Electronic health records (EHRs) are a critical tool in healthcare and capture a wide array of patient information that can inform clinical decision-making. However, the sheer volume and complexity of EHR data present challenges for healthcare providers, particularly in fast-paced environments such as intensive care units (ICUs). To address this problem, the automatic summarization of the main problems of patients from daily progress notes can be extremely helpful. Furthermore, by accurately predicting ICU patients’ lengths of stay (LOSs), resource allocation and management can be optimized, allowing for a more efficient flow of patients within the healthcare system. This work proposes a hybrid method to summarize EHR notes and studies the potential of these summaries together with structured data for the prediction of LOSs of ICU patients. Our investigation demonstrates the effectiveness of combining extractive and abstractive summarization techniques with a concept-based method combined with a text-to-text transfer transformer (T5), which shows the most promising results. By integrating the generated summaries and diagnoses with other features, our study contributes to the accurate prediction of LOSs, with a support vector machine emerging as our best-performing classifier with an accuracy of 77.5%, surpassing existing systems and highlighting the potential for optimal allocation of resources within ICUs.
29 October 2025 in Social Sciences & Humanities Open
We introduce a measurement framework for analysing Neural Network learning dynamics through representational geometry. Given hidden representations, we compute a global pairwise cosine-similarity matrix and summarise it with two observables: Q_w (mean within-class similarity) and Q_b (mean between-class similarity). These quantities define a low-dimensional representational state q=(Q_w, Q_b) and a corresponding training-time trajectory, enabling phase-portrait style analysis of learning. Across multiple datasets and architectures, we find that improvements in classification performance are more consistently associated with reductions in between-class similarity than with increases in within-class similarity. This empirical finding motivates the hypothesis that, in many settings, learning trajectories exhibit discrimination-dominant representational change. We provide experiments validating the robustness of the derived metric D_q = Q_w - Q_b and compare its descriptive power to alternative representation-level summaries.