At the AI in Life Sciences event in Leiden, organized by the Leiden AI Community, three very different perspectives on applying AI in healthcare were presented. Dr. Siri van der Meijden and Laurens Schinkelshoek introduced PERISCOPE, a platform that uses AI to predict postoperative infections. Dr. Lifeng Han of HealthcareNLP explored how natural language processing can unlock the potential of clinical text. Dr. Sadegh Shahmohammadi of TNO presented a portfolio of applied AI projects ranging from workplace stress monitoring to robotics in hospitals.
Postoperative infections affect up to 25% of surgical patients and can significantly delay recovery, increase readmissions, and burden healthcare systems. Current diagnoses often occur only after seven to nine days, leading to late treatment. Meanwhile, hospitals store large amounts of patient data—sometimes spanning over a decade and covering tens of thousands of surgical cases—that could be used to improve clinical decision-making.
AI model and dashboard
PERISCOPE has developed an AI model that estimates the risk of infection after surgery, presenting both a probability score and a risk category (low, medium, high). The tool integrates directly into the electronic health record (EHR), offering clinicians a consolidated dashboard. This dashboard not only displays infection risk but also highlights key predictive factors per patient and centralizes relevant medical information that would otherwise require navigating multiple EHR screens.
A key challenge in developing the model has been the diversity of hospital data systems. Each hospital uses different codes for procedures, conditions, and medications, requiring a combination of technical and medical expertise to map datasets into comparable categories. To ensure applicability across hospitals, PERISCOPE validated its model on more than 250,000 procedures in hospitals in the Netherlands, Belgium, and Denmark. The team found that local validation and, in some cases, retraining were necessary to maintain accuracy due to differences in patient populations, protocols, and data recording practices.
From model to medical device
Building a clinically usable product required more than AI development. The PERISCOPE team created a CE-certified software application that meets medical device regulations, incorporating explainability features and risk mitigation through software testing. They also conducted comparative studies between model predictions and clinician assessments, finding that the model performed similarly to experienced physicians and better than clinicians in training.
While the technology is ready, practical adoption depends on hospital procurement processes, IT integration, and clinician workflows. The PERISCOPE dashboard is now integrated with hospital EHR systems such as ChipSoft HiX, allowing infection risk predictions to appear directly within clinicians’ standard patient overviews. Upcoming pilots in Dutch hospitals will focus on clinical implementation, workflow integration, and measuring real-world impact on outcomes and costs.
From transformers to clinical applications
Dr. Han introduced the foundations of modern NLP, tracing the development from recurrent neural networks and attention mechanisms to transformer-based models such as BERT and GPT. These models, originally designed for tasks like machine translation, now underpin many clinical NLP applications. For example, while early systems often failed to capture the meaning of idioms across languages, transformer models allow for more nuanced interpretation and generation of medical text.
Electronic health records contain not only structured data but also large amounts of free text, where clinicians describe symptoms, diagnoses, treatments, and patient progress. Extracting information from these notes is a key challenge. Using techniques such as named entity recognition and entity linking, NLP systems can automatically identify relevant medical concepts—diseases, prescriptions, side effects—and connect them to standardized knowledge bases. This allows for the creation of clinical knowledge graphs, which can be used to compare treatments across populations, support decision-making, and detect drug safety issues.
Handling ambiguity and variation
One difficulty in clinical NLP is ambiguity: the same term may refer to different conditions or drugs depending on context. Dr. Han described methods to reduce errors by fine-tuning pre-trained models with domain-specific data and by using ensemble approaches that combine multiple models. These techniques improve accuracy compared to general-purpose systems.
A recurring issue in medical AI is limited access to data due to privacy concerns. Even anonymized health records often cannot be freely shared. To address this, researchers are developing synthetic data that mimics real patient records while preserving clinical relevance. Synthetic data can help train AI models without exposing sensitive patient information and also alleviate data scarcity in specialized medical domains.
Towards patient-friendly communication
Beyond supporting clinicians, NLP also has potential to help patients better understand their own medical records. Summarization and simplification tools can translate long and complex reports into accessible language, empowering patients to participate more actively in treatment decisions. Translation tools can further help patients who do not speak the local language.
Dr. Han emphasized that clinical NLP is not only about building accurate models, but also about integrating them into healthcare workflows. Applications range from coding and drug safety checks to patient engagement and decision support. With ongoing work on entity extraction, relation modeling, and synthetic data generation, the field is moving toward practical tools that can reduce administrative burden for clinicians while improving safety and transparency in healthcare.
From research to product-oriented development
Two years ago, TNO established an AI Lab for Health & Work, designed to move from short-term research projects toward product-oriented development. The goal is to create demonstrators that can be spun out into companies or licensed to partners. In the past year, this approach has already led to a spin-off in pathology, while several other projects are underway.
One of the most challenging and ambitious initiatives is the development of an AI model to support early detection of Duchenne muscular dystrophy and other childhood movement disorders. Duchenne is a rare and fatal disease, and while it cannot be cured, earlier recognition allows for improved care and quality of life. TNO is working with patient organizations and families to collect short video clips that capture movement patterns, which can then be analyzed by AI models.
Initial experiments suggest that even short recordings can provide clinically useful signals, but the project also highlights the complexity of working with children’s health data. Consent requirements, privacy regulations, and ethical scrutiny have slowed progress. Shahmohammadi noted that although synthetic data can help augment scarce datasets, real-world data collection is still essential, and gaining approval for that process remains an ongoing challenge.
Workplace health and stress monitoring
Beyond rare diseases, TNO is investing in workplace health. In one study, participants used WhatsApp to respond to daily questions about stress and energy levels, either by text or voice. Responses were automatically processed and classified using AI models, creating a fast and accessible way to capture mental health data in daily life. The approach proved effective, with most participants replying within minutes, demonstrating how AI can help address data-collection barriers in sensitive domains such as stress and well-being.
In clinical practice, TNO is experimenting with guideline retrieval tools for child healthcare. By connecting patient records to relevant parts of official guidelines, the system supports clinicians in quickly identifying appropriate actions and escalation paths. Shahmohammadi explained that these kinds of tools must be integrated into existing clinical systems to be useful, echoing the experience of other AI developers who found that products fail when they sit outside the physician’s normal workflow.
Another area of work concerns the safe integration of large language models in healthcare. In collaboration with academic and industry partners, TNO is studying how LLMs can be used in medical education and practice while remaining compliant with privacy and regulatory requirements. Because generative AI is so new, there is little case law to guide regulators or hospital legal teams. As a result, ethics committees often take a cautious approach. Shahmohammadi described this as a daily reality: researchers push for bold experimentation, while legal experts, facing uncertainty, default to risk aversion.
Robotics in the hospital
Robotics also forms part of TNO’s healthcare research agenda. In the CareBus project, TNO is responsible for ensuring the safety of a hospital logistics robot that will transport medicines and supplies between wards. The robot is not humanoid but a functional delivery platform, and TNO plans to use AI and digital twin technology to validate its physical safety before deployment.
TNO’s future strategy involves heavier investment in healthcare. The organization sees opportunities in early disease detection, workplace well-being, clinical decision support, safe application of large language models, and robotics safety. While technical results are promising, he stressed that adoption timelines will ultimately be determined by regulatory approvals, procurement processes, and workflow integration, rather than by algorithmic performance alone.
See also
Leiden AI Community Launches, Bridging Academia and Business
