Electronic Health Records (EHRs) contain comprehensive patient information that is often voluminous and challenging to review efficiently. This paper presents a systematic evaluation of multiple methods for compressing patient records into standardized, comparable formats. Four compression approaches are implemented and compared: two template-based methods (structured extraction, extractive key-phrase) and two LLM-based methods (LLM, and hybrid LLM with 8 different models). Using a synthetic cohort of 75 patient records generated with realistic clinical patterns, each method is evaluated on information preservation (diagnosis, medication, allergy, lab value recall, and vital accuracy), compression efficiency, and output quality. Across methods, diagnosis recall ranged from 0.637 to 1.000, with medication and allergy recall consistently exceeding 0.880. In the test setup, the template-based approach yielded the highest compression ratio (7.6×), while the hybrid methods provided the most balanced trade-off between compression and clinical utility. These results suggest that combining structured extraction with LLM-generated summaries can be an effective strategy for scenarios requiring both compact representations and contextual clinical information.

Nursing staff are increasingly confronted with extensive and detailed patient documentation, requiring much time to read through numerous possible care measures. Combined with rising patient loads, this underscores the need for a clearer and more immediately accessible overview of each patient’s situation. Patient-specific care pathway visualisations offer a promising approach to reduce cognitive load, support faster decision-making, and improve situational awareness. This work investigates two Artificial intelligence (AI)-assisted methods for generating such visualisations: (1) simple image generation based on structured textual prompts, and (2) automated code generation that produces graph-based representations of clinical pathways. Using a dataset of synthetic patient profiles and seven defined care pathways, evaluating multiple state-of-the-art foundation models. The results highlight clear differences between models and approaches, particularly in language sensitivity, structural consistency, and the level of detail achievable. Image-based outputs provided visually rich overviews but frequently introduced subtle logical inconsistencies, while code-based methods produced verifiable and structurally coherent pathways yet varied in their ability to preserve contextual and psychosocial information. Together, these findings indicate that AI-assisted visualisation can effectively support—but not yet fully automate—patient-specific pathway generation, and they point toward hybrid solutions that combine visual accessibility with logical robustness.

We analyse how a transformer-based language model learns the rules of chess from text data of recorded games. We show how it is possible to investigate how the model capacity and the available number of training data influence the learning success of a language model with the help of chess-specific metrics. With these metrics, we show that more games used for training in the studied range offers significantly better results for the same training time. However, model size does not show such a clear influence. It is also interesting to observe that the usual evaluation metrics for language models, predictive accuracy and perplexity, give no indication of this here. Further examination of trained models reveals how they store information about board state in the activations of neuron groups, and how the overall sequence of previous moves influences the newly-generated moves.