Causally-informed Deep Learning towards Explainable and Generalizable Outcomes Prediction in Critical Care

TL;DR

This work proposes causally-informed explainable early prediction model, which leverages causal discovery to identify the underlying causal relationships of prediction and thus owns two unique advantages: demonstrating the explicit interpretation of the prediction while exhibiting decent performance when applied to unfamiliar environments.

Abstract

Recent advances in deep learning (DL) have prompted the development of high-performing early warning score (EWS) systems, predicting clinical deteriorations such as acute kidney injury, acute myocardial infarction, or circulatory failure. DL models have proven to be powerful tools for various tasks but come with the cost of lacking interpretability and limited generalizability, hindering their clinical applications. To develop a practical EWS system applicable to various outcomes, we propose causally-informed explainable early prediction model, which leverages causal discovery to identify the underlying causal relationships of prediction and thus owns two unique advantages: demonstrating the explicit interpretation of the prediction while exhibiting decent performance when applied to unfamiliar environments. Benefiting from these features, our approach achieves superior accuracy for 6 different critical deteriorations and achieves better generalizability across different patient groups, compared to various baseline algorithms. Besides, we provide explicit causal pathways to serve as references for assistant clinical diagnosis and potential interventions. The proposed approach enhances the practical application of deep learning in various medical scenarios.

Figures (8)

• Figure 1: The overall architecture of the proposed approach. a, The cDEEP model is learned from a large collection of electronic health records, including dynamic clinical monitoring data and the patients' static demographic variables. During training, cDEEP optimizes the causal graphs and neural networks iteratively to unveil direct causal effects, enhancing the model's generalizability. b, In the inference process, cDEEP enhances the interpretability of deep learning models by providing explicit causal pathways for predictions, enabling clinicians to identify actionable variables throughout the chain. The detailed architecture and network design are provided in Extended Data Fig. \ref{['fig:train']}.
• Figure 1: Detailed architecture of the algorithm.a, Dataset allocation. We split the data from the MIMIC-IV and eICU databases into several subsets: training, validation, and in-distribution testing sets are randomly split from all patients with age $\leq 75$; out-of-distribution testing sets consist of all the patients with age $\geq 76$. $P$ denotes the number of patients in each subset. b, Dynamic prediction scheme. Data from the electronic health record (EHR) are transformed into temporally structured sequences with each time slice being 2 hours. Taking the historical data from the preceding 14 days, the model predicts the patient's risk of developing a specific outcome at each time point in the next 24 hours, thereby generating multiple prediction "samples". Here $S_p$ denotes the number of samples for patient $p$.
• Figure 2: Demonstration of cDEEP's interpretable predictions. Caption continues on the next page.
• Figure 2: Demonstration of cDEEP's interpretable predictions.a, Illustration of the causal probability matrix that shows the probability of each variable at time point $t$ being a direct cause of the outcome. Details of the causal probability matrix are described in Methods. b, Visualization of the causal pathways and controlled direct effect (CDE) values provided by our web-based visualization tool. This tool enables clinicians to interact with the nodes of the causal graph, allowing them to examine the CDE values and causal pathways, thereby clarifying the underlying relationships and the overall impact of each variable on the outcomes. Each circle in the graph signifies a variable, with color coding indicating the polarity of deviation from the average (red for above average and blue for below average), while the size of the circle reflects the magnitude of the deviation. Causal pathways are represented by arrows pointing from the cause to the effect, with the thickness of the arrows corresponding to the CDE values, and the color of the arrows indicating the polarity of the causal effect (red for positive and blue for negative). We deployed this tool on https://cdeep.icu/ and see Supplements C.2 for user guidance.
• Figure 2: Detailed architecture of the algorithm. our algorithm, cDEEP, comprised two main modules---one predicts outcomes and the other decomposes the inference path of the prediction, optimizes the causal graphs (V2O and V2V graphs) and neural networks (outcomes and variable prediction models) iteratively to unveil direct causal relationships, enhancing the model's interpretability and generalizability.

Theorems & Definitions (2)

• Definition 1
• Theorem 1