Deep Magnitude Management of Clinical Code Embeddings to Predict Unplanned Hospital Readmissions

Abstract The rate of unplanned hospital readmissions is a relevant indicator of the quality of care provided. From a financial point of view, unplanned readmissions are costly for patients and healthcare providers. Awareness of unplanned readmissions helps to mitigate the growth of healthcare costs. Several predictive models have been proposed to reduce unplanned hospital readmissions. However, most research on the topic does not consider the time elapsed between visits, when historical medical events are included in said research. Additionally, the focus on frequent medical events such as chronic illnesses that may lead to unplanned readmission is not explicitly captured in these models. The failure to consider elapsed time between visits and frequent or chronic illnesses can lead to a decline in the model's performance. We then introduce a model called 'Deep Magnitude Management' (D2M) that incorporates the above-mentioned aspects. D2M handles sequences of visits according to their corresponding date and incorporates an explicit transfer mechanism that enables it to focus on frequent medical events. To provide effective evidence, we compare the performance of D2M with the state-of-the-art models and conduct various ablation studies using the MIMIC-3 database. Furthermore, we provide a series of charts explaining the results of the predictions.