Recent progress in large language models (LLMs) has enabled the automated processing of lengthy documents even without supervised training on a task-specific dataset.Yet, their zero-shot performance in complex tasks as opposed to straightforward information extraction tasks remains suboptimal.One feasible approach for tasks with lengthy, complex input is to first summarize the document and then apply supervised fine-tuning to the summary.However, the summarization process inevitably results in some loss of information.In this study we present a method for processing the summaries of long documents aimed to capture different important aspects of the original document.We hypothesize that LLM summaries generated with different aspect-oriented prompts contain different information signals, and we propose methods to measure these differences.We introduce approaches to effectively integrate signals from these different summaries for supervised training of transformer models.We validate our hypotheses on a high-impact task -30-day readmission prediction from a psychiatric discharge -using real-world data from four hospitals, and show that our proposed method increases the prediction performance for the complex task of predicting patient outcome.

Predicting clinical outcomes such as mortality and length of hospital stay during the initial admission to the intensive care unit (ICU) is vital for enhancing patient survival rates and optimizing hospital resource allocation. Accurate diagnostic information is deeply correlated with mortality and hospitalization duration; however, initial admission data often lack sufficient diagnostic details and may even have discrepancies with discharge diagnoses. Previous studies have leveraged Electronic Health Record (EHR) data, encompassing clinical notes and vital sign monitoring data, to boost the performance of clinical outcomes prediction. Nevertheless, they have often overlooked the relevance between sentence information, resorting to simple averages, and have missed the importance of diagnostic information. To overcome this limitation, we propose the Diagnosis-Enhanced Sentence Attention Model for Clinical Prediction (DESAM-cp), a multimodal neural network. We employed an auxiliary loss to ensure that embeddings of clinical note were informative of the patient's discharge diagnosis information and used sentence-level self-attention to account for the relevance between sentences. Ultimately, clinical note embeddings containing diagnostic information were fused with monitoring data embeddings to predict clinical outcomes. Our experiments demonstrate that the methodology of DESAM-cp outperforms previous research methods. Additionally, through ablation tests, visualization analysis, and attention analysis, we have substantiated that our individual modules function effectively. Our analysis also suggests that our model focuses on medically relevant sentences in relation to diagnosis and clinical outcome predictions and can provide supporting evidence to assist clinicians in making informed decisions regarding diagnoses and treatments. The code for DESAM-cp used in this work is available at https://github.com/a11525/DESAM-cp