Jeff Leek

9 Key Pharma–Digital Health Deals from May You Should Know About >> 💊 Sanofi and Regeneron have partnered with Viz.ai to deploy an AI-powered platform for earlier detection and management of COPD. The system uses natural language processing and EHR data to flag high-risk patients and streamline triage and care coordination 💊 AstraZeneca has teamed up with Aptar Digital Health to develop AI tools that detect early signs of chronic kidney disease (CKD) via retinal scans. The goal is to support earlier diagnosis during routine eye exams and embed digital tools into CKD care 💊 Boehringer Ingelheim has signed a multi-year deal with Tempus AI to enhance its oncology pipeline using real-world data and advanced analytics. The partnership provides access to over 8 million patient records to help accelerate research and personalize treatment development 💊 Boehringer Ingelheim has also partnered with Datavant to securely integrate real-world patient data into more than 75 clinical trials. The goal is to accelerate development while preserving patient privacy through secure, de-identified data linkage 💊 Abbott is integrating glucose data from its FreeStyle Libre system into Epic’s EHR, allowing real-time access to patient trends during clinical care. The move streamlines diabetes management and brings wearable data into provider workflows 💊 Moderna is expanding its use of Benchling’s cloud R&D platform across its entire research organization to replace legacy tools and support AI-driven experiment design, sample tracking, and data analysis 💊 Merck Healthcare and Dawn Health have launched a digital ecosystem for growth hormone deficiency, combining child-friendly tools, caregiver support, and real-time clinical insights. Future updates will include AI-powered growth predictions and connected device integration 💊 Takeda has partnered with Salesforce to roll out its Life Sciences Cloud across medical, commercial, and patient support teams. The move also signals a shift away from Takeda’s longtime CRM provider, Veeva 💊 Johnson & Johnson MedTech has joined forces with Qure.ai to improve early detection of lung cancer in India. AI screening tools will be deployed across 30 clinics to identify lung nodules at earlier, more treatable stages using X-ray and CT scan analysis 👇 List of all news sources in comments #DigitalHealth #AI #Pharma

824

26 Comments

Replicating an Economic Model ≠ Evaluating an AI Model Many early GenAI use cases in health economics (HE) focus on replicating existing models. But replication isn’t evaluation. Some GenAI tools can replicate existing health economic models with results that perfectly align with human-built models in terms of outputs like ICERs. Sounds impressive — but is it enough? Or is this really the direction we should take when evaluating AI-assisted HE models? Is it feasible or realistic that every time in real HTA submissions we build two models (AI vs. human) just to compare? ✅ Replication may test a single model’s performance, but it doesn’t advance the evaluation framework that’s needed to assess trustworthy and transparent AI-assisted modelling. Let me ask a simple question to my HEOR and HTA colleagues: If there’s an AI tool that replicates a model in disease area X, producing exactly the same outputs (Cost, QALYs, ICER) as an existing human-built model — does that alone remove your concerns about its transparency and reliability? And if you use that same AI tool to build a de novo model from scratch, how can you guarantee it will still be 100% aligned with a hypothetical human-built model (if one exists)? ✅ Also — let’s be fair. Humans make errors too. Models are built with uncertainty, and decisions are probabilistic, influenced by evolving evidence. Deviation between AI-built and human-built models does not necessarily mean AI is wrong. Maybe new data is emerging. Maybe AI uses different statistical methods that are actually better than Excel. Maybe it applies updated assumptions. As the classic quote says: “All models are wrong, but some are useful.” ✅ When we talk about “evaluating AI-built HE models”, we’re actually talking about two layers of evaluation: Evaluating the HE model itself — calculations, parameters, assumptions, settings, as highlighted in the NICE HTA Lab report. Evaluating the AI tool that builds the model — its trustworthiness, transparency, and reproducibility. This second layer is foundational. Every pharma, biotech, consultancy can build their own AI tool. There’s no single “universal” AI tool for HE modelling — nor should there be. For trust to be earned, we must evaluate the AI itself: What AI architecture is being used? If it’s RAG, what knowledge base is feeding it? If it’s an Agentic model, what supporting evidence does it output (e.g., R/Python survival analysis code, documentation of decision steps)? Replication and comparison are useful starting points. But it’s time for the HEOR community to move beyond replication and build robust evaluation frameworks tailored to GenAI’s strengths and risks — especially around transparency, documentation, and human oversight. 👉 NICE HTA Lab project: https://lnkd.in/e7rXkcVH Curious how others in #HEOR and #HTA are approaching this. Let’s discuss. #GenAI #HTA #Transparency #AgenticAI # #HTAinnovation

10

Quick update on AI × biotech this week - over $300M deployed, and a clear pattern emerging. The standout isn't just the funding. It's how institutions are solving the data problem. Federated learning showed up in 4 of the 5 major announcements: pharma consortiums, cancer centers, and infrastructure platforms all using it to collaborate without exposing proprietary data. This matters because the same challenge exists inside our own organizations. Data siloed across divisions, regions, and functions. Privacy and compliance creating walls. Federated approaches could unlock internal collaboration the same way they're unlocking industry collaboration. Worth watching how this evolves and whether we can apply these models internally. Five developments from the past week: 1. Periodic Labs raised $300M for autonomous lab automation. Largest seed round I've seen in this space. 2. Bristol Myers Squibb, Takeda, and Astex joined AbbVie and J&J in a federated learning consortium to train OpenFold3 on protein-small molecule data. Data stays private, models learn collectively. 3. Illumina launched BioInsight, a new business unit focused on AI-driven analysis of multiomics data. Stock was up 6.6% on the news. 4. Dana-Farber, Fred Hutch, Memorial Sloan Kettering, and Johns Hopkins launched the Cancer AI Alliance (CAIA)—federated learning across 1M+ patient records without sharing raw data. 5. Mira Murati (former OpenAI CTO) released Tinker, an API for fine-tuning AI models for scientific applications. Sources in comments. #AIBiotech #DrugDiscovery #FederatedLearning

97

6 Comments

LLMs are getting commoditized. The pace of core model breakthroughs has clearly slowed. We are no longer seeing step-function jumps from “just a bigger model”. For most real systems today: • architecture matters more than the base LLM, • planning, orchestration, memory, and verification dominate quality, • the model itself is increasingly a replaceable component. In my own code factory, I can swap GPT, Claude, or Gemini with minimal impact. There are differences in latency and edge cases, but they no longer define the system. The next real breakthroughs will not come from GPT-6. They will come from new architectures built around models, not inside them. LLMs are becoming infrastructure. Systems are becoming the product.

1

The Temporal Context Trap: Why Naive RAG is a Clinical Liability Vector databases are the current "gold rush" in Healthcare AI, but most implementations ignore a fundamental clinical reality: Medical data is a stream, not a static lake. In a standard RAG pipeline, the retrieval mechanism treats a 2019 oncology report and a 2025 post-operative summary with equal semantic weight if they share similar keywords. Research in longitudinal EHR-RAG confirms that embedding-based similarity tends to overweight repeated historical patterns while underweighting critical state changes. If your system suggests a treatment plan based on a "highly similar" document that is five years out of date, the model hasn't necessarily hallucinated. Your engineering pipeline has failed to model temporal relevance. The Engineering Reality: Similarity Is Not Relevance. In healthcare, mathematical proximity in vector space does not equate to clinical utility. We need to move beyond simple cosine similarity. 1. The Decay Problem Standard top-k retrieval is time-blind. Recent work on TempRALM and time-sensitive retrieval suggests that we must adjust retrieval scores by weighting the delta between the current timestamp and the document timestamp. 2. The State-Change Failure Clinical trajectories are defined by transitions—moving from "Active" to "Remission" or "Stable" to "Acute." Naive RAG fails to identify these terminal events, often stuffing the context window with stale, pre-transition data. The Implementation Shift: Building Time-Aware Pipelines To build production-grade AI in this space, we must move from generic OSS templates to specialized healthcare architectures: • Weighted Recency in Re-ranking Implement a scoring function that combines semantic similarity with a decay factor. This forces the system to prioritize the current clinical state while maintaining the ability to "look back" when the query demands longitudinal history. • Temporal Partitioning Instead of one massive index, partition your vector store by encounter epochs or care episodes. This helps ground the LLM in the patient’s most recent "state" without losing the thread of their medical journey. • Hybrid Temporal Reasoning Engineering must distinguish between time-aware indexing (timestamped events) and time-aware reasoning (prompting the LLM to respect chronological order). The Bottom Line Stop focusing solely on the embedding model. If you are an engineering leader in this space, start talking about how your retrieval architecture handles the vector of time. In clinical decision support, the right answer at the wrong time is a safety risk.

4

Interesting paper alert - a commentary in Nature Medicine highlights a pivotal tension in healthcare today: while #AIagents promise transformative impacts — from autonomous triage to multi-modal clinical decision support — regulatory frameworks remain built for a different era (a recurring scheme in the world of Health AI which I do not see going away anytime soon: technical progress in AI will always outpace regulatory actions). Here are the take aways from the paper: 🔹 Current state of regulation of AI agents in health and medicine: Most regulatory pathways, like the EU Medical Device Regulation and the U.S. FDA’s 510(k), are optimized for narrow, static, non-adaptive tools. Approval of AI-driven devices with broad scope and high autonomy remains rare and challenging. 🔹 The problem (why existing regulatory frameworks fall short of capturing agentic AI applications): AI agents orchestrate multiple tools, access sensitive data, and make goal-directed decisions autonomously. Their adaptive nature, coupled with non-deterministic outputs and complex multi-component architectures, pushes them outside the assumptions underpinning today’s medical device regulation. 🔹 Proposed solutions (how can we explore and experiment safely with AI agents in clinical settings in the absence of comprehensive regulatory guidance): The paper explores several emerging regulatory strategies, including: Voluntary alternative pathways → tailored approvals for high-risk AI agents Adaptive oversight → dynamic, real-world performance monitoring Regulatory sandboxes → controlled environments to experiment safely Outcome-driven evaluation → focusing on patient-centered results rather than static technical compliance 💡 The takeaway (so where does this leave us): While regulation is still playing catch-up with the risks and complexities of AI agents in healthcare, it is possible to safely experiment and sandbox their use — provided there are clear guardrails, strong oversight, and rigorous evaluation frameworks. This balanced, sandbox-first approach could enable innovation without compromising patient safety. For first hand insights on how this is all playing out in the real world right now, have a read of these two recent papers from Eric Topol, MD (https://lnkd.in/gEycp__S) and Harsha Nori (https://lnkd.in/gKNxpxsA) John Lim Christopher Bain ZongYuan 宗元 Ge 戈 Deval Mehta #AI #Healthcare #Regulation #AIagents #DigitalHealth #NatureMedicine

136

13 Comments

Digital therapeutics have entered an accountability era. Boardrooms no longer ask if real-world evidence (RWE) is integrated but how reliably it’s governed. This framework defines 12 essential workflows for leadership, from endpoint mapping and data validation to regulatory documentation and audit simulations. The principle is clear: Innovation may open doors, but credibility keeps them open. Strong evidence systems, transparent governance, and continuous validation will define the next generation of trusted digital therapeutics. 🔔 Follow Rizwan Tufail for executive frameworks on building compliant, data-driven, and evidence-ready healthcare organizations.

31

13 Comments

🚀 New GitHub Repositories: Bioinformatics Algorithms from Scratch 🧬💻 I’ve recently published a series of bioinformatics and computational biology repositories implementing core sequence alignment and phylogenetic algorithms from first principles in R. These projects were developed with an emphasis on algorithmic transparency, biological interpretation, and computational complexity, making them useful for both education and research. These projects were completed as apart of my Bio Seq class at GMU for my PhD. 🔬 What’s included: Smith–Waterman (Local Sequence Alignment) and Needleman-Wunsch (Global Sequence Alignment) Dynamic programming implementation with full traceback, scoring matrices, and runtime complexity discussion. UPGMA & WPGMA Distance-based hierarchical clustering methods illustrating molecular clock assumptions and their impact on tree topology. Neighbor-Joining (NJ) A biologically realistic, clock-free phylogenetic method with explicit branch length estimation and Q-matrix construction. 📊 Each repository includes: Clean, well-annotated code Runtime and space complexity analysis GitHub-ready documentation (README.md) Visualizations 🎓 These projects reflect my ongoing work at the intersection of bioinformatics education, computational biology, and reproducible research, and align closely with my teaching and research interests. 🔗 GitHub: https://lnkd.in/eefXHdAg Always happy to connect with others interested in bioinformatics, genomics, phylogenetics, or computational methods in biology. #Bioinformatics #ComputationalBiology #Phylogenetics #Genomics #GitHub #RStats #STEMEducation #ReproducibleResearch #TeachingBioinformatics

31

1 Comment

Carlos Guzman’s #OpenMed walkthrough is a fantastic proof point: bring a healthcare-tuned model to your data, run it at scale, and only pay while it’s running. 1M clinical patient notes, 4 GPU’s, 5m30s, ~$2–3. This is how we vault healthcare forward with AI —by unlocking real use cases at low cost. #Snowflake #CortexAI

12

AI is quietly rewriting the math of clinical trials. In a recent 2026 outlook report, one line jumped out at me: early evidence suggests AI could roughly double clinical development success rates versus historical norms.  That’s not just incremental efficiency; that’s a fundamental shift in how we think about risk, timelines, and capital allocation in drug development. When you look at their phase-by-phase analysis, the impact becomes very real. - Phase I→II success rates are modeled to rise from 52% to 80%. - Phase II→III from 29% to 40%. - Overall probability of getting from first-in-human to approval moves from 7.9% to 17.7%.  In a world where a single failed Phase II can sink a program, those numbers are transformational. What’s driving this? It’s not “AI” in the abstract. It’s AI-native platforms that ingest multi-omic data, patient-level records, and real-world evidence to make better decisions before and during trials: - Smarter target selection to avoid dead ends - Better patient stratification to reduce noise - More precise safety signals earlier in the process The result: more “shots on goal” without blowing up the R&D budget. By compressing discovery timelines and de-risking early phases, AI-native biotechs can iterate faster, validate assets sooner, and recycle capital into the next wave of candidates. That’s why these companies are commanding premium valuations—and why early-stage assets suddenly have some of the best prospective return profiles we’ve seen in years. For sponsors and investors, this changes how you think about clinical trial innovation. It’s no longer just about operational efficiency (faster recruitment, fewer protocol deviations); it’s about intelligence baked into the trial design itself. Site selection, endpoint definition, adaptive trial structures—these are all becoming algorithmically informed decisions. And this is exactly where digital health in clinical trials comes in. Digital biomarkers, remote monitoring, AI-powered eligibility screening, virtual visits: these tools sit on top of the same data and intelligence stack that is lifting success rates. The challenge now is separating signal from noise—identifying which solutions are clinically validated, regulatory-ready, and capable of plugging into this new AI-native trial paradigm. 👉 To go deeper on the digital health side of this story, I strongly recommend reading Galen Growth | Redefining Digital Health Intelligence’s analysis of Digital Health in Clinical Trials—it’s one of the clearest lenses on where technology is truly reshaping trial design, operations, and evidence generation (and where it isn’t yet). Further reading: Galen Growth: Digital Health in Clinical Trials - https://lnkd.in/dEDJ_CPK Data: www.healthtechalpha.com Subscribe: www.galengrowth.com #DigitalHealth #ClinicalTrials #AIinHealthcare

7