MSGene: a multistate model using genetic risk and the electronic health record applied to lifetime risk of coronary artery disease.

Coronary artery disease (CAD) is the leading cause of death among adults worldwide. Accurate risk stratification can support optimal lifetime prevention. Current methods lack the ability to incorporate new information throughout the life course or to combine innate genetic risk factors with acquired lifetime risk. We designed a general multistate model (MSGene) to estimate age-specific transitions across 10 cardiometabolic states, dependent on clinical covariates and a CAD polygenic risk score. This model is designed to handle longitudinal data over the lifetime to address this unmet need and support clinical decision-making. We analyze longitudinal data from 480,638 UK Biobank participants and compared predicted lifetime risk with the 30-year Framingham risk score. MSGene improves discrimination (C-index 0.71 vs 0.66), age of high-risk detection (C-index 0.73 vs 0.52), and overall prediction (RMSE 1.1% vs 10.9%), in held-out data. We also use MSGene to refine estimates of lifetime absolute risk reduction from statin initiation. Our findings underscore our multistate model's potential public health value for accurate lifetime CAD risk estimation using clinical factors and increasingly available genetics toward earlier more effective prevention.

Fetuin-A and its genetic association with cardiometabolic disease.

Fetuin-A acts as both an inhibitor of calcification and insulin signaling. Previous studies reported conflicting results on the association between fetuin-A and cardiometabolic diseases. We aim to provide further insights into the association between genetically predicted levels of fetuin-A and cardiometabolic diseases using a Mendelian randomization strategy. Genetic variants associated with fetuin-A and their effect sizes were obtained from previous genetic studies. A series of two-sample Mendelian randomization analyses in 412,444 unrelated individuals from the UK Biobank did not show evidence for an association of genetically predicted fetuin-A with any stroke, ischemic stroke, or myocardial infarction. We do find that increased levels of genetically predicted fetuin-A are associated with increased risk of type 2 diabetes (OR = 1.21, 95%CI 1.13-1.30, P = < 0.01). Furthermore, genetically predicted fetuin-A increases the risk of coronary artery disease in individuals with type 2 diabetes, but we did not find evidence for an association between genetically predicted fetuin-A and coronary artery disease in those without type 2 diabetes (P for interaction = 0.03). One SD increase in genetically predicted fetuin-A decreases risk of myocardial infarction in women, but we do not find evidence for an association between genetically predicted fetuin-A and myocardial infarction in men (P for interaction = < 0.01). Genetically predicted fetuin-A is associated with type 2 diabetes. Furthermore, type 2 diabetes status modifies the association of genetically predicted fetuin-A with coronary artery disease, indicating that fetuin-A increases risk in individuals with type 2 diabetes. Finally, higher genetically predicted fetuin-A reduces the risk of myocardial infarction in women, but we do not find evidence for an association between genetically predicted fetuin-A and myocardial infarction in men.

MSGene: Derivation and validation of a multistate model for lifetime risk of coronary artery disease using genetic risk and the electronic health record.

Currently, coronary artery disease (CAD) is the leading cause of death among adults worldwide. Accurate risk stratification can support optimal lifetime prevention. We designed a novel and general multistate model (MSGene) to estimate age-specific transitions across 10 cardiometabolic states, dependent on clinical covariates and a CAD polygenic risk score. MSGene supports decision making about CAD prevention related to any of these states. We analyzed longitudinal data from 480,638 UK Biobank participants and compared predicted lifetime risk with the 30-year Framingham risk score. MSGene improved discrimination (C-index 0.71 vs 0.66), age of high-risk detection (C-index 0.73 vs 0.52), and overall prediction (RMSE 1.1% vs 10.9%), with external validation. We also used MSGene to refine estimates of lifetime absolute risk reduction from statin initiation. Our findings underscore the potential public health value of our novel multistate model for accurate lifetime CAD risk estimation using clinical factors and increasingly available genetics.

Selecting cardiac magnetic resonance images suitable for annotation of pulmonary arteries using an active-learning based deep learning model.

An increasing and aging patient population poses a growing burden on healthcare professionals. Automation of medical imaging diagnostics holds promise for enhancing patient care and reducing manpower required to accommodate an increasing patient-population. Deep learning, a subset of machine learning, has the potential to facilitate automated diagnostics, but commonly requires large-scaled labeled datasets. In medical domains, data is often abundant but labeling is a laborious and costly task. Active learning provides a method to optimize the selection of unlabeled samples that are most suitable for improvement of the model and incorporate them into the model training process. This approach proves beneficial when only a small number of labeled samples are available. Various selection methods currently exist, but most of them employ fixed querying schedules. There is limited research on how the timing of a query can impact performance in relation to the number of queried samples. This paper proposes a novel approach called dynamic querying, which aims to optimize the timing of queries to enhance model development while utilizing as few labeled images as possible. The performance of the proposed model is compared to a model trained utilizing a fully-supervised training method, and its effectiveness is assessed based on dataset size requirements and loss rates. Dynamic querying demonstrates a considerably faster learning curve in relation to the number of labeled samples used, achieving an accuracy of 70% using only 24 samples, compared to 82% for a fully-supervised model trained on the complete training dataset of 1017 images.

Hybridizing machine learning in survival analysis of cardiac PET/CT imaging.

BACKGROUND: Machine Learning (ML) allows integration of the numerous variables delivered by cardiac PET/CT, while traditional survival analysis can provide explainable prognostic estimates from a restricted number of input variables. We implemented a hybrid ML-and-survival analysis of multimodal PET/CT data to identify patients who developed myocardial infarction (MI) or death in long-term follow up. METHODS: Data from 739 intermediate risk patients who underwent coronary CT and selectively stress 15O-water-PET perfusion were analyzed for the occurrence of MI and all-cause mortality. Images were evaluated segmentally for atherosclerosis and absolute myocardial perfusion through 75 variables that were integrated through ML into an ML-CCTA and an ML-PET score. These scores were then modeled along with clinical variables through Cox regression. This hybridized model was compared against an expert interpretation-based and a calcium score-based model. RESULTS: Compared with expert- and calcium score-based models, the hybridized ML-survival model showed the highest performance (CI .81 vs .71 and .64). The strongest predictor for outcomes was the ML-CCTA score. CONCLUSION: Prognostic modeling of PET/CT data for the long-term occurrence of adverse events may be improved through ML imaging score integration and subsequent traditional survival analysis with clinical variables. This hybridization of methods offers an alternative to traditional survival modeling of conventional expert image scoring and interpretation.

ukbpheno v1.0: An R package for phenotyping health-related outcomes in the UK Biobank.

The complexity and volume of data associated with population-based cohorts means that generating health-related outcomes can be challenging. Using one such cohort, the UK Biobank-a major open access resource-we present a protocol to efficiently integrate the main dataset and record-level data files, to harmonize and process the data using an R package named "ukbpheno". We describe how to use the package to generate binary phenotypes in a standardized and machine-actionable manner. For complete details on the use and execution of this protocol, please refer to Yeung et al. (2022).

Multi-task Deep Learning of Myocardial Blood Flow and Cardiovascular Risk Traits from PET Myocardial Perfusion Imaging.

BACKGROUND: Advanced cardiac imaging with positron emission tomography (PET) is a powerful tool for the evaluation of known or suspected cardiovascular disease. Deep learning (DL) offers the possibility to abstract highly complex patterns to optimize classification and prediction tasks. METHODS AND RESULTS: We utilized DL models with a multi-task learning approach to identify an impaired myocardial flow reserve (MFR <2.0 ml/g/min) as well as to classify cardiovascular risk traits (factors), namely sex, diabetes, arterial hypertension, dyslipidemia and smoking at the individual-patient level from PET myocardial perfusion polar maps using transfer learning. Performance was assessed on a hold-out test set through the area under receiver operating curve (AUC). DL achieved the highest AUC of 0.94 [0.87-0.98] in classifying an impaired MFR in reserve perfusion polar maps. Fine-tuned DL for the classification of cardiovascular risk factors yielded the highest performance in the identification of sex from stress polar maps (AUC = 0.81 [0.73, 0.88]). Identification of smoking achieved an AUC = 0.71 [0.58, 0.85] from the analysis of rest polar maps. The identification of dyslipidemia and arterial hypertension showed poor performance and was not statistically significant. CONCLUSION: Multi-task DL for the evaluation of quantitative PET myocardial perfusion polar maps is able to identify an impaired MFR as well as cardiovascular risk traits such as sex, smoking and possibly diabetes at the individual-patient level.

Artificial Intelligence to Improve Risk Prediction with Nuclear Cardiac Studies.

PURPOSE OF REVIEW: As machine learning-based artificial intelligence (AI) continues to revolutionize the way in which we analyze data, the field of nuclear cardiology provides fertile ground for the implementation of these complex analytics. This review summarizes and discusses the principles regarding nuclear cardiology techniques and AI, and the current evidence regarding its performance and contribution to the improvement of risk prediction in cardiovascular disease. There is a growing body of evidence on the experimentation with and implementation of machine learning-based AI on nuclear cardiology studies both concerning SPECT and PET technology for the improvement of risk-of-disease (classification of disease) and risk-of-events (prediction of adverse events) estimations. These publications still report objective divergence in methods either utilizing statistical machine learning approaches or deep learning with varying architectures, dataset sizes, and performance. Recent efforts have been placed into bringing standardization and quality to the experimentation and application of machine learning-based AI in cardiovascular imaging to generate standards in data harmonization and analysis through AI. Machine learning-based AI offers the possibility to improve risk evaluation in cardiovascular disease through its implementation on cardiac nuclear studies. AI in improving risk evaluation in nuclear cardiology. * Based on the 2019 ESC guidelines.

Genomic insights in ascending aortic size and distensibility.

BACKGROUND: Alterations in the anatomic and biomechanical properties of the ascending aorta (AAo) can give rise to various vascular pathologies. The aim of the current study is to gain additional insights in the biology of the AAo size and function. METHODS: We developed an AI based analysis pipeline for the segmentation of the AAo, and the extraction of AAO parameters. We then performed genome-wide association studies of AAo maximum area, AAo minimum area and AAo distensibility in up to 37,910 individuals from the UK Biobank. Variants that were significantly associated with AAo phenotypes were used as instrumental variables in Mendelian randomization analyses to investigate potential causal relationships with coronary artery disease, myocardial infarction, stroke and aneurysms. FINDINGS: Genome-wide association studies revealed a total of 107 SNPs in 78 loci. We annotated 101 candidate genes involved in various biological processes, including connective tissue development (THSD4 and COL6A3). Mendelian randomization analyses showed a causal association with aneurysm development, but not with other vascular diseases. INTERPRETATION: We identified 78 loci that provide insights into mechanisms underlying AAo size and function in the general population and provide genetic evidence for their role in aortic aneurysm development.

Twenty-Five Novel Loci for Carotid Intima-Media Thickness: A Genome-Wide Association Study in >45 000 Individuals and Meta-Analysis of >100 000 Individuals.

OBJECTIVE: Carotid artery intima-media thickness (cIMT) is a widely accepted marker of subclinical atherosclerosis. Twenty susceptibility loci for cIMT were previously identified and the identification of additional susceptibility loci furthers our knowledge on the genetic architecture underlying atherosclerosis. APPROACH AND RESULTS: We performed 3 genome-wide association studies in 45 185 participants from the UK Biobank study who underwent cIMT measurements and had data on minimum, mean, and maximum thickness. We replicated 15 known loci and identified 20 novel loci associated with cIMT at P<5×10-8. Seven novel loci (ZNF385D, ADAMTS9, EDNRA, HAND2, MYOCD, ITCH/EDEM2/MMP24, and MRTFA) were identified in all 3 phenotypes. An additional new locus (LOXL1) was identified in the meta-analysis of the 3 phenotypes. Sex interaction analysis revealed sex differences in 7 loci including a novel locus (SYNE3) in males. Meta-analysis of UK Biobank data with a previous meta-analysis led to identification of three novel loci (APOB, FIP1L1, and LOXL4). Transcriptome-wide association analyses implicated additional genes ARHGAP42, NDRG4, and KANK2. Gene set analysis showed an enrichment in extracellular organization and the PDGF (platelet-derived growth factor) signaling pathway. We found positive genetic correlations of cIMT with coronary artery disease rg=0.21 (P=1.4×10-7), peripheral artery disease rg=0.45 (P=5.3×10-5), and systolic blood pressure rg=0.30 (P=4.0×10-18). A negative genetic correlation between average of maximum cIMT and high-density lipoprotein was found rg=-0.12 (P=7.0×10-4). CONCLUSIONS: Genome-wide association meta-analyses in >100 000 individuals identified 25 novel loci associated with cIMT providing insights into genes and tissue-specific regulatory mechanisms of proatherosclerotic processes. We found evidence for shared biological mechanisms with cardiovascular diseases.

Improving explainability of deep neural network-based electrocardiogram interpretation using variational auto-encoders.

Aims: Deep neural networks (DNNs) perform excellently in interpreting electrocardiograms (ECGs), both for conventional ECG interpretation and for novel applications such as detection of reduced ejection fraction (EF). Despite these promising developments, implementation is hampered by the lack of trustworthy techniques to explain the algorithms to clinicians. Especially, currently employed heatmap-based methods have shown to be inaccurate. Methods and results: We present a novel pipeline consisting of a variational auto-encoder (VAE) to learn the underlying factors of variation of the median beat ECG morphology (the FactorECG), which are subsequently used in common and interpretable prediction models. As the ECG factors can be made explainable by generating and visualizing ECGs on both the model and individual level, the pipeline provides improved explainability over heatmap-based methods. By training on a database with 1.1 million ECGs, the VAE can compress the ECG into 21 generative ECG factors, most of which are associated with physiologically valid underlying processes. Performance of the explainable pipeline was similar to 'black box' DNNs in conventional ECG interpretation [area under the receiver operating curve (AUROC) 0.94 vs. 0.96], detection of reduced EF (AUROC 0.90 vs. 0.91), and prediction of 1-year mortality (AUROC 0.76 vs. 0.75). Contrary to the 'black box' DNNs, our pipeline provided explainability on which morphological ECG changes were important for prediction. Results were confirmed in a population-based external validation dataset. Conclusions: Future studies on DNNs for ECGs should employ pipelines that are explainable to facilitate clinical implementation by gaining confidence in artificial intelligence and making it possible to identify biased models.

Improving patient identification for advanced cardiac imaging through machine learning-integration of clinical and coronary CT angiography data.

BACKGROUND: Standard computed tomography angiography (CTA) outputs a myriad of interrelated variables in the evaluation of suspected coronary artery disease (CAD). But an important proportion of obstructive lesions does not cause significant myocardial ischemia. Nowadays, machine learning (ML) allows integration of numerous variables through complex interdependencies that optimize classification and prediction at the individual level. We evaluated ML performance in integrating CTA and clinical variables to identify patients that demonstrate myocardial ischemia through PET and those who ultimately underwent early revascularization. METHODS AND RESULTS: 830 patients with CTA and selective PET were analyzed. Nine clinical and 58 CTA variables were integrated through ensemble-boosting ML to identify patients with ischemia and those who underwent early revascularization. ML performance was compared against expert CTA interpretation, calcium score and clinical variables. While ML using all CTA variables achieved an AUC = 0.85, it was outperformed by expert CTA interpretation (AUC = 0.87, p < 0.01 for comparison), comparable to ML integration of CTA variables with clinical variables. However, the best performance was achieved by ML integration of expert CTA interpretation and clinical variables for both dependent variables (AUCs = 0.91 and 0.90, p < 0.001). CONCLUSIONS: Machine learning integration of diagnostic CTA and clinical data may improve identification of patients with myocardial ischemia and those requiring early revascularization at the individual level. This could potentially aid in sparing the need for subsequent advanced imaging and better identifying patients in ultimate need for revascularization. While ML integrating all CTA variables did not outperform expert CTA interpretation, ML data integration from different sources consistently improves diagnostic performance.

Genome-Wide Association Study and Identification of a Protective Missense Variant on Lipoprotein(a) Concentration: Protective Missense Variant on Lipoprotein(a) Concentration-Brief Report.

[Figure: see text].

Advanced liver fibrosis but not steatosis is independently associated with albuminuria in Chinese patients with type 2 diabetes.

BACKGROUND & AIMS: Increasing evidence suggests that non-alcoholic fatty liver disease (NAFLD) may be an independent risk factor for chronic kidney disease (CKD). Given the high prevalence of NAFLD among patients with diabetes who are also at risk of CKD, we aimed to investigate the association between NAFLD and albuminuria, a marker commonly found in diabetic nephropathy. METHODS: This study included a cohort of Chinese patients with type 2 diabetes from the Hong Kong Diabetes Registry recruited between March 2013 and May 2014. Liver stiffness measurement (LSM), with probe-specific cut-offs, was used to detect advanced liver fibrosis. While controlled attenuation parameter (CAP) was used to assess liver steatosis using transient elastography. RESULTS: A total of 1,763 Chinese patients with type 2 diabetes were recruited in this analysis. The mean (standard deviation) age and duration of diabetes were 60.7 (11.5) years and 10.8 (8.5) years, respectively. The prevalence of albuminuria was higher in diabetic patients with liver steatosis and those with advanced fibrosis (no NAFLD vs. liver steatosis vs. advanced fibrosis: 41.4% vs. 46.2% vs. 64.2%, p <0.001). After adjustment for potential confounders including glycated hemoglobin, hypertension and body mass index, advanced fibrosis, but not liver steatosis, was associated with increased risk of albuminuria (odds ratio [OR] 1.52; 95% confidence interval [CI] 1.02-2.28; p = 0.039) in patients with eGFR ≥60 ml/min/1.73 m2. The odds of albuminuria increased with greater severity of liver fibrosis in a dose dependent manner, with the highest odds observed in patients with LSM scores ≥11.5 kPa assessed by M probe or ≥11.0 kPa assessed by XL probe (adjusted OR 1.53; 95% CI 1.07-2.20; p = 0.021). CONCLUSIONS: Advanced liver fibrosis, but not steatosis, is independently associated with albuminuria in Chinese patients with type 2 diabetes. Attention should be paid to liver fibrosis in patients with obesity and type 2 diabetes complicated with albuminuria. LAY SUMMARY: In this study, we assessed the link between non-alcoholic fatty liver disease (NAFLD) and albuminuria in a cohort of 1,763 Chinese patients with type 2 diabetes. This study shows that advanced liver fibrosis, a severe form of NAFLD, was independently associated with increased risk of albuminuria. The risk of albuminuria increased with greater severity of liver fibrosis.

Screening diabetic patients for non-alcoholic fatty liver disease with controlled attenuation parameter and liver stiffness measurements: a prospective cohort study.

OBJECTIVE: Type 2 diabetes is an important risk factor for non-alcoholic fatty liver disease (NAFLD), but current guidelines provide conflicting recommendations on whether diabetic patients should be screened for NAFLD. We therefore studied the strategy of screening diabetic patients by FibroScan. DESIGN: Liver fat and fibrosis were assessed by controlled attenuation parameter (CAP) and liver stiffness measurements (LSM) by FibroScan at a diabetic centre for patients from primary care and hospital clinics. Probe-specific LSM cut-offs were used to detect advanced fibrosis. RESULTS: Of 1918 patients examined, 1799 (93.8%) had valid CAP and 1884 (98.2%) had reliable LSM (1770 with the M probe and 114 with the XL probe). The proportion of patients with increased CAP and LSM was 72.8% (95% CI 70.7% to 74.8%) and 17.7% (95% CI 16.0% to 19.5%), respectively. By multivariable analysis, female gender, higher body mass index, triglycerides, fasting plasma glucose and alanine aminotransferase (ALT) and non-insulin use were associated with increased CAP. Longer duration of diabetes, higher body mass index, increased ALT and spot urine albumin:creatinine ratio and lower high-density lipoprotein-cholesterol were associated with increased LSM. Ninety-four patients (80% had increased LSM) underwent liver biopsy: 56% had steatohepatitis and 50% had F3-4 disease. CONCLUSIONS: Diabetic patients have a high prevalence of NAFLD and advanced fibrosis. Those with obesity and dyslipidaemia are at particularly high risk and may be the target for liver assessment. Our data support screening for NAFLD and/or advanced fibrosis in patients with type 2 diabetes.