Adaptive time-aware LSTM for predicting and interpreting ICU patient trajectories from irregular data

CISSOKO Mamadou Ben Hamidou

Abstract:

Abstract
In personalized predictive medicine, accurately modeling a patient’s illness and care processes is crucial due to the inherent long-term temporal dependencies. However, Electronic Health Records (EHRs) often consist of episodic and irregularly timed data, stemming from
sporadic hospital admissions, which create unique patterns for each hospital stay. Consequently, constructing a personalized predictive model necessitates careful consideration of these factors to accurately capture the patient’s health journey and assist in clinical
decision-making.

Long Short-Term Memory (LSTM) networks are effective for handling sequential data like EHRs, but they face two significant limitations: the inability to interpret prediction results and to take into account irregular time intervals between consecutive events. To address these limitations, we introduce novel deep dynamic memory neural networks called Multi-Way Adaptive and Adaptive Multi-Way Interpretable Time-Aware LSTM (MWTA- LSTM and AMITA-LSTM) designed for irregularly collected sequential data. The primary objective of both models is to leverage medical records to memorize illness trajectories and care processes, estimate current illness states, and predict future risks, thereby providing a high level of precision and predictive power. To enhance their capabilities, both models extend the standard LSTM in two key ways. Firstly, they incorporate frequency measurement and the most recent observation to enhance personalized predictive modeling of patient illnesses, enabling a more accurate understanding of the patient’s condition. Secondly, they parameterize the cell state to handle irregular timing effectively, utilizing both elapsed times and a frequency-based decay factor.

These enhancements allow the models to comprehend the impact of interventions on the course of illness, facilitating the memorization of illness courses and improving the ability to capture the temporal dynamics of healthcare data, thus accommodating variations and
irregularities in event and observation timing. The effectiveness of our proposed models is validated through empirical experiments
conducted on two real-world clinical datasets and three time series datasets for forecasting. The results demonstrate the superiority of our framework over current state-of-the-art models and other robust baselines. This showcases the potential of our approach in advancing personalized predictive medicine by offering a more accurate and comprehensive method for modeling patient health trajectories, ultimately aiding in more informed clinical decision-making.