Jørgen Kim Kanters

Deep learning has achieved expert-level performance in automated electrocardiogram (ECG) diagnosis, yet the "black-box" nature of these models hinders their clinical deployment. Trust in medical AI requires not just high accuracy but also transparency regarding the specific physiological features driving predictions. Existing explainability methods for ECGs typically rely on post-hoc approximations (e.g., Grad-CAM and SHAP), which can be unstable, computationally expensive, and unfaithful to the model's actual decision-making process. In this work, we propose the ECG-IMN, an Interpretable Mesomorphic Neural Network tailored for high-resolution 12-lead ECG classification. Unlike standard classifiers, the ECG-IMN functions as a hypernetwork: a deep convolutional backbone generates the parameters of a strictly linear model specific to each input sample. This architecture enforces intrinsic interpretability, as the decision logic is mathematically transparent and the generated weights (W) serve as exact, high-resolution feature attribution maps. We introduce a transition decoder that effectively maps latent features to sample-wise weights, enabling precise localization of pathological evidence (e.g., ST-elevation, T-wave inversion) in both time and lead dimensions. We evaluate our approach on the PTB-XL dataset for classification tasks, demonstrating that the ECG-IMN achieves competitive predictive performance (AUROC comparable to black-box baselines) while providing faithful, instance-specific explanations. By explicitly decoupling parameter generation from prediction execution, our framework bridges the gap between deep learning capability and clinical trustworthiness, offering a principled path toward "white-box" cardiac diagnostics.

Background and Aims: Atrial cardiomyopathy (AtCM) is increasingly recognized as an important substrate for atrial fibrillation (AF). This study aimed to examine potential markers and risk factors of AtCM, and associations with incident AF, heart failure (HF), and stroke. Methods: Individuals from the UK Biobank with cardiac magnetic resonance imaging and electrocardiographic information were included. Atrial cardiomyopathy markers included left atrial dilation, left atrial mechanical dysfunction, P-wave prolongation, and abnormal P-wave terminal force. Risk factors for AtCM were assessed using logistic regressions. Incident AF, HF, and stroke according to AtCM markers were assessed in multivariable Cox-regression and cumulative incidence models. AF risk according to AtCM markers, clinical and genetic risk factors was evaluated by integrating the HARMS2-AF score and a polygenic risk score for AF. We used net reclassification improvement (NRI) to evaluate reclassification of risk when considering AtCM markers. Results: Among 26 467 individuals, 4145 (15.7%) had ≥1 marker and 619 (2.3%) had ≥2 markers of AtCM. Age, coronary artery disease, and hypertension were consistently associated with AtCM. Having one AtCM marker conferred a hazard ratio (HR) for AF of 1.88 [95% confidence interval (CI): 1.54–2.31; P < .001], with higher rates observed in individuals with ≥2 markers (HR: 4.59; 95% CI: 3.52–5.99; P < .001). Addition of AtCM markers was associated with an NRI of 13.7% (95% CI: 9.2%–18.3%). Integration of clinical and genetic risk factors indicated an additive effect on AF rates. Having ≥2 markers associated with HF (HR: 3.08, 95% CI: 2.03–4.66, P < .001), and stroke (HR: 3.07, 95% CI: 1.78–5.28, P < .001). Conclusions: One in seven individuals had at least one marker of AtCM. Atrial cardiomyopathy markers were associated with AF, HF, and stroke, supporting AtCM as a common substrate for all three outcomes.

Background Synthetic data can be the solution to privacy requirements, can enrich datasets limited by underrepresentation of certain subgroups/minorities, combat data shortage, and reduce annotation costs, to facilitate the development of data-hungry machine-learning applications. An important limitation of current synthetic data is the missing link to patient characteristics and outcomes. This metadata is essential for synthetic data to be useful for solving real-world problems, and we aimed to generate novel synthetic data with clinical characteristics and outcomes. Methods We designed a novel generative method for generating 1:1 personalized synthetic electrocardiograms (ECGs) with associated patient characteristics and outcomes. The architecture of the model is a U-net neural network, which creates patient-specific ECGs to allow attachment of patient characteristics, comorbidities, and outcomes. We developed the model on the General Suburban Population Study and the Lolland-Falster Study cohorts and validated the model on the Danish Inter99 cohort. We compared original and synthetic ECGs usings Bland-Altman plots, by comparing associations with sex, age, and body mass index using linear models, and by comparing associations with all-cause mortality using Cox models. We defined that at most 5% of synthetic ECGs should have their paired original ECG as nearest neighbor in Euclidean space. Results We generated 6,612 novel, personalized electrocardiograms. Although Bland-Altman plots showed a high level of agreement between synthetic and original ECGs, only 3.7% of synthetic ECGs had their original paired ECG as their nearest neighbor. Synthetic ECGs had preserved relations between heart rate, R-wave amplitude, and T-wave amplitude and age, sex, and body mass index. Corrected QT interval was about 9.5 ms shorter in men compared to women in both the original and synthetic cohorts. The associations between PR interval and clinical characteristics were attenuated in the synthetic cohort. Heart rate and corrected QT interval were each associated with increased mortality with similar hazard ratios in the original and synthetic populations. Conclusions We demonstrated the ability of a novel neural network method to generate personalized synthetic ECGs with preserved associations with many patient characteristics and with all-cause mortality. This method may facilitate data sharing for the development of better risk prediction models.

Background: Wearables with integrated electrocardiogram (ECG) acquisition have made single-lead ECGs widely accessible to patients and consumers. However, the 12-lead ECG remains the gold standard for most clinical cardiac assessments. In this study, we developed a neural network to reconstruct 12-lead ECGs from single-lead and dual-lead ECGs, and evaluated the mathematical accuracy. Methods: We used lead I or leads I and II from 9514 individuals from the Physikalisch-Technische Bundesanstalt (PTB-XL) cohort and a generative adversarial network, with the aim of recreating the missing leads from the 12-lead ECG. ECGs were divided into training, validation, and testing (10%). Original and recreated leads were measured with a commercially available algorithm. Differences in means and variances were assessed with Student’s t-tests and F-tests, respectively. Calibration and bias were assessed with Bland-Altman plots. Inter-lead correlations were compared in original and recreated ECGs. Results: The variability of precordial ECG amplitudes is significantly reduced in recreated ECGs compared to real ECGs (all p < 0.05), indicating regression-to-the-mean. Amplitude averages are recreated with bias (p < 0.05 for most leads). Reconstruction errors depend on the real amplitudes, suggesting regression-to-the-mean (R ² between target and error in R-peak amplitude in lead V3: 0.92). The relations between lead markers have a similar slope but are much stronger due to reduced variance (R-peak amplitude R ² between leads I and V3, real ECGs: 0.04, recreated ECGs: 0.49). Using two leads does not significantly improve 12-lead recreation. Conclusions: AI-based 12-lead ECG reconstruction results in a regression-to-the-mean effect rather than personalized output, rendering it unsuitable for clinical use.

Background: Ventricular repolarization time (ECG QT and JT intervals) is associated with malignant arrhythmia. Genome‐wide association studies have identified 230 independent loci for QT and JT; however, 50% of their heritability remains unexplained. Previous work supports a causal effect of lower serum calcium concentrations on longer ventricular repolarization time. We hypothesized calcium interactions with QT and JT variant associations could explain a proportion of the missing heritability. Methods and Results: We performed genome‐wide calcium interaction analyses for QT and JT intervals. Participants were stratified by their calcium level relative to the study distribution (top or bottom 20%). We performed a 2‐stage analysis (genome‐wide discovery [N=62 532] and replication [N=59 861] of lead variants) and a single‐stage genome‐wide meta‐analysis (N=122 393, [European ancestry N=117 581, African ancestry N=4812]). We also calculated 2‐degrees of freedom joint main and interaction and 1‐degree of freedom interaction P values. In 2‐stage and single‐stage analyses, 50 and 98 independent loci, respectively, were associated with either QT or JT intervals (2‐degrees of freedom joint main and interaction P value −8 ). No lead variant had a significant interaction result after correcting for multiple testing and sensitivity analyses provided similar findings. Two loci in the single‐stage meta‐analysis were not reported previously ( SPPL2B and RFX6 ). Conclusions: We have found limited support for an interaction effect of serum calcium on QT and JT variant associations despite sample sizes with suitable power to detect relevant effects. Therefore, such effects are unlikely to explain a meaningful proportion of the heritability of QT and JT, and factors including rare variation and other environmental interactions need to be considered.

Purpose: The association between thyroid dysfunction and exudative age‐related macular degeneration (AMD) is unknown.Methods: In this Danish longitudinal nationwide registry‐based cohort study we included all Danish residents aged 50–100 between 2008 and 2018. Using the Danish national registries, we studied the association between thyroid dysfunction and exudative AMD. Thyroid dysfunction was classified as two consecutive redeemed prescriptions of thyroid hormones (hypothyroidism) or anti‐thyroid medication (hyperthyroidism). Exudative AMD was classified as an ICD diagnosis of AMD and a code for anti‐VEGF treatment. All patients are treated for exudative AMD in a hospital in Denmark, and we therefore have complete registration of this patient group.Results: We included 2 087 305 individuals, of which 1 072 567 (51.4%) were women; 59 318 (2.8%) had hypothyroidism, and 33 922 (1.6%) had hyperthyroidism. During a median follow‐up of 11 years, 26 998 (1.3%) people developed exudative AMD. Hypothyroidism (adjusted hazard ratio [HR]: 1.17; 95% confidence interval [CI] 1.10–1.25; p < 0.001) and hyperthyroidism (HR: 1.23; 95% CI:1.13–1.34; p < 0.001) were both associated with the development of exudative AMD. The age‐stratified analyses yielded similar results to the main analyses, except that the risks were exaggerated in the older part of the population.Conclusion: This is the first longitudinal nationwide study showing that both hypo‐ and hyperthyroidism are associated with an increased risk of exudative AMD. AMD is a quantitative problem in the population and our findings could have a public health impact. Further studies are needed to study the underlying mechanisms of the association.

To date only a fraction of the genetic footprint of thyroid function has been clarified. We report a genome-wide association study meta-analysis of thyroid function in up to 271,040 individuals of European ancestry, including reference range thyrotropin (TSH), free thyroxine (FT4), free and total triiodothyronine (T3), proxies for metabolism (T3/FT4 ratio) as well as dichotomized high and low TSH levels. We revealed 259 independent significant associations for TSH (61% novel), 85 for FT4 (67% novel), and 62 novel signals for the T3 related traits. The loci explained 14.1%, 6.0%, 9.5% and 1.1% of the total variation in TSH, FT4, total T3 and free T3 concentrations, respectively. Genetic correlations indicate that TSH associated loci reflect the thyroid function determined by free T3, whereas the FT4 associations represent the thyroid hormone metabolism. Polygenic risk score and Mendelian randomization analyses showed the effects of genetically determined variation in thyroid function on various clinical outcomes, including cardiovascular risk factors and diseases, autoimmune diseases, and cancer. In conclusion, our results improve the understanding of thyroid hormone physiology and highlight the pleiotropic effects of thyroid function on various diseases.

Background Some evidence suggests gene-treatment interactions might cause persistent symptoms in individuals receiving levothyroxine (LT4) treatment. We investigated, as previously hypothesized, if single nucleotide polymorphisms (SNPs) in rs225014 (Thr92Ala), rs225015, or rs12885300 (ORFa-Gly3Asp) in the Deiodinase 2 gene (DIO2), or rs17606253 in the Monocarboxylate Transporter 10 gene (MCT10) were associated with outcomes indicative of local tissue hypothyroidism in LT4-treated patients and controls. Methods We included 18,761 LT4-treated patients and 360,534 controls in a population-based cross-sectional study in UK Biobank. LT4 treatment was defined as a diagnosis of hypothyroidism and self-reported use of LT4 without use of T3. Outcomes were psychological well-being, cognitive function, and cardiovascular risk factors. Associations were evaluated by linear, logistic, or ordinal logistic multiple regression. Adjustments included sex, age, sex-age interaction, and genetic principal components 1-10. Results Compared to controls, LT4 treatment was adversely associated with almost all outcomes, most noteworthy: Increased frequency of tiredness (p < 0.001), decreased well-being factor score (p < 0.001), increased reaction-time (p < 0.001), and increased BMI (p < 0.001). Except for a significant association between the minor rs225015 A allele and financial dissatisfaction, there was no association of rs225014, rs225015, rs12885300, or rs17606253 with any outcomes in LT4-treated patients. For all outcomes, carrying the risk allele at these four SNPs did not amplify symptoms associated with LT4 treatment compared to controls. Conclusions rs225014, rs225015, rs12885300, and rs17606253 could not explain changed psychological well-being, cognitive function, or cardiovascular risk factors in LT4-treated patients. Our findings do not support a gene-treatment interaction between these SNPs and LT4 treatment.