Death comes but why: A multi-task memory-fused prediction for accurate and explainable illness severity in ICUs

World Wide Web https://doi.org/10.1007/s11280-023-01211-w Death comes but why: A multi-task memory-fused prediction for accurate and explainable illness severity in ICUs Weitong Chen1 · Wei Emma Zhang1 · Lin Yue2 Received: 12 February 2023 / Revised: 15 February 2023 / Accepted: 13 September 2023 © The Author(s) 2023 Abstract Predicting the severity of an illness is crucial in intensive care units (ICUs) if a patient‘s life is to be saved. The existing prediction methods often fail to provide sufficient evidence for time-critical decisions required in dynamic and changing ICU environments. In this research, a new method called MM-RNN (multi-task memory-fused recurrent neural network) was developed to predict the severity of illnesses in intensive care units (ICUs). MM-RNN aims to address this issue by not only predicting illness severity but also generating an evidencebased explanation of how the prediction was made. The architecture of MM-RNN consists of task-specific phased LSTMs and a delta memory network that captures asynchronous feature correlations within and between multiple organ systems. The multi-task nature of MM-RNN allows it to provide an evidence-based explanation of its predictions, along with illness severity scores and a heatmap of the patient’s changing condition. The results of comparison with state-of-the-art methods on real-world clinical data show that MM-RNN delivers more accurate predictions of illness severity with the added benefit of providing evidence-based justifications. Keywords Personalized healthcare · Illness severity prediction · Explainable prediction · Time series This article belongs to the Topical Collection: APWeb-WAIM 2022 Guest editors: Calvanese Diego, Toshiyuki Amagasa and Bohan Li B Lin Yue Weitong Chen Wei Emma Zhang 1 The University of Adelaide, Adelaide 5000, South Australia, Australia 2 The University of Newcastle, Newcastle 2308, New South Wales, Australia 123 World Wide Web Figure 1 Two different SOFA trajectories for two ICU patients. According to their SOFA score, Patient ID: 80030 (red) was initially in critical condition but gradually improved and was eventually discharged. In contrast, the condition of Patient ID: 45767 (blue) deteriorated, and they ultimately passed away 1 Introduction The exponential growth of electronic health records (EHRs) has drawn significant interest from the machine learning and data mining communities. With a wealth of information from multiple sources and formats, these EHRs offer a vast dataset for developing evidence-based clinical decision-making tools. For example, the My Health Record System1 stores over 41.9 million EHRs on over 6.4 million patients. Despite criticisms by some that EHRs are often vendor-specific and sometimes limited in scope [1], their sheer size and diversity make them a valuable resource for deep learning technology. This is especially true in the intensive care Unit (ICU), where critical decisions are driven by forecasts of patient outcomes based on pathological and physiological values [1]. As such, deep learning technology has been broadly applied to advance research in ICU decision support, particularly mortality estimation [2] and phenotype analysis [3]. In general, clinical decisions in ICUs are time-critical and highly dependent on physiological data. However, making accurate and rapid decisions in these fastchanging environments without enough real-time information on the severity of a patient‘s illness can be very challenging for clinicians. As a result, numerous scoring systems have been developed and progressively refined to assist with rapid patient assessment. Examples include the sequential organ failure assessment score (SOFA) [4], APACHE II [5], and SAPS II [6]. The produced scores reflect the current clinical condition of a patient based on a set of basic physiological indicators. 1.1 Motivation These scoring systems serve as a simple calculation of a patient’s vital signs at various times but do not provide real-time information for critical decision-making in the ICU. The longer the time between updated information, the less opportunity there is to respond to a deteriorating patient, which is why a continuous monitoring of key indicators such as heart rate is essential. According to Bouch and Thompson [7], an instantaneous scoring system that covers a wider range of indicators is urgently needed to support better decision-making in the ICU. To demonstrate the potential impact of such a system, we present an example of two ICU patients, whose SOFA scores are charted over time to show changes in their condition (as illustrated in Figure 1). If linked to medical interventions, these high-frequency 1 https://myhealthrecord.gov.au 123 World Wide Web SOFA measures could provide insight into the effectiveness of each treatment. This example highlights both the potential and the need for continuous prediction of illness severity scores as a new tool for patient monitoring. Over the years, recurrent neural networks (RNNs) and their variants [8–12], have been explored as deep models for handling time series data, and many have achieved significant results with clinical prediction tasks like mortality risk. Given a sequence of multivariate features, the typical outlook of mortality risk with the prediction techniques of today is about 24 hours -barely enough time for clinicians to intervene. More importantly, short-term mortality risk predictions may have ethical implications. For example, the mortality risk to a patient over the next week may be, say, 80% but the prediction for the next 24 hours may only be 5%. If faced with an unaffordable treatment, many patients and caregivers may choose not to continue with clinical services unbeknownst to the consequences of that decision beyond tomorrow. Thus, continuously predicting the medical trajectory not only offers more detailed information at a finer time granularity but could also help caregivers concentrate on planning effective treatments with better consideration of an illness’s true severity. Despite their solid results to date, learning models have some deficiencies. For instance, they normally treat all multivariate time-series variables as an entire input stream without considering the correlations between the physiological variables. However, human organs are highly correlated to each other and to a patient‘s deterioration. When one or two organs start to malfunction, others tend to follow over a short period. For example, systolic blood pressure is positively correlated with diastolic blood pressure and pulse pressure, whereas diastolic blood pressure is inversely correlated with pulse pressure. Also, a deterioration in the fraction of inspired oxygen can asynchronously affect cerebral blood flow. Thus, exploiting correlations between medical time-series variables can further improve classification performance for ICU prediction tasks. There are few research works that h (...truncated)