BrainGPT, Pulse Loss Detection from Smartwatch, Predicting Drug Combination Outcomes 🚀
Health Intelligence (HINT)
2025-03-17
🚀
New Developments in Research
VoxelPrompt: AI-Powered Vision-Language Agent for Medical Image Analysis
A new study introduces VoxelPrompt, a multimodal vision-language agent designed to process 3D medical images (MRI, CT) and execute a wide range of radiological tasks through natural language interaction. Unlike traditional models limited to single-task segmentation or classification, VoxelPrompt dynamically plans, executes, and interprets medical imaging analyses, integrating structured reasoning with visual processing.
Achieved near-parity with fine-tuned single-task models for segmentation and pathology-based visual question-answering, while enabling a much broader range of tasks.
Supports open-ended medical imaging tasks, from tumor segmentation and longitudinal disease tracking to complex anatomical measurements and lesion characterization.
Uses a language-driven planning agent that predicts executable instructions for image processing, segmentation, and volumetric feature extraction, making AI-driven radiology more interactive and adaptive.
Processes images without resampling, preserving native scan resolutions and voxel spacings for higher fidelity analysis across different imaging modalities.
Evaluated on a diverse set of neuroimaging tasks, demonstrating strong generalization across multiple MRI and CT acquisitions, anatomies, and disease types.
VoxelPrompt represents a major advancement in AI-assisted radiology, unifying medical image analysis into a flexible, multimodal framework that simplifies clinician workflows while maintaining high diagnostic accuracy​.
Physics-Based AI for Drug Residence Time Prediction
A new study introduces an automated physics-based framework for predicting absolute drug residence times, a critical factor in drug efficacy and safety. The method integrates Random Acceleration Molecular Dynamics (RAMD) for identifying ligand egress pathways and Infrequent Metadynamics (iMetaD) to estimate precise unbinding kinetics, balancing accuracy, computational efficiency, and automation.
Achieved state-of-the-art residence time prediction across five major drug targets, including kinases, GPCRs, and proteases, with a 1.22 RMSE and R² of 0.80 in log-scaled predictions.
Demonstrated superior ranking and absolute time estimation of ligand dissociation, surpassing existing empirical and simulation-based models.
Optimized computational efficiency, reducing GPU time by over 70%, making residence time calculations feasible for high-throughput drug discovery campaigns.
Successfully replicated known drug unbinding pathways, accurately modeling the impact of protein flexibility, water displacement, and molecular interactions.
Proposed an exploration-exploitation framework, allowing the use of other AI and enhanced sampling methods, making the approach adaptable to various drug design workflows.
This AI-driven physics-based approach represents a breakthrough in computational drug discovery, offering a scalable and reliable tool for optimizing drug candidates with precise kinetic profiles.
Smartwatch-Based AI for Detecting Sudden Loss of Pulse
A new study introduces a machine learning-powered detection system for identifying sudden loss of pulse using consumer smartwatches. This system leverages photoplethysmography (PPG) and motion sensors to monitor cardiovascular health in real time, with the potential to improve survival rates for out-of-hospital cardiac arrest (OHCA) by enabling immediate emergency response.
Achieved 67.23% sensitivity for detecting loss of pulse in a simulated cardiac arrest model, with 1 false emergency call per 21.67 user-years, balancing sensitivity with false positive risk.
Developed and validated an AI-driven algorithm capable of distinguishing true loss of pulse from benign signal artifacts, ensuring minimal disruption to emergency medical services.
Demonstrated real-time detection capabilities, running passively on a smartwatch, with a multi-stage alert system including haptic, audio, and visual warnings before emergency escalation.
Showed strong generalizability across diverse populations, with no significant performance differences across age, sex, skin tone, or demographics.
Highlighted the potential for mass deployment, converting unwitnessed cardiac arrests into functionally witnessed events, a critical step in improving survivability from sudden cardiac events.
This study marks a major advancement in wearable AI for health monitoring, offering a scalable and automated approach to real-time cardiac event detection, with profound implications for emergency response and preventive medicine​.
Medical Hallucinations in Foundation Models: Risks, Causes, and Mitigation Strategies
A new study investigates the prevalence and impact of medical hallucinations in large foundation models used for healthcare applications. Medical hallucinations—AI-generated misinformation that appears clinically plausible—pose significant risks for diagnosis, treatment planning, and patient safety. This research presents a comprehensive framework for understanding, detecting, and mitigating hallucinations in medical AI.
Introduced a taxonomy of medical hallucinations, categorizing errors such as factual inaccuracies, spurious correlations, and fabricated sources, with real-world case studies demonstrating their clinical impact.
Conducted a benchmark evaluation of medical AI models, revealing that retrieval-augmented generation (RAG) and chain-of-thought (CoT) reasoning significantly reduce hallucination rates but do not eliminate them entirely.
Analyzed a physician-annotated dataset of medical AI responses, identifying persistent hallucination patterns even in state-of-the-art models, with diagnostic and treatment errors appearing in up to 24% of cases.
Surveyed healthcare professionals across multiple countries, finding that 91.8% had encountered AI-generated medical hallucinations, with 84.7% believing these could negatively impact patient health.
Proposed mitigation strategies, including data curation, external knowledge integration (RAG, medical knowledge graphs), and uncertainty quantification, to enhance reliability and regulatory compliance in AI-driven clinical decision support.
This study highlights the urgent need for robust safeguards in medical AI, ensuring patient safety through improved transparency, regulatory oversight, and clinician-in-the-loop validation​.
MADRIGAL: Multimodal AI for Predicting Drug Combination Outcomes
A new study introduces MADRIGAL, a multimodal AI model designed to predict the clinical outcomes of drug combinations using preclinical data. Unlike traditional models that rely solely on molecular structures or target-based features, MADRIGAL integrates structural, pathway, cell viability, and transcriptomic data, enabling more accurate predictions of efficacy and safety for both approved drugs and novel compounds.
Trained on 21,842 compounds and 953 clinical outcomes, surpassing existing drug combination models in both accuracy and generalizability.
Utilizes a transformer bottleneck module to unify diverse drug modalities while effectively handling missing data during both training and inference.
Outperformed single-modality and state-of-the-art AI models in detecting adverse drug interactions and predicting transporter-mediated drug effects.
Successfully predicted resmetirom (the first FDA-approved MASH drug) as having a superior safety profile compared to other investigational compounds.
Supports personalized cancer therapy, accurately modeling drug synergies in acute myeloid leukemia patient-derived xenografts, improving treatment selection based on patient genomic profiles.
Integrates with large language models (LLMs) to enable natural language descriptions of clinical outcomes, facilitating safety assessments and regulatory decision-making.
MADRIGAL represents a breakthrough in AI-powered drug discovery, providing a scalable, multimodal framework for optimizing combination therapies, managing polypharmacy, and reducing late-stage clinical failures.
NeuroBench: A Standardized Benchmark for Neuromorphic Computing
A new study introduces NeuroBench, the first comprehensive benchmarking framework for neuromorphic computing algorithms and hardware. Unlike conventional AI models that rely on deep learning efficiency benchmarks, NeuroBench provides hardware-independent and hardware-dependent evaluation tracks tailored for spiking neural networks (SNNs), neuromorphic processors, and bio-inspired AI systems.
Developed through a collaborative effort between academia and industry, ensuring open-source accessibility and community-driven standardization.
Establishes dual-track evaluation:
Algorithm Track assesses neuromorphic algorithms independently of hardware, measuring computational complexity, efficiency, and task performance.
System Track evaluates real-world deployment on neuromorphic hardware, quantifying latency, energy efficiency, and compute performance.
Benchmarks a diverse set of neuromorphic AI tasks, including few-shot continual learning, event-based vision, motor decoding, and chaotic time-series prediction.
Introduces new performance metrics, such as synaptic operations, connection sparsity, activation sparsity, and memory footprint, offering fine-grained insights into neuromorphic efficiency.
Showcases strong energy efficiency gains, with neuromorphic systems consuming up to 37× less power than conventional deep learning models in certain tasks.
Designed for continuous evolution, with a centralized leaderboard, workshops, and competitions, fostering progress in neuromorphic AI.
NeuroBench sets a new standard for evaluating neuromorphic computing, offering a scalable, unbiased framework that enables fair comparisons across hardware and algorithmic approaches, accelerating progress in bio-inspired AI​.
BrainGPT: Multimodal AI for 3D Brain CT Radiology Report Generation
A new study introduces BrainGPT, a clinically visual instruction-tuned (CVIT) multimodal LLM designed to generate structured radiology reports from 3D brain CT scans. Unlike prior models trained on 2D medical images, BrainGPT processes volumetric scans, offering anatomy-aware insights with improved diagnostic accuracy.
Trained on 18,885 3D brain CT scans with corresponding reports, creating the largest dataset for multimodal CT report generation.
Achieved 74% indistinguishability from human-written reports in a physician-led Turing test, validating its clinical relevance.
Introduced Feature-Oriented Radiology Task Evaluation (FORTE), a new evaluation metric that surpasses traditional NLP-based scoring in assessing diagnostic completeness and accuracy.
Demonstrated high recall for key radiology terms (e.g., infarcts, hemorrhage, atrophy), enabling better structured clinical documentation.
Validated on the CQ500 external dataset, maintaining strong generalization across different brain pathologies.
BrainGPT represents a major step forward in AI-powered radiology, providing a scalable framework for automated, high-quality radiology reporting in 3D medical imaging.
MorphLDM: AI-Generated Brain MRI Using Deformable Templates
A new study introduces MorphLDM, a latent diffusion model (LDM) designed to generate morphologically accurate 3D brain MRIs by applying learned deformation fields to a template, rather than synthesizing images directly. Unlike traditional GAN or diffusion-based models, MorphLDM enables structural variations aligned with attributes such as age, sex, and disease state.
Achieved state-of-the-art FID (202.49) and MS-SSIM (0.73) scores, indicating superior image fidelity and diversity compared to existing generative models.
Improved age and sex accuracy in synthetic MRIs, reducing mean absolute error (MAE) by 60% compared to baseline LDMs.
Used a learned universal template, dynamically modified through generated deformation fields, ensuring data-driven realism without precomputed biases.
Outperformed GANs, autoregressive models, and previous LDMs on voxel-based morphometry, reducing effect size discrepancies in brain region volumes.
Demonstrated robust generalization across ages and conditions, particularly improving MRI synthesis for older populations with fewer training samples.
MorphLDM marks a significant advancement in generative neuroimaging, providing a more biologically plausible and attribute-controlled approach for synthetic brain MRI generation and augmentation.
Love Health Intelligence (HINT)? Share it with your friends using this link: Health Intelligence.
Want to contact Health Intelligence (HINT)? Contact us today @ lukeyunmedia@gmail.com!
Thanks for reading, by Luke Yun