Prediction of Clinical Complication Onset using Neural Point Processes

TL;DR

This work reframes adverse clinical event onset prediction as a neural temporal point process problem, aiming to improve interpretability by modeling sequences of continuous-time events preceding onset. It evaluates six state-of-the-art neural TPP models (three RNN-based and three attention-based) across multiple critical-care events using $\lambda_k(t|x_{[0,t)})$-based modeling and negative log-likelihood training, with inference via an end-to-end MLP for next-event type and time. Across eICU-derived datasets for pneumonia, sepsis, cardiac arrest, acute renal failure, respiratory failure, and cardiogenic shock, the results show broadly comparable performance among models, with intensity-free approaches often delivering strong log-likelihood and short-term horizon forecasts (via metrics like optimal transport distance) outperforming long-term forecasts. The study identifies saturation of performance across architectures and suggests future improvements through richer feature spaces (including demographics and biomarkers), few-shot learning for rare sequences, and clinician-guided amber-flag sequence construction to enhance interpretability and predictive utility in critical care.

Abstract

Predicting medical events in advance within critical care settings is paramount for patient outcomes and resource management. Utilizing predictive models, healthcare providers can anticipate issues such as cardiac arrest, sepsis, or respiratory failure before they manifest. Recently, there has been a surge in research focusing on forecasting adverse medical event onsets prior to clinical manifestation using machine learning. However, while these models provide temporal prognostic predictions for the occurrence of a specific adverse event of interest within defined time intervals, their interpretability often remains a challenge. In this work, we explore the applicability of neural temporal point processes in the context of adverse event onset prediction, with the aim of explaining clinical pathways and providing interpretable insights. Our experiments span six state-of-the-art neural point processes and six critical care datasets, each focusing on the onset of distinct adverse events. This work represents a novel application class of neural temporal point processes in event prediction.

Figures (5)

• Figure 1: Illustration of dependency calculation in RNN, CNN, and transformer-based models with the current event (red triangle) and historical events (blue circles).
• Figure 2: Illustration of determining an event stream from a neural TPP. Each event intensity function is a continuous parametric curve determined by the observed history of event occurrences. Event intensity curves kee updating as events occur.
• Figure 3: A sample series of amber flag events (orange) preceding the onset of the adverse event (red) within 24 hours of the adverse event onset.
• Figure 4: Log-likelihoods of validation dataset. Higher log-likelihoods indicate better performance.
• Figure 5: optimal transport distances (OTD), measuring the cost of transforming the predicted sequence to the ground truth. Low OTD values indicate better predictions.