Two-step interpretable modeling of Intensive Care Acquired Infections

TL;DR

The paper tackles dynamic prediction of ICU-acquired infections with interpretability by integrating minute-level EHR signals into a landmarking competing risks framework (LM-CR). It introduces a two-step approach: Step 1 uses a convolutional neural network (CNN) to extract a high-frequency risk score $Z_{CNN}$ from 1-minute vital signs, and Step 2 embeds this score within a Deep LM-CR model to produce time-dependent CIFs $F_{j,LM}(t_{hor}|\cdot)$. Explainability is provided via Saliency Map Order Equivalent (SMOE) scale-based saliency maps to link CNN patterns to clinical conditions, with data-driven clustering identifying patterns corresponding to tachycardia, hypotension, desaturation, and hyperventilation. Compared with a full ANN model, the Deep LM-CR offers similar accuracy with greater stability and interpretable covariate effects, achieving a global AUROC around 0.75 and a slightly improved Brier score, thereby enabling earlier and more reliable infection warnings in ICU settings.

Abstract

We present a novel methodology for integrating high resolution longitudinal data with the dynamic prediction capabilities of survival models. The aim is two-fold: to improve the predictive power while maintaining interpretability of the models. To go beyond the black box paradigm of artificial neural networks, we propose a parsimonious and robust semi-parametric approach (i.e., a landmarking competing risks model) that combines routinely collected low-resolution data with predictive features extracted from a convolutional neural network, that was trained on high resolution time-dependent information. We then use saliency maps to analyze and explain the extra predictive power of this model. To illustrate our methodology, we focus on healthcare-associated infections in patients admitted to an intensive care unit.

Figures (16)

• Figure 3: Schematic illustration of the CNN model. The input signal is processed by a convolutional layer (128 filters of size 3). The ReLU function is applied before a max-pooling operator which reduces the size of the features. After each max-pooling layer follows a dropout layer whose dropout rate is 0.25. This sequence of hidden layers is repeated five times. The feature maps are then flattened into an array (flatten layer) and then propagated through a fully-connected layer (dense layer) with a sigmoid activation function.
• Figure 4: Schematic representation of the inclusion of the CNN-based risk score $Z_{\textnormal{CNN}}(t_{LM})$ in the ICU cohort data.
• Figure 5: Competing risks model for ICU-AI.
• Figure 6: Distribution of the CNN risk score at three different landmark points ($t_{LM}^k\in\{ 3,6,8\}$ days), stratified for the cause of failure.
• Figure 7: Correlation plot: CNN risk score vs. the vital signals (averaged in the 24 hours before the landmark).