This week in Cardiovascular AI
Summarizing the recent literature on Cardiovascular AI
Welcome back to This Week in Cardiovascular AI - a bi-monthly email newsletter for followers and subscribers that summarizes recent research in Cardiovascular AI and Digital Health that we find important and unique.
The biggest article to discuss this week comes from Nature Cardiovascular Research discussing the development of MAARS, an advanced AI tool for more accurately predicting sudden cardiac death risk in the hypertrophic cardiomyopathy population - something that has historically been very difficult. Plus, in our bonus section, What else I read…, we review new advances in AI drug discovery and fully autonomous robotic surgery!
Not a subscriber? Join now to stay up-to-date and consider upgrading to support our work!
Multimodal AI to Forecast Arrhythmic Death in Hypertrophic Cardiomyopathy
Nature Cardiovascular Research
This study introduces MAARS (Multimodal AI for ventricular Arrhythmia Risk Stratification), an advanced deep learning model trained to predict sudden cardiac death from arrhythmia (SCDA) in patients with hypertrophic cardiomyopathy (HCM). Traditional risk prediction tools underperform in this population for a number of reasons. MAARS integrates three modalities—structured electronic health records (EHR), contrast-enhanced cardiac MRI (LGE-CMR), and cardiac imaging reports (CIR)—via transformer-based architectures with late multimodal bottleneck fusion. The model is trained and validated on a Johns Hopkins HCM cohort and externally validated on an independent registry from Sanger Heart & Vascular Institute.
Key Findings:
Model Design:
Uses separate transformer encoders per modality (ViT for LGE-CMR, dense feedforward for EHR and CIR).
Employs late bottleneck fusion across modalities.
Incorporates attention rollout and Shapley value decomposition for interpretability.
Internal Validation (JHH-HCM Cohort):
N = 553 patients with LGE-CMR from Johns Hopkins
Event rate: 3.4% (19 SCDA events)
AUROC: 0.88 [95% CI: 0.83–0.93]
Balanced Accuracy: 0.81 [0.66–0.91]
MAARS outperformed all clinical guidelines and risk tools (AUC 0.41–0.72 for comparators)
External Validation (SHVI-HCM Cohort):
N = 286 patients from Atrium Health
Event rate: 9% (25 SCDA events)
AUROC: 0.87 [95% CI: 0.79–0.92]
Balanced Accuracy: 0.76 [0.61–0.91]
Maintained strong calibration and predictive accuracy across subgroups by age, sex, and imaging vendor
Feature Insights:
Most important predictors included LGE burden, septal wall thickness, diastolic blood pressure, and key radiology report terms.
Model maintained high performance with reduced feature sets, showing feasibility for future real-world simplification.
Clinical Takeaways:
MAARS provides the first externally validated, multimodal AI tool for SCDA risk stratification in HCM, addressing limitations of traditional tools that rely heavily on LVEF or are poorly calibrated for this population.
Superior to Guidelines: MAARS outperforms current HCM risk calculators (e.g., HCM Risk-SCD, ACC/AHA, ESC tools), offering a much-needed alternative for patients with preserved LVEF.
Potential use case: pre-implantation risk assessment for ICDs in HCM patients with preserved EF but other risk features (e.g., fibrosis, abnormal vitals, family history).
The external validation on a geographically and demographically distinct cohort (SHVI-HCM) enhances its generalizability.
Read it here: Link
Artificial Intelligence-Based Screening of Cardiac Amyloidosis from Standard Transthoracic Echocardiograms
European Heart Journal
This study developed and validated a deep learning (DL) model that identifies cardiac amyloidosis (CA) using only a single apical four-chamber transthoracic echocardiographic (TTE) video clip. The AI model was trained and calibrated using data from Mayo Clinic and externally validated in a large, diverse, international cohort. The model outperformed conventional CA risk scoring systems and has the potential to enable automated, scalable screening in both high- and low-resource settings.
Key Findings:
Model Architecture:
An ensemble of five 3D convolutional neural networks (CNNs) trained via five-fold cross-validation.
Each model generated prediction trajectories across video frames; these were time-averaged and further ensemble-averaged.
The final output included:
A CA probability score
A binary classification (likely/unlikely CA)
A flag for uncertainty, based on instability and entropy thresholds.
Training & Validation:
Training set: 1,349 CA + 1,263 controls.
Tuning set: 7,174 patients with and without HFpEF.
External validation: 2,719 patients from five continents (597 CA; 2,122 controls).
The model used only grayscale echo videos without Doppler, ECG gating, or patient metadata.
Performance:
Internal tuning: AUC = 0.93, Sen= 89.0%, Spec = 96.2%, NPV = 99.8% in low-prevalence settings (~1%)
External validation: AUC = 0.93, Sen = 85.0%, Spec = 93.1%, PPV = 78.0%, NPV = 95.6%
Uncertainty Estimation:
The model labels predictions as "uncertain" (~13.4% of cases) when prediction entropy or instability is high, improving robustness and clinical safety.
Comparative Performance:
The DL model outperformed:
Increased Wall Thickness (IWT) score
Transthyretin Cardiac Amyloidosis Score (TCAS)
It improved net clinical benefit and decision curve analysis over both traditional models.
Using the model for screening prior to technetium-PYP (Tc-PYP) scans:
Increased true positives by 36.4%
Reduced false positives by 6.9%
Improved test efficiency across amyloid subtypes (ATTRwt, ATTRv, AL)
Clinical Takeaways:
Transformative Screening Potential: Enables automated, early identification of cardiac amyloidosis using only a single echo view — applicable even in routine TTE workflows.
Scalable Across Institutions: The model generalizes well across geographies, scanner vendors, and echo technicians, supporting widespread clinical use.
Enhances Diagnostic Yield: Outperforms conventional scores in both high- and low-prevalence settings, improving diagnostic confidence while minimizing over-testing.
Supports Uncertainty-Aware AI: Incorporating entropy and instability allows the model to safely defer when uncertain, a major step toward trustworthy AI in echocardiography.
Read it here: Link
Novel CAC Dispersion and Density Score to Predict Myocardial Infarction and Cardiovascular Mortality
Circulation: Cardiovascular Imaging
This study introduces the CAC-DAD score, a deep learning–driven coronary artery calcium (CAC) metric that incorporates both dispersion (spatial location) and density (attenuation in Hounsfield units) of coronary calcifications. Compared with the traditional Agatston score (AS), the CAC-DAD score demonstrated superior prognostic accuracy for short-term major adverse cardiovascular events (MACE), particularly in patients undergoing cardiovascular risk or perioperative assessment.
Key Findings:
Study Design & Population:
961 patients undergoing cardiac CT for cardiovascular risk assessment or preoperative evaluation.
Median age: 67 years; 61% male.
Outcomes tracked: 61 MACE events (nonfatal MI or CV death).
CAC-DAD score calculated using fully automated deep learning algorithms.
Model Performance:
AUC (C-statistic) for MACE:
CAC-DAD model = 0.65
AS model = 0.61
Combined model = 0.66
Multivariable predictors of MACE:
CAC-DAD score ≥2050: HR 2.57 (P=0.002)
Age, diabetes, and statin use also significant
AS ≥400 lost significance in multivariable models
Risk Reclassification:
Patients with high CAC-DAD but low AS had higher MACE risk (HR 3.00)
Patients with low CAC-DAD and high AS had no observed MACE
Suggests CAC-DAD reclassifies risk better than AS alone
Clinical Simulation:
In statin-naive patients, CAC-DAD remained predictive
In statin users, CAC-DAD was not discriminatory, possibly due to plaque modification effects
Clinical Takeaways:
CAC-DAD = A Next-Gen Risk Tool: It refines CAC scoring by accounting for plaque location and attenuation, two factors neglected in traditional Agatston scoring.
Outperforms Agatston Score: Especially in scenarios where high calcium density or proximal dispersion alters risk — including patients previously deemed low risk by AS.
Supports Personalized Risk Stratification: Offers potential to up-classify overlooked high-risk patients and de-escalate low-risk ones with high AS but stabilizing features.
AI-Augmented, Fully Automated: The model automates image segmentation, lesion mapping, and metric calculation, enabling rapid and reproducible application across CT vendors.
Read it here: Link
Improving Clinical Utility of Fetal Cine CMR Using Deep Learning Super-Resolution
Circulation: Cardiovascular Imaging
This prospective, single-center study evaluated the impact of deep learning (DL)-based super-resolution reconstruction on fetal cardiovascular magnetic resonance (CMR) imaging. By applying a 2-step convolutional neural network (CNN) pipeline to low-resolution cine acquisitions (cineDL), the authors demonstrated significantly shortened scan times and preserved or improved image quality compared to conventional high-resolution cine imaging (cineNR). The study found that cineDL improved contrast, was less sensitive to fetal motion, and led to additional diagnostic findings in fetuses with congenital heart disease (CHD).
Key Findings:
Study Design:
42 third-trimester pregnant individuals with suspected or confirmed fetal CHD.
CineDL sequences reconstructed from compressed sensing-based low-resolution acquisitions.
CNN pipeline included Adaptive-CS-Net (denoising) and Precise Image Net (artifact suppression and upscaling).
Acquisition Time:
Reduced by 46.8% using cineDL (134±9.6 s vs. 252±8.8 s; P<0.001).
Two slices could be acquired per breath-hold with cineDL.
Quantitative Metrics:
No significant difference in apparent signal-to-noise ratio (aSNR) (P=0.395).
Higher contrast-to-noise ratio for cineDL (23.1±11.3 vs. 19.4±9.4; P=0.021).
Sharper edge rise distance in cineDL (1.8±0.2 mm vs. 2.1±0.5 mm; P=0.002).
Volumetric & Functional Analysis:
No significant differences between cineDL and cineNR for metrics like LV/RV EDV or EF.
Excellent reproducibility (ICC > 0.95 for most parameters).
Diagnostic Yield:
CineDL enabled visualization of additional cardiac structures in 24% (10/42) of fetuses.
Additional pathologies identified in 12% (5/42), including VSDs, d-TGA, and partial anomalous pulmonary venous return.
Some findings (e.g., tricuspid valve visualization, anomalous veins) were confirmed postnatally.
Artifact Profile:
CineDL had more pronounced backfolding and aliasing artifacts, but these were distinguishable and did not impede clinical assessment.
CineDL was less affected by fetal and respiratory motion than cineNR.
Clinical Takeaways:
Shorter Exams with Better Quality: Deep learning super-resolution reconstruction substantially reduces scan time while preserving diagnostic quality, addressing a key barrier in fetal CMR.
Enhanced Diagnostic Confidence: CineDL increased the ability to detect structural abnormalities, especially in challenging cases where fetal motion or maternal breath-holds limit conventional image clarity.
Operationally Feasible: Reconstructions were performed in under 1 minute directly on the scanner’s hardware, making it scalable for clinical settings without requiring off-line processing.
Potential Clinical Impact: Broader adoption of DL-enhanced cine CMR could support more timely and accurate prenatal planning in high-risk CHD pregnancies.
Caveats: Slightly more imaging artifacts and the need for robust hardware may still limit some settings, though overall clinical interpretation was not impaired.
Read it here: Link
What else I read…
AI Breakthrough in Drug Discovery: A New Era for Idiopathic Pulmonary Fibrosis Treatment
From Nature Medicine
Artificial intelligence (AI) has long held the promise of revolutionizing drug discovery, and a recent clinical milestone is turning that promise into reality. A randomized, double-blind, placebo-controlled Phase 2a trial of an AI-discovered drug, rentosertib, for idiopathic pulmonary fibrosis (IPF) has demonstrated both safety and early signs of efficacy. This marks a significant step forward for AI-enabled drug discovery in a clinical setting.
The Challenge of IPF and the AI Solution
Idiopathic pulmonary fibrosis (IPF) is a debilitating, progressive lung disease with limited treatment options. Current standard-of-care therapies can slow disease progression but don't reverse lung damage, highlighting an urgent need for novel treatments. This is where AI steps in.
Insilico Medicine's generative AI-based biology and chemistry platforms were instrumental in this breakthrough. The AI-driven PandaOmics platform analyzed transcriptomic and multi-omic datasets to identify Traf2-and Nck-interacting kinase (TNIK) as a novel therapeutic target for IPF. TNIK is a signaling protein involved in WNT and TGFβ pathways, which are implicated in fibrosis, but it had not been previously explored as an IPF target.
Following target identification, the Chemistry42 generative AI platform designed and optimized rentosertib, a small-molecule inhibitor of TNIK. This entire process, from de novo target discovery to the completion of Phase 1 trials, took less than 30 months, significantly faster than industry averages.
Clinical Trial Results: Safety and Promising Efficacy
The Phase 2a trial involved 71 IPF patients who received either a placebo or rentosertib at three different dosing regimens over 12 weeks. The primary endpoint, the incidence of treatment-emergent adverse events, was similar across all groups, demonstrating an acceptable safety profile, even for patients concurrently receiving standard-of-care antifibrotic therapies.
Encouraging trends were observed in secondary endpoints. Patients on the highest dose of rentosertib experienced a mean improvement in forced vital capacity (FVC) of +98.4 ml, compared to a decline of -0.3 ml in the placebo arm. While modest, this FVC improvement is notable given the short treatment duration and the progressive nature of IPF, where stabilization is considered a meaningful clinical outcome. Subgroup analyses suggested that rentosertib might be most beneficial for patients not receiving concurrent antifibrotic therapy, indicating an area for future investigation regarding potential drug-drug interactions.
Beyond clinical endpoints, serum proteomics profiling revealed that rentosertib modulated the expression of fibrosis- and aging-related proteins, providing mechanistic evidence of target engagement and suggesting potential biomarkers for treatment response. This data can be used to further train AI models, creating closed-loop systems for improved drug design based on real-world treatment outcomes.
Why is this an AI Success Story?
So, why did rentosertib succeed where many other AI-generated drug candidates have not? The key lies in the sophisticated design of the AI platforms. The platform used by Xu et al. was designed to:
Validate targets using a "time machine" framework: Models trained on historical data were tested on their ability to predict targets later pursued by the pharmaceutical industry.
Integrate diverse data: The system combined multi-omic data, network analysis, and biomedical literature using complementary AI techniques, converging on TNIK.
Employ intelligent compound design: The AI system targeted both the TNIK hinge region and less-conserved allosteric pockets for improved selectivity. Compounds were filtered for drug-likeness, novelty, and synthetic accessibility, with lead optimization focusing on absorption, distribution, metabolism, and excretion (ADME) properties.
These strategic design choices, including prospective validation against future data and early attention to drug-likeness, likely contributed to rentosertib's clinical progress.
The Future of AI in Drug Development
The success of rentosertib affirms that AI can do more than just generate molecules faster and cheaper; it can truly guide discovery and de-risk development. This trial represents a turning point, demonstrating that AI can reshape how we develop medicines. As AI continues to integrate across the drug development pipeline, with advancements in areas like biologically informed representation learning and multimodal data integration, we can anticipate even more transformative impacts on drug discovery and personalized medicine.
…
Google Blog Post Gives More Details about MedGemma
From Google Research
MedGemma is Google's new open-model collection designed to foster innovation in health AI. It's a key part of the Health AI Developer Foundations (HAI-DEF) initiative, which emphasizes providing lightweight, accessible models. This approach gives developers crucial control over aspects like data privacy, infrastructure choices, and the ability to modify the models to fit specific needs.
Key Models within MedGemma:
MedGemma 27B Multimodal: This model is a powerhouse for interpreting complex health data. It's particularly adept at handling multimodal information, meaning it can process and understand data from various sources, and longitudinal electronic health records (EHRs), which track patient information over time. Its strength lies in tasks that require generating free-text responses, such as summarizing medical notes or providing detailed descriptions from imaging data.
MedSigLIP: In contrast to MedGemma's free-text generation, MedSigLIP is a lightweight image and text encoder ideal for tasks that yield structured outputs, like classification or search functionalities. It's designed for efficiency in image-related applications.
Accessibility and Performance:
Both MedGemma and MedSigLIP are engineered to be resource-efficient. They can run on a single GPU, making them accessible to a wider range of developers and institutions. Furthermore, the smaller MedGemma 4B variant and MedSigLIP are adaptable for mobile hardware, opening up possibilities for on-the-go health AI applications. A significant aspect of these models is their adaptability: MedGemma 4B, for instance, can achieve state-of-the-art performance after fine-tuning for specific tasks.
Benefits of an Open-Source Approach in Healthcare AI:
The open nature of the MedGemma collection is a critical advantage, especially in the sensitive domain of healthcare. It offers:
Flexibility: Developers can tailor the models to their unique requirements.
Privacy: Control over the models means developers can better manage data privacy, a paramount concern in healthcare.
Customization: The ability to modify the models allows for specialization for diverse medical applications.
Reproducibility: Open models promote reproducibility in research and development.
Stability: An open and collaborative environment often leads to more stable and robust models over time.
Real-World Applications and Getting Started:
Developers are already putting MedGemma and MedSigLIP to use in practical scenarios. MedSigLIP has been employed for tasks like triaging chest X-rays and detecting nodules, aiding in early diagnosis and patient management. MedGemma has been utilized for working with traditional Chinese medical literature and summarizing patient progress notes, demonstrating its versatility across different medical contexts.
Google is actively supporting developers by providing resources such as detailed notebooks and deployment options within Vertex AI, making it easier to integrate these powerful AI models into health applications. This initiative aims to democratize access to advanced health AI tools, fostering a collaborative environment for breakthroughs in medical technology.
…
Major Milestone for Autonomous Robotic Surgery
From Science Robotics
Researchers at Johns Hopkins University have achieved a significant breakthrough in robotic surgery, successfully using an AI-trained robot to autonomously remove gallbladders from dead pig organs. This accomplishment suggests that fully autonomous surgical procedures on humans could be trialed within the next decade.
The AI-powered robots were trained by observing video footage of human surgeons performing operations. During the trials, the robots demonstrated impressive capabilities, including the ability to correct mistakes, request different tools, and adapt to anatomical variations. Although slightly slower than human surgeons, the robots exhibited smoother movements and plotted shorter trajectories between tasks. Out of eight operations conducted on pig organs, the robots achieved a 100% success rate.
Currently, most of the 70,000 robotic procedures performed annually in the NHS in England are human-controlled, with only a small number, such as bone-cutting for hip and knee operations, being semi-autonomous. The UK's Health Secretary, Wes Streeting, has highlighted increasing robotic surgery as a core part of a 10-year plan to reform the NHS and reduce waiting lists, aiming for nine out of ten keyhole surgeries to be robot-assisted within a decade, up from one in five today.
While this achievement is a major step forward, experts like John McGrath, chair of the NHS Robotics Steering Committee, emphasize that fully autonomous surgery is still some years away from clinical application. Challenges remain in accurately testing a robot's ability to respond to dynamic factors in living patients, such as movement, breathing, blood, and accidental trauma. However, the successful trials on pig organs provide concrete evidence for the potential of AI to enhance surgical precision and potentially transform healthcare by reducing waiting times and improving outcomes.
…
Alphabet’s Isomorphic Labs Gearing Up for Clinical Trials
From Fortune
Alphabet's Isomorphic Labs, a subsidiary of Google DeepMind, is preparing to launch its first human clinical trials for AI-designed drugs, marking a significant step towards its ambitious goal of revolutionizing disease treatment and potentially solving all diseases with artificial intelligence.
The company leverages advanced machine learning algorithms, particularly building upon the Nobel Prize-winning AlphaFold system (which accurately predicts protein structures and molecular interactions), to analyze vast biological data, identify drug targets, and design novel compounds. This AI-driven approach aims to accelerate drug discovery, enhance treatment precision and efficacy, and significantly reduce the time and cost associated with traditional drug development, which can take years and have success rates as low as 10% in trials.
Colin Murdoch, President of Isomorphic Labs and Google DeepMind's chief business officer, confirmed that human trials are "very close" and that the company is "staffing up" for this new phase. While the trials will initially focus on oncology (cancer) and immunology candidates, the long-term vision is to extend research to a broader range of diseases.
Isomorphic Labs was spun out of DeepMind in 2021. In 2024, it formed major research collaborations with pharmaceutical giants Novartis and Eli Lilly, partnerships valued at nearly $3 billion. Additionally, in April 2025, the company secured $600 million in its first external funding round, led by Thrive Capital, to further develop its AI drug-design engine and advance its internal drug candidates into clinical development.
The success of these trials could transform healthcare by offering more targeted and effective treatments, leading to improved patient outcomes and more efficient use of resources in the pharmaceutical industry. The company envisions a future where AI models could generate effective drug designs "at the click of a button". However, the journey from AI-designed drugs to approved treatments still involves navigating regulatory hurdles and addressing potential safety concerns.