Ligduurvoorspelling voor Ziekenhuismanagement

Providing compassionate, accessible, and high-quality care while also containing cost is the core mission of a hospital. For this to be achieved, all healthcare services must be efficiently planned from admission up to discharge. Effective handling of these services is determinant to the hospital throughput and can be quantified via measures like the Length of inpatient Stay (LoS). The LoS or the duration of stay is simply, the length of time a bed is occupied between admission and discharge, within a single admission episode. This thesis, carried out in collaboration with the University Hospital of Leuven, contributes to the literature by examining the LoS prediction problem from the perspective of a real-world hospital environment. Firstly, this work introduces a unique application of domain adaptation and discriminative learning for LoS prediction across heterogeneous feature spaces to facilitate data accessibility and handle issues related to data heterogeneity across multiple domains. Domain adaptation enables to leverage shared patient information across different hospital units by extracting and transferring learned weights from one hospital unit (the source domain) to another (the target domain). By recognizing the singularity of each unit of admission, this work accommodates domain adaptation, i.e., weights fine-tuning across a heterogeneous feature space between the two domains. Discriminative learning then enables the simultaneous training of the two sets of features (coinciding and non-coinciding features), by updating a feature set weights at a slower rate than the others. In a practical hospital setting with d number of medical departments (each defined by its specialty), the heterogeneous feature space across units, domain adaptation and discriminative learning are the three key elements of this work that can enhance its transferrability and use. Secondly, we address the important problem of explainability in deep learning models by visualizing the latent spaces of sequential models used for predicting hourly remaining Length of Stay (LoS). Through dimensionality reduction techniques, we demonstrate how these models learn patterns in the data over time, revealing that even with just 5-6 hours of early admission data, it is possible to differentiate short-stay patients from those with extended stays who are at high risk of in-hospital mortality. Finally, we illustrate how these prediction systems can be integrated into a clinical decision support system, providing real-time insights that can enhance capacity planning. Finally, this work evaluates the contribution of different data modalities for the continuous prediction of remaining LoS, by utilizing cross-attention mechanisms to extract the latent correlations among these modalities. Our approach handles missing modalities both at the time and the patient levels. The modalities considered include, timestamped thoracic X-rays, timestamped clinical reports and timestamped physiological parameters (such as laboratory results and vital parameters), along with static admission and demographic information. A time-dependent multimodal fusion is implemented to simulate a hospital environment where information is processed at the charted time. The contributions made in this thesis aim at taking an initial step towards the real-time implementation and adoption of AI-driven clinical decision support systems within healthcare facilities regarding LoS.