Several prior studies have developed pipelines generating personalized heart models, (e.g., Vadakkumpadan et al., 2010); however, these processes have generally been time-consuming and manual. We developed and implemented a semi-automatic pipeline for generating patient-specific ventricular models (Figure 1). All steps are fully automated, with the exception of MRI segmentation, which required manual intervention. The entire pipeline is open-source and available to the public. This semi-automated pipeline involves segmentation from MRI medical images of the heart, finite element model generation, virtual myocardial fiber generation, and node reordering as preparation for continuum model electrophysiological simulations.

We assessed the ability of machine learning classification algorithms to correctly classify virtual patient arrhythmia risk (R and UR correspond to arrhythmia while NR corresponds to no arrhythmia) as based on simple virtual patient model-derived features. In each patient model (baseline plus augmented), global ischemia volume and global ventricular volume was measured, and for each of the 17 AHA LV segments the ischemic percentage and tissue volume was measured (4 features, Figures 3A,B).

In the baseline patient model population, 17 segments in 30 patients were evaluated for global arrhythmic vulnerability. Of these 510 segments, arrhythmia appeared in 111 segments during the protocol (no arrhythmia in 399; a ratio of 0.218). In the augmented patient model population, 17 segments in 129 total patient models were evaluated. Of these 2,193 segments, global arrhythmia appeared in 285 segments (no arrhythmia in 1,908; ratio 0.130) during our protocol. A summary of the results are given in Table 1.

The full set of results of the virtual vulnerability protocol for all patient models as specified in Materials and Methods can be found in Supplementary Table 4. Supplementary Table 5 summarizes the differences between the arrhythmic and non-arrhythmic groups. In general, hearts with larger global ischemic volumes were more inducible after the pacing protocol. Additonally, pacing from segments with larger percentage of ischemic tissue were also more likely to induce arrhythmia (Oliveira et al., 2018; Martinez-Navarro et al., 2019, 2021). This result suggests a mechanistic link between location of pacing site and arrhythmia inducibility in post-MI patients. These results are consistent with other studies that have shown that ectopic beats originating from the borders of ischemic tissue are more likely to result in wavebreak and reentry formation. Additionally, a positive correlation between ischemic volume and vulnerability to arrhythmia has been widely reported in the literature (Rubenstein et al., 2008; Klem et al., 2012).

Segment-specific myocardial volume and segment-specific ischemic percentage as well as total myocardial volume and total estimated ischemic volume were calculated for each patient model in the baseline and augmented populations as detailed in Methods and Materials. Statistics on these features as well as associated arrhythmia outcomes can be found in Supplementary Table 5 (model input statistics).

Accuracy of all ML models trained and tested on data from both the baseline and augmented populations is shown in Table 2. For each of the seven ML models tested, mean predictive accuracy improved and accuracy standard deviation decreased when employing data from the augmented patient population. For all models trained and tested on data from the baseline patient population alone, SVM and logistic regression performed best in terms of mean accuracy (0.86; results among models ranged from 0.83 to 0.86). When using results from the augmented patient population, all ML models improved in accuracy (to 0.88 to 0.89; accuracy and variance of accuracy among all model trials between the baseline and augmented populations was statistically significant; p-values shown in Supplementary Table 8. Notably, 3- and 4-hl NN matched the performance of logistic regression, all performing best when considering the augmented population results.

Average precision (AP) of all ML models trained and tested on data from both the baseline and augmented populations is shown in Table 3. AP stayed the same, or modestly increased, for all ML models tested, with the exception of SVM, which decreased slightly. Sensitivity and precision of all ML models trained and tested on data from both the baseline and augmented populations at a threshold of 0.5 is additionally shown in Supplementary Tables 6, 7, respectively.

Figures 4, 5 present the ROC curves for the highest-performing models tested in both the baseline and augmented populations, knn, decision tree, and logistic regression are shown in Figure 4, while neural network ROC curves are presented in Figure 5. Supplementary Figure 2 shows results for SVM and xgboost. Corresponding AUC for all ML models trained and tested on data from both the baseline and augmented populations is shown in Table 4. While differences among models are relatively modest, logistic regression, 3-hl NN, and 4-hl NN performed similarly best in class, with confidence intervals as shown in Figures 4B, 5A,B.

Here, we have presented a combined in silico and machine learning methodology to assess risk for dangerous arrhythmia in post-infarct patients. We have aimed to briefly demonstrate that simulation-supported data augmentation can improve prediction models and overall accuracy for arrhythmia risk assessment. Briefly, we used a semi-automated image-based patient-specific modeling and simulation pipeline to create both baseline and augmented patient populations, and assessed vulnerability to reentry in both populations via a virtual programmed stimulation protocol. We then calculated specific geometric features in all patient models, and trained seven machine learning algorithms (3 classification, 1 clustering, and 2 neural networks) to predict arrhythmia outcome directly from these geometric patient model features alone.

We found that this combined approach can provide insight with high accuracy, rapidly and efficiently, with accuracy ranging from 83 to 86% for all ML models for the baseline population. Furthermore, all ML models improved in accuracy to 88–89% (accuracy and accuracy variance was statistically significant; p-values shown in Supplementary Table 8) when using results from the augmented patient population, demonstrating that, particularly in the case of small cohorts and/or sparse patient data, simulation-supported data augmentation can be employed to further improve results of predictive machine learning models.

Previous research has identified LGE volume (Klem et al., 2012) and LV mass (Haider et al., 1998) as predictors for sudden cardiac death. Because we implemented a decision tree as one of the ML models evaluated and this performed reasonably similar to other models, we were conveniently able to directly probe the decision-making in this algorithm to assess which feature(s) this model deemed as most-important for its decision making. Again, the hand-picked geometric features were: segment-specific myocardial volume and ischemic percentage, as well as total myocardial volume and total estimated ischemic volume, calculated for each patient model and segment in both the baseline and augmented populations. In both populations, the most important of the four input features tested was estimated total ischemic volume (Supplementary Figures 3, 4, respectively). However, in the augmented population, the other three features (total myocardial volume, segment-specific ischemic percentage, and segment-specific myocardial volume) were more important for decision making than in the baseline population. The decision trees for the baseline and augmented populations can be seen in Supplementary Figures 5, 6, respectively.

ML models have been utilized successfully and extensively in arrhythmia risk assessment (Feeny et al., 2020; Krittanawong et al., 2020) and in cardiovascular imaging, to diverse ends (Prakosa et al., 2013; Bernard et al., 2018; Sermesant et al., 2021). More recently, compound, explainable ML models have demonstrated improved risk prediction for ventricular arrhythmias as compared to traditional biomarkers (i.e., left ventricular ejection fraction, LVEF), as validated retrospectively in large clinical cohorts, (e.g., Ly et al., 2021). However, others in recent years have also pioneered the combination of biophysical modeling and ML approaches in arrhythmia risk assessment (Lamata, 2018). In cardiac electrophysiology and arrhythmias, applications include but are not limited to techniques for electrical mapping of the myocardium, research to uncover the basic mechanisms of arrhythmia, and arrhythmia treatment planning and management, as recently reviewed in Trayanova et al. (2021). A key utilization of biophysical model-enabled data creation in this space has been for feature augmentation to improve performance of learning schemes (Lozoya et al., 2019; Shade et al., 2020). Notably, Shade et al. (2020) used ML and personalized computational modeling in concert to accurately predict whether a patient was likely to experience AF recurrence following pulmonary vein isolation (PVI), using only pre-PVI LGE-MRI scans as input. This work shares some notable methodological similarities with the present study: only patient LGE-MRI were used as input for electrophysiological computational models, and the (baseline) patient cohort size was similar (32 vs. 0.30 in the present study). Also similarly, Shade et al. found reasonable predictive performance with traditional ML approaches and a simulation-augmented feature set. The divergence in the current study is that the input data was augmented via patient model population expansion, rather than introduction of additional features, e.g., from our electrophysiological simulations. Corrado et al. (2021) also recently demonstrated the use of a virtual patient cohort to assess risk for sustained atrial arrhythmia; an ML model (SVM) trained on local conduction velocity and action potential duration was able to accurately predict whether an arrhythmia would tether to that tissue region. In the present study, simulations also provided the ground-truth regarding patient vulnerability to arrhythmia, as required for this proof-of-concept in post-MI patients. In future work, we will indeed employ electrophysiological features from simulations themselves to assess how their incorporation improves and/or alters model performance.

Despite the successful proof-of-concept executed in this study, there are acknowledged limitations to this work. In order to expand the number of patient hearts in the augmented population, the ischemic tissue in each patient heart was reduced several times, as described in Methods and Materials. A naive approach was thus adopted as a first step and ischemic volume was not altered symmetrically, given practical limitations in terms of computational time and tractability for patient-specific biophysical simulations.

The patient population can be further augmented in several ways to explore the empirical role of class balance in classifier performance, as well as to create data of sufficient volume to explore the improved performance of vanilla NN and deep learning approaches e.g., convolutional neural networks. Approaches to be used for augmenting patient populations (the space of the patient-specific, image-derived geometries and concomitant features) include shape modeling approaches (Balaban et al., 2021) as well as generative adversarial networks (Gholami et al., 2018; Shaker et al., 2020). Next steps for this and related work research may include combining multi-organ systems for joint study (e.g., Banus et al., 2021), to both better constrain the parameter space of a personalized model and to subsequently capture plausible physiologically mechanisms.

Furthermore, we have employed LGE-MRI from patients 5 days post-MI and have considered the damaged tissue region as ischemic in the present study, rather than as an evolving necrotic/fibrotic scar region. It is known, however, that the initial region of ischemic injury evolves rapidly, spatially and functionally, and may change significantly by the time of imaging 5 days later (Anversa and Sonnenblick, 1990; Holmes et al., 1994; Ertl and Frantz, 2005; Geerse et al., 2009; Wan Ab Naim et al., 2020), introducing uncertainty into our assumptions regarding the modeling of damaged tissue.

Finally, to demonstrate potential clinical utility of the method, validation of trained ML models with e.g., paired clinical follow-up data for arrhythmia incidence for the non-augmented population would be critical. Simulation results as presented here and real clinical scenarios may be quite different; for instance, the overall clinical arrhythmia rate may differ between specific patient groups and from in silico incidence, and should be taken into consideration. Presently, while comparison between simulation results here and clinical data has not been possible due to lack of appropriate data, appropriate follow-up data and model validation is ultimately crucial for the method's translational utility.