Tim Biersa

I write about Medical AI. Topics I focus on • clinical AI research • radiology and imaging AI • AI regulation • AI safety in healthcare Goal Understanding how AI moves from research into real clinical workflows. #MedicalAI #ClinicalAI

@PhilipsHealth has received FDA 510(k) clearance for Spectral CT Verida in the US. Philips positions it as an AI-powered detector-based spectral CT system that combines always-on spectral imaging with AI-based reconstruction. The relevant point is the system design. High- and low-energy data are captured in a single acquisition, so conventional and spectral results are available together without separate scans or workflow changes. Philips says this supports tissue characterization, material differentiation, and faster CT workflow. This is less about AI as a standalone feature and more about AI embedded into the imaging chain. philips.com/a-w/about/news… #HealthcareAI #Radiology #CT #MedTech #SpectralCT

GE HealthCare receives FDA clearance for deep learning CT reconstruction tool dotmed.com/news/story/662…

The MHRA has expanded its AI Airlock programme with £3.6 million over three years. AI Airlock is the UK’s first regulatory sandbox for AI as a Medical Device. It is intended to address practical regulatory questions around AI in healthcare, including risk management, explainability, post-market monitoring, change control plans, and evolving intended use. The important point is that findings from the programme are intended to inform future AI regulation in healthcare directly. For AIaMD companies, this is a relevant signal. Regulatory expectations are increasingly being shaped through practical experience with real product cases. gov.uk/government/new… #HealthcareAI #MedTech #AIRegulation #AIaMD #MHRA

One of the clearest real-world use cases for AI in healthcare today is not diagnosis. It is documentation. A recent JAMA study across 5 US academic health systems found that AI scribes were associated with less total EHR time, less documentation time, and slightly more patient visits per week. The key takeaway is simple. AI creates real value when it reduces administrative burden and improves clinical workflow. This is not about replacing clinicians. It is about making care delivery more efficient. Important nuance. The effects were meaningful, but not transformational. After-hours EHR time did not significantly improve. So this is not a fix for burnout. It is a strong example of practical workflow value from generative AI in real care settings. jamanetwork.com/journals/jama/… #HealthcareAI #DigitalHealth #MedTech #AmbientAI #EHR #GenerativeAI

AISAP and Cardiology Consultants of Philadelphia announced a partnership to deploy an AI-enabled cardiac diagnostic platform. Focus areas heart failure valvular disease The platform supports echocardiographic analysis and clinical decision-making at the point of care. Deployment is planned across selected sites within the network, with additional joint research activities. Source prnewswire.com/news-releases/… #MedicalAI #Cardiology #ClinicalAI

A large paired noninferiority trial in breast screening evaluates AI as part of the workflow, not as a standalone reader. 31,301 women AI triage removed low-risk exams from human reading remaining cases double read with AI support Results 63.6 percent workload reduction CDR increased from 6.3 to 7.3 per 1,000 recall rate increased and not noninferior Modality matters DM showed higher detection and recall DBT maintained detection with stable recall nature.com/articles/s4159… #MedicalAI #RadiologyAI #BreastImaging

@Medtronic received FDA clearance to expand Stealth AXiS to cranial and ENT procedures. The system combines navigation, imaging and robotics. For cranial use, it applies AI to generate brain maps and highlight neural pathways. Previously cleared for spine surgery, it is now extended to additional surgical domains. Source reuters.com/business/healt… #MedicalAI #MedTech #Surgery

Federated learning does not solve the core problem of medical AI deployment. Data across hospitals is heterogeneous, imbalanced and shifts over time. This paper introduces TrustFed, focusing on something often missing reliability guarantees under distribution shift without centralizing data The key point Scaling medical AI is not only about accessing data it is about maintaining performance across sites Preprint arxiv.org/abs/2603.21656 #MedicalAI #TrustworthyAI #FederatedLearning

A key problem in medical multimodal AI is not just hallucination, but lack of grounding. This paper shows that hallucinated answers often have two signatures unstable token-level confidence minimal influence of the image on the response The important point is this A model can give a clinically plausible answer that is not driven by visual evidence The proposed method detects this directly from the model’s own probabilities no sampling no external models fully deterministic This shifts the focus from accuracy to a more relevant question Is the model actually using the image Preprint arxiv.org/abs/2603.21693 #MedicalAI #ClinicalAI #PatientSafety

A new cardiology AI preprint takes a different approach to multimodal interpretation. MARCUS combines ECG, echocardiography and CMR using modality-specific models and an orchestration layer that integrates findings across modalities. The key point is not single-model performance. It is the ability to handle clinical reasoning across modalities, where each signal provides only part of the picture. In multimodal cases, MARCUS shows a clear gap vs general-purpose models, which often fail to integrate conflicting signals. It also reduces cases of plausible but unsupported explanations through structured cross-modal verification. This suggests that progress in clinical AI may depend less on scaling one model and more on coordinating specialized models across data sources. Preprint, not peer reviewed arxiv.org/abs/2603.22179 #MedicalAI #ClinicalAI #Cardiology #MultimodalAI

Foundation models in medical imaging are not just learning disease patterns. They can also encode latent demographic and biometric signals, which may enable re-identification when combined with external data. The key point from this npj Digital Medicine commentary: The risk is not only at the level of data access or outputs, but at the level of learned representations. That changes how we need to think about privacy in medical AI: • not just data protection • but representation control • and governance of what models retain internally Paper (open access) nature.com/articles/s4174… #MedicalAI #HealthcareAI #MedicalImaging #ClinicalAI

A practical use of AI in liver MRI beyond image classification. A recent RSNA paper shows that deep learning–based multiphase arterial MRI improves • late arterial phase capture • image quality • HCC detection It also enables more reliable assessment of arterial hypervascularity, a key feature in HCC imaging. The interesting point is not the model itself, but the target: AI is improving phase timing and image acquisition quality, not just interpretation. In liver MRI, getting the arterial phase right often determines diagnostic confidence. This suggests a broader direction for imaging AI: Impact may come from improving how images are acquired, not only how they are read. Paper pubs.rsna.org/doi/epdf/10.11… #RadiologyAI #MedicalImaging #HealthcareAI #ClinicalAI

A recurring issue in radiology AI is not model performance, but site-specific variability. Models validated on curated datasets often degrade when exposed to differences in scanners, acquisition protocols, workflows, and patient populations. Recent announcements from HOPPR and @nvidia point to where the field is moving: • platforms for local fine-tuning and validation • integration of foundation models with imaging pipelines • infrastructure to support continuous performance monitoring The implication is clear: AI in medical imaging is becoming an infrastructure and deployment problem, not just a modeling problem. Sources hoppr.ai/news/does-your… prnewswire.com/news-releases/… investor.nvidia.com/news/press-rel… #HealthcareAI #RadiologyAI #MedicalImaging #ClinicalAI #MedTech

New medical AI preprint worth watching. Meissa introduces a 4B multimodal medical agent model designed for offline / on-prem deployment rather than relying on cloud APIs. Key points from the paper: • ~25× fewer parameters than typical frontier agent models • ~22× lower end-to-end latency vs API deployments • Designed for environments with strict data governance and infrastructure constraints The interesting signal is not only model performance. It reflects a growing push toward locally deployable medical AI agents that can operate inside hospital infrastructure instead of relying entirely on remote foundation models. Still a preprint (not peer reviewed) but a direction worth watching. Paper arxiv.org/abs/2603.09018 #MedicalAI #HealthcareAI #AgenticAI #EdgeAI #ClinicalAI #MedTech

New in @Nature: Merlin, a 3D CT vision–language foundation model. Trained on 6.3M CT images from 15,331 scans, paired with radiology reports and EHR diagnosis codes, the model learns joint representations of volumetric imaging and clinical text. Evaluated across 752 tasks including findings classification, report generation and disease prediction, Merlin showed strong cross-site generalization across hospitals. Source nature.com/articles/s4438… Paper nature.com/articles/s4158… #MedicalAI #RadiologyAI #FoundationModels

Interesting move from Amazon in the digital health space. Amazon has introduced Health AI on Amazon.com and in the Amazon Shopping app. The system provides health guidance, helps interpret lab results, manages medications, and can assist with booking care. According to Amazon, eligible U.S. Prime members can also access direct-message care visits with One Medical for more than 30 common conditions. It’s another example of how large tech platforms are starting to integrate AI-driven health navigation directly into consumer ecosystems. Source aboutamazon.com/news/retail/am… #MedicalAI #HealthcareAI #DigitalHealth #HealthTech

New review in @Nature on AI agents in healthcare. The paper highlights how LLM-based agents differ from standalone models by orchestrating multi-step clinical tasks through tool use and coordination, enabling goal-directed reasoning across heterogeneous medical data. Key application areas discussed • diagnostic assistance • clinical decision support • medical report generation • hospital workflow and operations Important focus of the review: evaluation, safety, and controllability of agent systems in clinical environments. Source nature.com/articles/s4438… #MedicalAI #ClinicalAI #AIAgents

First FDA-approved AI-enabled imaging device for breast cancer surgery. @PerimeterMed AI’s technology uses optical coherence tomography combined with AI to help surgeons visualize tissue margins during breast-conserving surgery. Goal: identify residual cancer during the procedure and potentially reduce repeat surgeries. Source advamed.org/industry-updat… #MedicalAI #MedTech

@Medtronic and @GEHealthCare are expanding their patient monitoring alliance. The partnership integrates Medtronic pulse oximetry and capnography into GE HealthCare monitoring platforms used in hospitals. Source medtechdive.com/news/medtronic… #MedTech #DigitalHealth

New dataset for evaluating medical LLMs. ClinConsensus introduces 2,500 open clinical cases across 36 medical specialties to test AI systems in more realistic clinical scenarios. This is a preprint, not yet peer reviewed. Source arxiv.org/abs/2603.02097 #MedicalAI #ClinicalAI