Improving Clinical Trials

🚨 4,609 Studies on LLMs in clinical Medicine. Only 19 were randomized trials. A new Nature Medicine systematic review just mapped the entire evidence landscape for clinical AI -> 4,609 studies identified, but only 1,048 used real patient data and just 19 were RCTs. Most relied on simulated scenarios or exam-style tasks. -> LLMs outperformed humans in only 33% of 1,046 head-to-head comparisons, with performance dropping on more realistic tasks and against more experienced clinicians. -> OpenAI models dominated 65.7% of evaluations, and at least 25% of studies had sample sizes under 30. 👉 We need robust validation through RCTs and real-world evidence before LLMs enter clinical practice. But here's the challenge: with trial and regulatory timelines averaging many years, the model under evaluation may already be obsolete by the time results are published. So how do we generate the rigorous evidence we need for efficacy and safety without falling behind the technology curve? This is the defining question for clinical AI governance right now. How should we adapt our evidence generation and regulatory frameworks to keep pace with generative AI? Ref: Chen et al. LLM-assisted systematic review of large language models in clinical medicine. Nature Medicine 2026

Today FDA announced a real-time, continuous clinical trial monitoring initiative. The details are still emerging (this is an RFI with a broader pilot to follow via RFP) but the directional signal is clear, and it matters. I want to be precise about what’s significant here, because it’s easy to over-interpret early-stage announcements. What actually matters here isn’t about any specific schema or data architecture. It’s about two things that have always been the hardest to deliver in drug development: (1) speed of decision-making and (2) predictability of the regulatory path. Here’s how it works: pre-specified safety and efficacy signals are captured directly from sites and EHR-connected workflows, evaluated algorithmically in near real time, and transmitted continuously to sponsors and FDA (not raw data but the signal that matters to regulatory decision making), which is a meaningful departure from submitting data only after key milestones or trial completion. Prototypes are already active. When sponsors and FDA can establish shared expectations upfront (what success looks like, what triggers a decision, what gets reported and when) the entire development process becomes less adversarial and more legible. You stop losing time to uncertainty. That’s true regardless of the precise implementation details. I spent my time at FDA (and since!) working on how AI can accelerate development and improve internal efficiency, which are adjacent problems. The underlying logic is the same. Essentially, the regulatory system works better when information flows faster and expectations are set earlier. Now at Parexel, I not only think about how sponsors actually operationalize these shifts but can help operationalize them. Through our partnership with Paradigm Health, we’re actively working through what this looks like in practice. Paradigm’s pre-validated eSource infrastructure combined with Parexel’s global trial execution capabilities and sponsor relationships is precisely the combination sponsors will need as the broader pilot opens through an RFP process. And with Boris Braylyan’s operational intensity on our side (he IS warp speed), I know this initiative has the potential to be transformative if implemented correctly. One last though, technology is only part of the story. The real enabler is regulatory expertise combined with deep therapeutic area knowledge, knowing which assets to prioritize for this pilot and what success criteria actually matter in a given indication. I cannot think of any regulatory group more equipped to do that than the one we have under Paul Bridges and Parexel. The organizations that engage with this as an operational question today will be better positioned than those who wait for the final guidance. Congratulations to Jeremy Walsh for getting this out the door, and to Sri Mantha for bringing FDA’s technological infrastructure to a place where technology is no longer the barrier.

Astellas Pharma becomes latest pharma giant to join Evinova's AI platform, following Bristol Myers Squibb and parent AstraZeneca in backing cross-industry clinical trial collaboration >> 🔘 Three major pharma companies are now sharing operational clinical trial data with Evinova's AI platform, marking a rare moment of cross-industry collaboration in drug development 🔘 The platform uses multi-agent AI to tackle one of pharma's most persistent problems: fragmented systems and manual processes that drag out timelines and inflate costs. 🔘 It converts protocols into machine-readable formats and generates optimized study designs in minutes, benchmarked across cost, timelines, patient experience, and even carbon footprint, replacing weeks of manual work. 🔘 A single clinical trial requires over 200 interconnected document types. AI authoring agents now handle intelligent recommendations across regulatory, scientific, and operational inputs, cutting costly protocol amendments 🔘 Early results show 5 to 7 percent savings minimum per study, translating to hundreds of millions of dollars across a top-10 pharma portfolio 🔘 The architecture is modular and cloud-native, letting organizations plug in their own AI models with built-in privacy and regulatory compliance across global markets 💬 The broader signal here: clinical development is finally moving from a document-heavy, siloed process to an AI-first workflow, and the opt-in data sharing model could set a new industry standard for how sponsors learn from each other #digitalhealth #pharma #AI

AI could make clinical trials faster, cheaper, and more inclusive, but success depends on explainability, interoperability, and trust. 1️⃣ 80% of trials face recruitment delays, and 50% of datasets contain quality issues; AI aims to fix both. 2️⃣ Machine learning improves protocol design accuracy (80% vs. 65%) and accelerates site selection and feasibility assessments. 3️⃣ AI tools boost enrollment by up to 65% and cut screening time by 78%, though real-world deployment can be costly and complex. 4️⃣ NLP and digital systems help identify underrepresented groups, supporting more diverse and inclusive recruitment. 5️⃣ AI-driven digital biomarkers enable 90% sensitivity in real-time safety monitoring, improving adverse event detection. 6️⃣ Risk-based monitoring powered by AI detects data integrity issues within 48 hours, much faster than manual reviews. 7️⃣ Predictive models achieve 85-90% accuracy in forecasting outcomes and enable adaptive, personalized trial designs. 8️⃣ High-dimensional, noisy, and heterogeneous data challenge AI systems; success requires strong data harmonization and validation. 9️⃣ Regulatory gaps, stakeholder distrust, and lack of explainability remain major barriers to clinical adoption. 🔟 Real-world trials show AI's promise, but also its high cost, customization demands, and integration hurdles. ✍🏻 David Olawade (MPH, FRSPH, FHEA), Sandra Chinaza Fidelis (RN, BNSc, MSc, MPH), Sheila Marinze, Eghosasere Egbon, Ayodele Osunmakinde, Augustus Osborne. Artificial intelligence in clinical trials: A comprehensive review of opportunities, challenges, and future directions. International Journal of Medical Informatics. 2026. DOI: 10.1016/j.ijmedinf.2025.106141

❌ *Don't* read this if you are looking for yet another "magical AI" story. ✔ *Do* read it if you are interested in understanding how one of the most complex processes in a heavily regulated industry could potentially be accelerated, one step at a time, using #ai Too often, those excited about AI's potential look for a single "game-changing" idea. That might work sometimes, but most industries have complex, multi-stakeholder processes, refined over decades. Using AI to transform them requires painstaking work at the level of individual steps, and then pulling it all together coherently. With #clinicaltrials, this could involve separate AI use cases in * Trial design * Patient selection and recruitment * Monitoring & improving patient compliance with trial conditions * Extracting relevant patient & trial data from numerous sources (like charts and lab results) * Data cleansing * Generating early drafts of clinical trial reports for submission to regulators Each of these in isolation would have incremental value, even if they succeeded. They are also, for the most part, "boring" uses of AI compared to the flashy stuff that makes headlines (e.g., AI-enabled drug discovery). But taken together, they could well start making a substantive difference. https://lnkd.in/dJPHv9_U

Patient‑centricity in healthcare has grown up. And that’s a good thing.   In healthcare and life sciences, we’re moving from engagement to co‑creation.   Patients are no longer being “looped in” late. Co‑creation isn’t an occasional workshop anymore—it’s becoming part of trial‑design muscle memory. When patient input is embedded early, clinical trials see ~25% faster enrollment and significantly fewer late‑stage amendments. Decentralized, patient‑friendly designs are also delivering ~20% higher retention. That’s impact—not intent.   The second shift is equally important: we’ve moved from good intentions to measurable outcomes. Patient experience is now treated as an operational lever. It’s measured, tied to KPIs, and discussed alongside timelines, cost, and risk. That signals true maturity.   The third evolution is how we use technology. We’re seeing a move from digital tools for novelty to responsible AI with purpose—designed to reduce patient burden, not add complexity. Simpler protocols. Smarter scheduling. Better listening to patient signals.   Taken together, this marks a fundamental change in mindset. Patients are being recognized for what they truly are— co‑experts in healthcare design, not just end users.   The question for leaders is no longer why patient partnership matters. It’s how deeply we’re willing to embed it into how we work, decide, and build. #PatientCentricity #PatientExperience

Clinical Research Needs a Reality Check, R3 Is Here Wake-Up Call: The new ICH-GCP R3 guidelines just dropped, and if you’re still running trials like it’s 2010, you’re already behind. R3 demands risk-based approaches, decentralized elements, and true patient-centricity. Yet, the industry keeps dragging its feet. Why? Because disruption is uncomfortable. What Needs to Change, Now: 1. Stop Wasting Time on Outdated Monitoring R3 prioritizes risk-based monitoring (RBM). If you’re still obsessed with 100% SDV, you’re part of the problem (minus some early phase oncology- if you know, you know). Solution: CRAs need to evolve into data-driven strategists. Equip yourself with skills in data analytics and centralized monitoring tools to spot trends before they become risks. Learn to read the signals, screen failure rates, dropout patterns, and query spikes tell a story. CRAs who identify these trends early will be the ones leading trials, not just monitoring them. 2. Decentralized Trials Are the Standard, Not a Nice-to-Have Still forcing patients into endless site visits? R3 says adapt or get left behind. Solution: Break into roles shaping the future: - Decentralized Trial Coordinator - Telehealth Study Manager - Remote Monitoring CRA 3. Patient-Centricity: Less Lip Service, More Action R3 is clear: trials must fit patients, not the other way around. Solution: Target roles like: Patient Engagement Lead, Design protocols around real lives. Your Next Move: Master R3: Knowledge of ICH-GCP R3 guidelines = competitive advantage. Target Future-Proof Roles: RBM specialists, DCT experts, and patient-centric strategists are the future of research. Think Like a Trendspotter: The best CRAs don’t just report data, they predict the next move. The Real Question: Are you disrupting the industry, or waiting to be replaced by those who will?

Doctors in London have used a gene therapy treatment to restore vision in children born with one of the most severe inherited eye conditions, AIPL1-related Leber congenital amaurosis. Before treatment, these children could barely distinguish light from darkness. They were unable to track objects, recognize faces, or interact visually with the world around them. The procedure involved a single, highly precise eye surgery, where surgeons injected a healthy copy of the faulty gene into the back of one eye. This allowed retinal cells to begin functioning more normally. Following treatment, the changes were significant. Some children were able to see shapes, recognize their parents, draw, and even begin reading and writing. Long-term follow-up showed that the treated eye maintained improved vision, while the untreated eye continued to decline, consistent with the natural progression of the condition. This development represents a major step forward in treating genetic forms of childhood blindness, offering the possibility of restoring vision rather than only managing symptoms. Researchers emphasize that the treatment is still being studied to understand how long the effects last and the best timing for intervention, but the early results provide strong evidence that vision loss in certain conditions may be reversible. Source: Clinical gene therapy trials for AIPL1-related Leber congenital amaurosis, Moorfields Eye Hospital / UCL Institute of Ophthalmology Photo: Moorfields Hospital Disclaimer: This content is for informational purposes only and should not be considered medical advice.

🧠 Last month the BMJ celebrated a decade of its patient partnership, but since then something's been bugging me: Why are we still ignoring the most valuable experts in neurodegenerative diseases—the patients themselves? 🤔 While we've seen pockets of patient involvement, traditional research models often overlook the valuable insights patients provide, bringing them in only as research participants or looking for a rubber stamp on studies designed only by scientists. By failing to keep patients at the centre, we're missing a trick. Here's why: 👀 Real-world insights: Patients live with their conditions every day, which means their experiences provide critical insights that can drive more effective user design. Ignoring these voices means missing out on the real experts. 🎯 Improved clinical outcomes: You can get the science and engineering right, but the patient-centricity wrong, and experience gadgets that gather dust in drawers, prescriptions that go unfilled, and trials with high attrition. We have to involve patients *before* that happens to get the right data to improve outcomes 🔨 Respecting patient-developed innovation: Too often "innovation" is a hammer in search of a nail - but many of the smartest "health hacks" I've seen were developed by patients and caregivers - we should be scaling these rather than always building a new app, a new device, or relying only on the peer-reviewed literature when that's not where these things come from or are documented 😇 Ethical imperative: Patients deserve to have a say in the research that affects their lives. Their participation ensures that research is conducted with empathy and respect for their experiences. There are some savvy organisations that understand this, such as LifeArc MND Insights Group, the Co-Design roles available at Parkinson's UK, Rare Dementia Support Champions programme for co-design, and from the other side of the fence the incredible advocates at Genetic ALS & FTD: End the Legacy who have started transforming the way a whole field thinks about those at risk of genetic ALS/FTD. 💬 I'd love to hear your thoughts! You know the drill: Comment below or shoot me a note to discuss. And feel free to share—it's important to me that this message spreads to the people who need to see it. ♻️

🧠 What if doctors could talk to clinical AI tools like a colleague? That’s the vision behind a fascinating new paper out in Frontiers in Artificial Intelligence — and it's one of the most compelling cases yet for how LLMs could transform clinical workflows, not by replacing doctors, but by making AI tools actually usable. Instead of LLMs making risky predictions on their own, this study explores a smarter role: 👉 LLMs as interactive interfaces that guide clinicians through trusted tools like QRisk3, AutoPrognosis, and clinical guidelines — all via natural language. Here’s why this matters: 💬 Usability is the real bottleneck Digital tools fail when they’re clunky. A conversational interface that can explain, adapt, and assist — in real time — is a usability leap forward. 🔍 Trust comes from transparency The LLM doesn't make the call — it pulls from validated models and guidelines, showing its sources along the way. That’s huge for explainability. 📉 Hallucinations? Nearly gone. By grounding responses in external tools and documents, the system answered >99% of clinical questions correctly — vs. ~44–75% for standalone LLMs. 📊 It works — and scales From calculating personalized risk to simulating “what if” scenarios, the system supports deep, patient-specific reasoning — all in plain language. 🔄 From tool → teammate It’s not just about automation. It’s about augmentation — giving clinicians a smarter, more natural way to work with digital tools. This feels like a glimpse of where clinical AI is actually headed: Not flashier algorithms — but better interfaces. 📄 Full paper attached #LLMs #AIinHealthcare #DigitalHealth #ClinicalAI #NLP #HealthTech #ExplainableAI #HumanCenteredAI #MedicalInnovation #MachineLearning