Tamas Madl

Gewetensvraagje: wemelt het in uw processor ook van de valse correlaties? 🤓 Novel technique overcomes spurious correlations problem in AI | TechXplore AI models often rely on "spurious correlations," making decisions based on unimportant and potentially misleading information. Researchers have now discovered these learned spurious correlations can be traced to a very small subset of the training data and have demonstrated a technique that overcomes the problem. The work has been published on the arXiv preprint server. (...) Spurious correlations are generally caused by simplicity bias during AI training. Practitioners use datasets to train AI models to perform specific tasks. For example, an AI model could be trained to identify photographs of dogs. The training dataset would include pictures of dogs where the AI is told a dog is in the photo. During the training process, the AI will begin identifying specific features that it can use to identify dogs. However, if many of the dogs in the photos are wearing collars, and because collars are generally less complex features of a dog than ears or fur, the AI may use collars as a simple way to identify dogs. This is how simplicity bias can cause spurious correlations. https://lnkd.in/esssZ57N #AI #spurious #correlations #simplicity #bias #unimportant #information #learning #machinelearning

1

There was an excellent post this morning shared by Jan Beger on multimodal AI. It was a commentary on a scoping review by Daan Schouten, Geert Litjens et al https://lnkd.in/gBgrP-_i that catalogs technical challenges and clinical applications in a well-written paper. That got me thinking about the definition of multimodal AI in the context of healthcare. While most of emerging literature and conceptual thinking still gravitates towards clinical data and imaging, there is a much broader view of multimodal AI that is imperative here and does not get enough attention yet. My definition of applied multimodal AI in healthcare refers to AI systems that can process, integrate, and reason across diverse, heterogeneous data types—including clinical, behavioral, environmental, socioeconomic, genetic, and self-reported information—to generate a comprehensive, context-aware understanding of an individual's health status and risk trajectory. It is not limited to clinical markers and signals as current agentic approaches allow a much wider address of risk markers across a patient's life journey - something that was hitherto impossible given the width of data this generates. Again, I speak from the perspective of a practitioner and at #CareCentra we have had some early success in synthesizing such multimodal signal data working with partners Intermountain Health and U.S. Department of Veterans Affairs in building such multimodal signal streams using constantly connected, remote care models to both predict and pre-empt exacerbations of risk in patients with a variety of chronic conditions. Thus for a congestive heart failure patient, multimodal signals in the applied world would include the following: Clinical: Ejection fraction, BNP levels, medications Behavioral: Step count trending down, sodium intake increasing Patient-Reported: Mild ankle swelling, reduced appetite Environmental: High heat + humidity index (risk of dehydration) Socioeconomic: Lives alone, limited transportation access Multimodal AI can integrate all this to predict risk of decompensation, trigger early intervention (e.g., virtual visit, medication titration), and recommend lifestyle changes personalized to the patient's motivation and ability. The implications of this broader definition are that: 1. It yields a richer digital phenotype: Moves us toward a more accurate and dynamic understanding of health and risk beyond traditional diagnoses. 2. It addresses continuous, not episodic care: Enables proactive care models that adapt based on real-time or near-real-time patient context. 3. It allows personalization at scale: By incorporating lifestyle, environment, and preferences, AI can tailor interventions and nudges; and 4. It builds towards equity & inclusion: Socioeconomic and environmental data help flag structural barriers to health—important for addressing disparities.

55

7 Comments

I’m excited to share that our scoping review preprint has been published on medRxiv: A Scoping Review of Algorithmic Equity, Data Diversity, and Inclusive Design in the Transformer Era of Clinical NLP https://lnkd.in/eWvWM8h6 Many thanks to Abeer Badawi, Ph.D., Elham Dolatabadi, and Farah Ahmad, MBBS, MPH, PhD, for their collaboration, guidance, and support throughout this work. The rapid digitization of healthcare has positioned transformer-based NLP models as powerful tools for managing clinical text. However, their growing integration into practice raises important and unresolved questions about equity and inclusivity. This review synthesizes 56 studies (2017–2024) and examines how equity is addressed across three dimensions: ➡️ Algorithmic equity — fairness audits are largely post hoc and fragmented ➡️ Data diversity & representativeness — persistent underrepresentation creates what we define as Data Diversity Debt ➡️ Participatory design — observed in only 11% of studies, revealing a major gap beyond clinician involvement. To move beyond descriptive audits, we propose an equity-by-design roadmap to embed fairness, inclusivity, and accountability across the full lifecycle of clinical NLP systems—and to retire Data Diversity Debt. #HealthAI #ClinicalNLP #AlgorithmicEquity #DigitalHealthEquity #AIethics #ResponsibleAI #medRxiv

26

2 Comments

Hidden Bias in AI: LLMs Overextend Research Conclusions #AI #LLM #ScienceLiteracy • 4900 LLM-generated summaries were compared to original studies. [Why important] Demonstrates large-scale evaluation across 10 leading models. • 26–73% of summaries featured conclusions broader than the source. [Why important] Highlights pervasive overgeneralization risk in AI-driven communication. • DeepSeek topped the list with 73% overgeneralizations; no model stayed under 26%. [Why important] Underlines that even the “best” LLM can mislead more than one-quarter of the time. • LLM outputs were 4.85× (95% CI [3.06–7.70], p < 0.001) more prone to broad claims than human-authored summaries. [Why important] Quantifies nearly fivefold higher risk of misinterpretation when relying on AI alone. • 2 mitigation strategies recommended: lower temperature settings & benchmark generalization accuracy. [Why important] Offers concrete steps to curb AI bias and safeguard public understanding. Link: https://lnkd.in/eHEQ4far

1

“An additional limitation of these guidelines is the lack of diversity in the steering groups responsible for developing these AI publishing guidelines, particularly in terms of gender and geographical representation. Demographics analysis show that most members of these steering groups are male, white, and from high-income countries.” 👏🏽Very important research in an era of fast adoption of LLM applications in healthcare, for which more guidelines need to be developed! #LLMs #Healthcare #ResponsibleAI

11

Last week at World Summit AI in Amsterdam, I attended a powerful talk by Karen Hao about the hidden costs of AI. One story that really stayed with me: the invisible human labor behind “clean” datasets. Karen shared accounts of people who spend their days labeling and filtering toxic content to keep it out of training data for large models from companies like OpenAI and Google’s Gemini. The toll is real—constant exposure to harmful material can lead to serious mental health impacts. She also highlighted another often-overlooked cost: the physical footprint of AI. Data centers consume enormous quantities of electricity and water for cooling. In some cities, these demands compete directly with local communities—diverting water and grid capacity, especially during heat waves and droughts. How we can keep innovating while avoiding these harms: ✅ Mandatory mental health support: Paid counseling, decompression time, and rotations to limit exposure windows. ✅ Minimize exposure: Layer automated pre-filters before any human review so people see the least toxic content possible. ✅ Build with Ethics in Mind: Encourage open discussions about the societal impact of AI, involve diverse voices in tech development, and push for policies that protect both people and the environment. #AI #WorldSummitAI #EthicalAI #EmpireOfAI

32

2 Comments

Loooove this post by Veerle Eeftink - van Leemput on Shiny and LLMs! 🤖 https://lnkd.in/dshZtXMd I've also been developing training materials on LLMs in Shiny recently, and it's become apparent very quickly that thinking about user experience is even more important when integrating LLMs into apps 👨‍💻 Some of the thing to consider: ⚡️ Handle requests asynchronously so multiple users don't block each other ⚡️ Provide visual feedback when components are loading ⚡️ Catch errors and display user-friendly messages #rstats #llms #shiny

29

1 Comment

🩺 Explainable AI (XAI) in Healthcare: Turning Black Boxes into Trusted Tools AI is transforming healthcare — predicting disease risk, interpreting scans, and summarizing patient records. But in medicine, accuracy alone isn’t enough. We need trust. That’s where Explainable AI (XAI) steps in — making AI’s decisions transparent so clinicians, patients, and regulators understand why a model made a decision. In healthcare, explainability depends on the type of data: 📊 Tabular, 🖼️ Image, or 📝 Text. 📊 1️⃣ Tabular Data — The Backbone of Health Records EHRs, lab results, and vital signs often live in table format. Models like XGBoost or neural networks can predict readmission or flag high-risk patients. 💡 Why it matters: Clinicians need to know which variables (e.g., age, BP, lab results) drove the decision. 🛠️ Common Methods: Feature Importance (e.g., SHAP values) — ranks influential features. Partial Dependence Plots — show variable impact. Counterfactuals — “What if” scenarios for prediction changes. 🖼️ 2️⃣ Image Data — Seeing What the Model Sees From tumor detection in MRIs to fracture identification in X-rays, CNNs dominate image tasks — but can act like black boxes. 💡 Why it matters: Radiologists must confirm AI focuses on the right regions. 🛠️ Common Methods: Grad-CAM — heatmaps highlighting important areas. Occlusion Tests — test the impact of hiding regions. Integrated Gradients — pixel-level contributions. 📝 3️⃣ Text Data & LLMs — Understanding Clinical Language LLMs power chatbots, note summarizers, and decision-support systems in healthcare. 💡 Why it matters: Outputs must be factual, source-backed, and bias-aware. 🛠️ Common Methods: Attention Visualization — shows which words influenced output. Chain-of-Thought Transparency — reveals reasoning steps. Citation Tracking (RAG) — links answers to original sources. 💡 Final Thought: Explainable AI isn’t just a feature — it’s a requirement for safe, ethical, and compliant AI in healthcare. By opening the black box, we build trust between technology and human decision-makers. 📚 References Samek et al., Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, Springer, 2021. Ghassemi et al., The false hope of current approaches to explainable artificial intelligence in health care, The Lancet Digital Health, 2021. ✨ Thanks to GPT-5 for grammar refinement and paragraph reconstruction — because even AI explanations should be crystal clear! 😄 #AI #XAI #ExplainableAI

5

1 Comment

Just published a guide on synthetic data generation, combining classic tools and GenAI approaches. Built for anyone working with sensitive, hard-to-access datasets, especially in healthcare. Key takeaways: • Synthetic data must be validated — it is not automatically safe • Bias in real data = bias in synthetic data if not addressed • Open-source tools can cover most needs I built this after years of using these methods and seeing how fragmented the resources were. Sharing my open-source repo with the full guide for anyone to use: https://lnkd.in/e7mjts4u

61

3 Comments

💭 “What if we could not only see language drift, but actually predict it?” After countless late nights, that question led us to a patent, a paper, and a milestone I’ll never forget. 🌍 Our work “Koopman Spectral Dynamics for Interpretable and Predictive Semantic Drift in Language Models” has just been accepted at IEEE ICDM 2025 (A* tier). This isn’t just a paper, it comes alongside our first published patent: the world’s first application of Koopman operator theory in NLP. No better venue than IEEE ICDM to share this. Its value will be proven with time. 💡 The problem we tackled: Language models are powerful, but words don’t stay still, their meanings drift. Models capture this drift, but we rarely understand or predict why and how it happens. 🔑 Our solution: We used Koopman operators (a tool from dynamical systems) to model semantic drift as if meanings were “trajectories” evolving in time. This lets us both forecast drift and interpret the hidden forces shaping it. 👦 Satyam’s Explanation: Imagine words are little boats floating on a big river of meaning. Over time, the current shifts: “cloud” once meant sky fluff, then computers. Our method builds a “map of the river” so we can see where each boat is heading, and why it drifts that way. 🙏 Gratitude: Huge thanks to my co-authors Vivian Phung, Vũ Phúc, SHIVAM MISHRA, Sundaram Mishra, and Prof. Dr. Vishwanath Bijalwan. This journey was long, but your teamwork and creativity made it possible. Excited to present this at ICDM 2025 and see how the community builds on it. ✨ This one’s for every sleepless night that finally paid off. #ICDM2025 #AI #MachineLearning #KoopmanOperators #NLP #satmis

23

4 Comments

The Valyu API is the leading search API for scientific research. PubMed is powerful, but limited. Most tools only surface abstracts, leaving out the figures, tables, and full-text detail that actually matter. Valyu has the entire PubMed corpus, updated regularly as new papers are released, and our search retrieves full-text chunks, tables, figures, and equations directly from biomedical papers. Now AI agents can finally access all the research they need 👇

28

8 Comments

I’m at the Dutch AI Conference (AI Coalitie 4 Nederland | AIC4NL ) and couldn’t help but notice Willem Jonker’s opening remark that “data is plutonium.” It’s a powerful image, but let’s be real—not every bit of data deserves to be treated like a ticking time bomb. Yes, sensitive personal info absolutely needs top-tier protection—think nuclear-level security. But anonymous stats, aggregated trends, and synthetic datasets? They’re the fuel for innovation, not radioactive waste. We’re racing toward tougher rules in Europe, but lumping all data together as a nuclear hazard risks slowing down breakthroughs without actually making people safer. I love AIC4NL’s drive to take on the tech giants, but the European players we often highlight are still small fry next to Google, Microsoft, and Amazon. If we really want to compete on a global stage, our regulations need to be smart and balanced: protect privacy and competition without killing creativity. Let’s kick things off with a framework that treats data with the nuance it deserves, that drives innovation, and that helps the Netherlands (and the EU) punch above its weight.

51

4 Comments

NeurIPS 2026 call for papers is live now. Make your AI agents get to work. Papers won't write themselves. Tell them to hurry! /s

1

Why Machine Learning in Biology is HARD At the Thinking About Thinking London conference two weeks ago in Friends House. Thore Graepel in his keynote, gave one of the clearest explanation yet of why building AI for biology isn't just technically challenging … it's fundamentally constrained.   The 𝐁𝐢𝐚𝐬-𝐕𝐚𝐫𝐢𝐚𝐧𝐜𝐞 𝐃𝐢𝐥𝐞𝐦𝐦𝐚 in Biological Models When we try to understand aging, disease, and clinically significant effects via the lenses of machine learning, there is a big trade-off in terms of data and quality. What do I mean? For direct-human health conditions, we're often data constrained.  Privacy, sensitive personal info and siloed systems are part of this issue. Medical info IS quality information that we need but fundamentally, also there is a large variability in between humans that make it difficult for machines to learn.    To overcome this, we can simplify things by going down to the cellular level. We can easily get massive amounts of data points and control the environment well in the lab. But that leads to increased bias (in machine learning terms) For example, even if we know the drug works on a cell - will it still work on humans?     That's the trap! Want to study something relevant to human health? You need complex models. But complex models give you small sample sizes, uncontrollable variables, and massive variance. Your model might be "right" but your confidence intervals are useless. Want statistical power and clean data? Use simple models. But now you're so far from human biology that your findings might not translate at all. Low variance, high bias. It is not as simple as ‘gathering more data’ :(   It's appreciating the landscape and  thinking about how we fundamentally design around these limitations. And work our systems with full knowledge of the variance/biases.   For those building clinical AI... this is why bench-to-bedside is so hard. What are challenges that you had recently in machine learning?

43

2 Comments

🚀 New preprint: Hybrid Code v2 I’m excited to share the latest version of my work: “Hybrid Code v2” (arXiv) 🔗 https://lnkd.in/e7_agH7h This work explores a hybrid paradigm—bringing together complementary computational approaches to address limitations in current AI/ML systems, particularly around reliability, scalability, and real-world deployment. Hybrid methods are increasingly emerging as a practical bridge between theoretical advances and operational systems. Recent research shows that hybrid architectures can improve efficiency, robustness, and performance compared to purely classical approaches (Hugging Face). In this v2 update, I focused on: • Strengthening the core framework   • Improving methodological clarity   • Expanding experimental/implementation insights   • Moving closer to real-world applicability  💡 The broader question: How do we design systems that are not just powerful—but trustworthy, deployable, and aligned with real-world constraints? I’d really value feedback from the community—especially from those working at the intersection of AI, healthcare, and systems design. #AI #MachineLearning #HybridSystems #Research #arXiv #HealthcareAI

6

An important development - how to get an LLM to produce the same result every time? Even if you set temperature = 0, the LLM will NOT generate the same output every single time. Try it 100 times and you will see this. Mira Murati's new company, Thinking Labs, has figured out how to fix this. Mind you this doesn't mean that the LLM will be ACCURATE, it just means that using the techniques her team has uncovered, you could get the LLM to generate the same output, given an identical input, every single time.

8

Excited to share this fascinating article from wim blommaert at ING about our collaboration and scaling The Synthetic Data Vault Enterprise where it is deployed in more than 20 applications and already has been used to improve AI models performance by almost 20% (more on this soon). What I love about this article is how it goes from a broader perspective to concrete examples, proof points and weaves in a personal journey. (Link in the comments) Be it testing software applications or improving AI model performance or for that matter testing deployed AI applications; synthetic data is proving its value in increasing AI adoption. This is a major milestone for us. As everyone recognizes, Generative AI tools and their uptake in Global 2000 is challenging but with right partners and a laser sharp focus on what features will make or break the product, it is doable. Looking forward to the next 100 applications!  Thank you wim blommaert and ING for being such amazing partners. Thankful to our team at DataCebo who day-in day-out keep the focus on the product and pushing the boundaries in making synthetic data a reality. Recently we hit another milestone, more on that coming soon.

35

1 Comment

Many people treat AI prompting as a technical skill. But it’s not. It’s a mirror into your thinking. Most people prompt like they use Google: • Simple questions • Yes or no type response • I ask for X, I expect Y • Lack of context And then they blame the model when the output lacks clarity. But here’s the truth: AI exposes the gaps in your thinking. The ambiguity, the assumptions, and the lack of clarity. Gen AI surfaces all of it. That’s not a flaw. It’s the feature. If writing a great prompt feels hard, it because you need to reframe your thinking. Change your approach. Treat AI as an incredibly fast and intelligent assistant that you can throw ideas at, challenge you to challenge your thinking, and most importantly, augment whatever it is you are trying to do. Next time you go to prompt AI, sit there a little longer. Think a little bit more. Ask questions before seeking answers. Do this, and you’ll be surprised at what is possible.

5

Beyond excited to share our new editorial in BioData Mining. In high-stakes settings like medicine, the most dangerous failure mode is being wrong while sounding absolutely certain. We dig into why today’s alignment methods can inflate confidence, and we highlight practical ways to bring confidence back in line with reliability. Thank you for all your help Jake Patock Asma Nawaz and of course to my brilliant PIs Graciela Gonzalez Hernandez, MS PhD and Nicholas Tatonetti Paper: https://lnkd.in/gAmnjqWY #ClinicalAI #LLMs #AIinHealthcare #TrustworthyAI #ModelCalibration

17

3 Comments