## Overview This case study comes from a presentation by Kolina Tondel at the NLP Summit conference, representing Wroclaw Medical University in Poland. The project is part of a doctoral program in medical sciences and focuses on developing a clinical decision support system (CDSS) for managing neonatal sepsis in intensive care units. The work represents an implementation doctorate program, which is a special format designed to transfer scientific research directly into practical technology implementation in real hospital settings. The project involves a collaborative ecosystem comprising three key partners: Wroclaw Medical University for scientific expertise, the Institute of Mother and Child in Warsaw for clinical settings and validation, and a technology partner (referred to as "Brw Medical University" in the transcript, likely another technical institution) for the technological implementation. This multi-stakeholder approach is notable as it brings together the scientific, clinical, and technical expertise needed to develop healthcare AI systems that can actually be deployed in production. ## The Clinical Problem Sepsis represents a significant global health challenge. The presenter emphasizes that every three seconds somewhere in the world, sepsis contributes to a patient's death, with approximately 3 million deaths annually. For neonatal patients, the challenge is even more complex because premature babies are often born with existing health challenges. The key issues with sepsis management include: - Difficulty in rapid diagnosis due to non-specific symptoms - Rapid disease progression requiring time-sensitive intervention - Overlap with symptoms of other diseases - Long waiting times for definitive test results (24 hours for microbiological blood culture results) - The common use of empirical antibiotic treatment, which European health strategies aim to reduce The presentation frames sepsis management as "a race against time" where making the right diagnosis and decision is a fight for every minute. This creates a compelling use case for AI-powered decision support systems. ## Technical Approach ### Data Challenges in Healthcare The presenter identifies a fundamental challenge that is common across healthcare systems globally: medical data is unstructured, scattered across multiple systems, in varying formats, and the same processes may be named differently across institutions. This lack of standardization makes potentially valuable data difficult to use for AI applications. This is where natural language processing becomes essential. The project uses NLP techniques including concept extraction and keyword recognition to convert unstructured clinical notes and records into structured data formats that can be analyzed by machine learning models. The presenter explicitly mentions that without electronic medical records and proper data processing capabilities, the advancement of AI in medicine would not be possible. ### NLP Pipeline for Data Extraction The system is designed to work with complete patient medical records, not just fragments. Natural language processing tools are used to: - Extract relevant clinical information from unstructured text - Structure the extracted information into analyzable formats - Label data appropriately (e.g., sepsis present vs. healthy patient) - Create properly formatted datasets for machine learning model training The presenter notes that in neonatal intensive care unit settings, data extraction happens in real-time, pulling vital signs, laboratory results, and clinical observations from electronic medical records. This real-time aspect is critical for the system's clinical utility. ### Machine Learning Model Development The project follows a standard but carefully considered ML development pipeline. The presenter emphasizes the importance of data quality with the maxim "garbage in, garbage out," highlighting that significant effort goes into ensuring the first element (data preparation) is reliable and well-prepared. The current status of the project, as described in the presentation, is that they are "in the middle of this process"—after initial testing and in the model-building phase. The presenter acknowledges that collecting data, labeling, and preparing datasets is time-consuming, which reflects the reality of healthcare AI development where data preparation often constitutes the majority of project effort. The machine learning models are trained to identify early symptoms in patients by detecting very small changes in parameters—changes that may indicate trends or potential future events before they become visible to the human eye or detectable by standard medical equipment. ## Integration and Deployment Considerations The presenter outlines the planned implementation approach for the clinical decision support system: - The system analyzes information formally found in patient charts and medical records - NLP tools extract and structure information from complete medical documentation - The AI model provides predictions or recommendations based on labeled datasets - The output serves as additional information for doctors, similar to an additional laboratory test or ultrasound result Importantly, the presenter repeatedly emphasizes that the decision support system does not make decisions on its own or perform any medical procedures independently. This "human-in-the-loop" approach is positioned as collaborative rather than competitive—AI augmenting rather than replacing clinical judgment. This framing is important for clinical acceptance and regulatory compliance in healthcare settings. ## Results and Evidence While the presenter's own project is still in development, they reference published research showing promising results from similar systems: - AI models can detect very small variations in basic laboratory parameters that suggest a patient's future trajectory - Some systems have demonstrated the ability to reduce diagnosis time from 24 hours (standard microbiological testing) to 2 hours - High sensitivity and specificity levels have been achieved in referenced publications - Pattern recognition through AI algorithms shows potential for reducing unnecessary antibiotic use The presenter is careful to note that "AI is not better or worse than a doctor"—the goal is collaboration, not replacement. This balanced perspective on AI capabilities in clinical settings reflects a mature understanding of how these systems should be positioned for healthcare deployment. ## Broader Context and Adoption Trends The presentation references growing scientific interest in machine learning and deep learning in recent years, describing it as an "exploding" trend in healthcare research. The presenter also cites a "Health Future Index" report for Poland (also published for other countries) indicating that social attitudes toward AI technology are changing positively, with medical staff increasingly wanting and appreciating technology in their daily work. An interesting observation from the presentation is that for younger medical staff, technology and AI are becoming key elements in career development and workplace choice decisions. This suggests that healthcare providers investing in AI technology may have advantages in recruiting and retaining talent. ## Critical Assessment It's important to note that this case study represents work in progress rather than a fully deployed production system. The project is in the model-building and testing phase, with results yet to be validated in actual clinical practice. The impressive statistics cited (24 hours to 2 hours diagnosis reduction, high sensitivity and specificity) come from other published studies rather than this specific implementation. The collaborative ecosystem approach (scientific, clinical, and technical partners) is a strength that increases the likelihood of successful real-world deployment. However, the challenges of healthcare AI deployment—including regulatory approval, clinical workflow integration, continuous monitoring, and model maintenance—are not extensively discussed in the presentation. The NLP components described appear to focus on traditional NLP techniques (concept extraction, keyword recognition, text structuring) rather than large language models specifically. While valuable for the stated purposes of data extraction and structuring, this represents a more established approach to clinical NLP rather than cutting-edge LLM applications. The machine learning models for sepsis prediction would likely be classification models trained on the structured data rather than generative AI systems. Overall, this case study illustrates the practical challenges and careful approach required when developing AI systems for critical healthcare applications, with appropriate emphasis on data quality, human oversight, and collaborative rather than replacement paradigms for clinical AI.