or diastolic blood pressure of ≥90 mmHg (Williams et al., 2018). The corresponding ICD-10 code is I10. Figures 2A,B show that women have, on average, a lower systolic and diastolic blood pressure than men. In the Biobank cohort, 35.2% of men are diagnosed, while 52.3% of men meet the cut-off criteria. For women, 26.6% are diagnosed, while 40.2% meet the cut-off criteria. This means that women and men are underdiagnosed for essential primary hypertension at the same rate, 1.5×, when non-sex-specific criteria are used.

Hypertrophic cardiomyopathy is diagnosed with a wall thickness of > 15 mm (Elliott et al., 2014); the ICD-10 codes are I42.1 and I42.2. Figure 2C shows that none of the approximately 900 people with cardiac magnetic resonance images met the criteria for hypertrophic cardiomyopathy or were diagnosed. Women, on average, have a distinctly smaller wall thickness than men.

The first-degree AV block is diagnosed with a PQ interval of > 200 ms (Holmqvist and Daubert, 2013); the ICD-10 code is I44.0. As shown in Figure 2D, women have a smaller PQ interval than men on average. The first-degree AV block is generally asymptomatic but is no longer considered entirely benign, with nearly double the risk of developing atrial fibrillation and triple the risk of needing a pacemaker (Holmqvist and Daubert, 2013). As such, the current recommendation is to monitor patients regularly to see if the conduction delay continues to widen or if they are developing atrial fibrillation (Oldroyd et al., 2022). In the UK Biobank, 0.81% of men are diagnosed and 12.6% of men meet the cut-off. For women, 0.18% of them are diagnosed, while 5.4% meet the cut-off. So, women are underdiagnosed 30×, while men are underdiagnosed 15.6×, meaning that women are nearly 2× more underdiagnosed relative to men for a first-degree AV block with the given non-sex-specific criteria.

Dilated cardiomyopathy is diagnosed by a left ventricle ejection fraction of < 45% and a left ventricle end-diastolic diameter of > 112% of the predicted diameter based on the age and sex (Orphanou et al., 2022). Left ventricle fractional shortening less than 25% can be used in place of the ejection fraction criteria, but these data were not available in the Biobank. Because the left ventricle end-diastolic volume is reported, we used the Teichholz formula (Arora et al., 2010), LVEDV = 7 ( LVEDD cal ) 3 / ( 2.4 + LVEDD cal ) , to calculate the end-diastolic diameter from the volume and the formula, LVEDD_pre = 45.3(BSA)^0.3–0.03 (age) − 7.2, to predict the end-diastolic diameter from the BSA and age. If LVEDD_cal/LVEDD_pre >1.12, the individual would meet the criteria and either be assigned a red or purple dot, as shown in Figure 2E,F, depending on whether they had also been diagnosed with dilated cardiomyopathy or not, respectively. If they did not meet this criterion, they were assigned an orange or blue dot, where blue indicates that they had been diagnosed and orange indicating that they had not been. In the Biobank cohort, 55 people were diagnosed with dilated cardiomyopathy, but only 35 met the cut-off using these calculations, with nearly all of these discrepancies for not meeting the ejection fraction criteria, as shown by the blue dot. Women, on average, have a slightly lower end-diastolic diameter and higher ejection fraction than men. Out of the men in the cohort, 0.23% of them were diagnosed, while 5.71% met the cut-off criterion. For women, 0.06% of them were diagnosed, while 2.02% met the cut-off criterion. As such, women are underdiagnosed 33.7×, men are underdiagnosed 24.8×, and women are 1.4× more underdiagnosed than men when non-sex-specific criteria are used.

Figure 3 shows the ROC and AUC values for the five model types, MLP, untuned XGBoost, tuned XGBoost, SAINT, and XGBoost, with only Framingham Risk Score features across the 12 sex and disease categories. A large AUC score is designed to minimize false negatives (predicted healthy but actually diseased) and maximize true positives (predicted diseased and actually diseased). Table 4 summarizes the performance metrics for all classifiers, where we report the test set accuracy, precision, and recall in addition to the AUC score. In terms of the AUC metric, the SAINT model performed best on all datasets, except the female-only conduction disorder datasets, where only the corresponding tuned XGBoost model performed better. For accuracy and precision, XGBoost (tuned) models were the best performing in 11/12 cases and 8/12 cases, respectively. The SAINT had the best-performing recall in 9/12 cases.

Although all 60 models were measurably better than a random classifier, none of the models demonstrated a high AUC score. Since we did not observe training set overfitting, this might be indicative of a high Bayes error rate and low feature-output correlations in the datasets. The ischemic disease and conduction disorder models performed rather poorly, most likely caused by the small training set sizes and the increasingly significant class imbalance in their test sets.

Figure 3 suggests that including features from ECG, magnetic resonance imaging, pulse wave analysis, and carotid ultrasound, along with Framingham Risk Score features significantly increased the AUC score of the corresponding 48 models, as compared to the AUC scores for the Framingham-only XGBoost models. The XGBoost classifiers trained and tuned on the Framingham-only features were always the lowest or second-lowest performing models for a given dataset.

Figure 4 shows the performance of the 12 individually trained XGBoost classifiers on individual sexes. First, the classifiers trained on both sexes perform the best for all female-only datasets, top row. The best AUC values for the female-only data are lower than the best AUC values for the male-only data for all disease categories, except conduction disorders. Second, the male-only classifiers perform the best for the male-only datasets for any disease and hypertension categories, while the both-sex classifiers perform the best for ischemic and conduction diseases. Third, the performance of all classifiers is fairly similar for most sex- and disease-specific categories, except for three cases: 1) the female-only classifier is significantly worse at predicting male cases of ischemic diseases, 2) the male-only classifier is worse at predicting female cases of ischemic diseases, and 3) at predicting conduction diseases as compared to the female-only and both-sex classifiers.

Figure 5 shows the 10 most predictive features for any type of cardiovascular disease for both sexes combined, women only, and men only. A more positive SHAP value indicates a larger contribution to the positive class, diagnosed with a cardiovascular disease, while a negative SHAP value indicates the opposite. Each dot represents an individual person in the dataset, while the red color means that a person had a high value of that feature, e.g., older, while blue means a lower value, e.g., younger. For the binary categories of sex, smoking status, and diabetes status, red represents a male subject, a person who smokes, and a person with diabetes, respectively. Traditional risk factors refer to the Framingham Risk Score features plus body mass index, as shown in Table 2.

For both sexes combined, six of the top 10 features, namely, age, body mass index, cholesterol, HDL cholesterol, sex, and blood pressure, are the factors that are traditionally associated with an increased cardiovascular risk. The other four features are ECG features. For the female-only dataset, five of the top 10 features are traditional risk factors but the rest of the features include a mix of ECG, pulse wave analysis, and carotid ultrasound features. For the male-only dataset, six of the traditional risk factors make up the top 10 features. The other four features are ECG features. Interestingly, the male-only dataset is the only place where the smoking status and the diabetes status make up the top 10 features. Age, body mass index, and cholesterol, either HDL or total, are consistently the top three features, regardless of sex. The ECG features of the PQ interval and T-axis appear in all three categories as well.

Table 5 shows the top 10 features based on the SHAP value for the tuned XGBoost model prediction of cardiovascular disease. Across all groups, age is the most important feature in predicting risk. When trained on both sexes, the body mass index and HDL cholesterol also appear for all disease groups. Sex is the most important for ischemic heart disease, but interestingly, it is not in the top 10 for conduction disorders. Out of the traditional risk factors in the Framingham Risk Score, the diabetes status appears only once for the prediction of hypertension and the smoking status does not appear at all. A measure of blood pressure also only appears for any disease and hypertension. The ECG, pulse wave analysis, and magnetic resonance imaging features of the PQ interval, T-axis, pulse rate, R-axis, QRS duration, and LV ejection fraction also appear in two disease categories, each in the top 10.