Sex-specific virtual population for the prediction and assessment of arrhythmia risk

Figures

• 

  Figure 1. Workflow for constructing the virtual populations and the QTc distributions. (A) Overview of the population generation pipeline. Each virtual subject is represented by a 1D ventricular cable model comprising Endo, M, and Epi cells. Model parameters are modulated by sex hormones and sex-specific ion-channel expression. After sequential filtering, the final virtual populations are obtained. The right panel shows the mean pseudo-ECG waveforms and standard deviations (SDs) for the final male and female populations. (B) QTc distributions of the raw virtual populations and the reference clinical QTc distributions (dashed lines). The clinical male QTc distribution was modelled as a Gaussian with μ = 388.45 ms and σ = 25.52 ms, and the clinical female distribution with μ = 407.01 ms and σ = 23.38 ms. (C) QTc distributions of the male and female virtual populations were obtained by distribution-matched sampling from the raw populations.

• 

  Figure 2. Evaluation of the consistency between clinical and simulated ECG feature changes over time and concentration after dosing. (A) Time courses of ECG feature changes within 24 h after dosing for four drugs, comparing clinical observations and simulation predictions. The column directions are presented in the following sequence (left to right): clinical males, simulated males, clinical females, and simulated females. Rows (top to bottom): dofetilide, quinidine, ranolazine, and verapamil. The y-axis shows feature changes (ΔΔ), the error bars in the clinical data indicate 95% confidence intervals, and the x-axis shows time after dosing. Simulated trajectories are summarized by the population median, and the vertical bars denote the pointwise 25th–75th percentile across virtual individuals. (B) Concentration–effect relationships for ECG features, comparing clinical data (solid lines) and simulations (dashed lines). The clinical curves are fitted from subject-level ΔΔ data, whereas the dashed simulated lines are linear fits to population-mean simulated responses at each concentration-time point. Grey shading denotes 95% confidence intervals for the clinical fits. Distributions of…

• 

  Figure 3. Parameter distributions in the virtual populations. (A) Heatmap of differences between female-specific and male-specific parameter means (female minus male). Red indicates higher parameter values in females, while blue indicates higher values in males. (B) Violin plots of four selected parameters across the three myocardial regions (Endo, M, and Epi). The left half of each violin denotes the male distribution and the right half denotes the female distribution.

• 

  Figure 4. QTc distributions and risk prediction in sex-specific LQTS populations. (A) QTc distribution in the male LQTS populations. (B) QTc distribution in the female LQTS populations. (C) The ratio of risk of LQTS in the female group to that in the male group. The blue lines represent our simulation prediction results, white the other colours represent clinical data. The first three columns of clinical data (LQT1–3) are from the same study, and the last column (LQT, without subtype distinction) is from other studies. The grey dotted line is the reference line for a risk ratio of 1.

• 

  Figure 5. Changes in ECG features following hormone interventions in sex-specific populations. (A–D) Simulation results: distributions of JTmax before and after testosterone administration in females (A), JTmax before and after testosterone removal in males (B), QTc before and after testosterone administration in females (C), and QTc before and after testosterone removal in males (D). Panels A and B show 100 randomly sampled individual models and panels C and D show 30. Differences were assessed using two-sided t tests; ***P < 0.001. (E–F) Clinical findings reported by Bidoggia et al. (2000): shortening of JTmax in androgenized females (E) and prolongation of JTmax in castrated males (F). (G) Clinical QTc before and after transdermal testosterone therapy in women with chronic heart failure (Schwartz et al., 2011); the light grey dashed line indicates the placebo group. (H) Clinical QTc before and after androgen deprivation therapy in men (Hasegawa et al., 2021), showing significant QTc prolongation during treatment.

• 

  Figure 6. Drug-induced arrhythmia risk in normal male and female virtual populations. (A) Arrhythmic event risk for 109 drugs in the normal male and female virtual populations. Drugs are ordered by male risk from high to low. The x-axis lists drug names, and colours denote clinical risk categories: Known Risk (red), Possible Risk (orange), Conditional Risk (yellow), and No Risk (green). Bars indicate simulated risk in males (blue) and females (pink). (B) Female-to-male risk ratio aggregated by clinical risk category, and its association with risk category (Spearman ρ = −0.512, P = 1.263 × 10−8). (C) Relationship between drug-induced ΔQTc and event risk across the 109 drugs, with linear fits. Regression parameters in females: k = 0.00401, b = 0.0285 (R2 = 0.869); in males: k = 0.00258, b = 0.0137 (R2 = 0.951).

• 

  Figure 7. Sex differences for drugs with ΔQTc < 10 ms. (A) Scatter plot of QTc changes (upper bound < 10 ms) vs. event risk. (B) Sex differences in risk for drugs labelled as Known Risk and Possible Risk. Horizontal black lines denote medians, and vertical black lines show interquartile ranges. Paired Wilcoxon test, P < 0.001.

• 

  Figure 8. Sex-specific risk drivers and channel-sensitivity differences under multichannel block. (A) Interaction coefficients γk from the ridge regression interaction model, which included sex, main effects of channel inhibition, and sex-by-channel interaction terms. The response was log(1 + risk), where risk is the probability of arrhythmic events for each drug in the corresponding sex-specific population. Points indicate bootstrap medians and error bars show bootstrap 95% confidence intervals. Positive γk indicate that the channel block produces a larger risk increase in females than in males, while a negative γk indicate that the channel block tends to be more risk-reducing in females. (B) Channel-sensitivity analysis within the observed domain. Bar heights represent bootstrap medians and error bars show bootstrap 95% confidence intervals.