Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A tutorial on discovering and quantifying the effect of latent causal sources of multimodal EHR data

Marco Barbero-Mota, Eric V. Strobl, John M. Still, William W. Stead, Thomas A. Lasko

TL;DR

The paper tackles causal effect estimation in large, imperfect multimodal EHR data where latent confounding is pervasive. It presents a pipeline that first preprocesses and time-aligns diverse EHR modalities, then uses FastICA to identify probabilistically independent latent sources $S$ with mixing matrix $A$, situating them as exogenous factors in a linear structural causal model. A causally predictive model $H_c(S)$ is trained to predict outcomes $Y$ from these sources, with sample-level ITEs derived via SHAP on the causal model, offering interpretable, patient-specific causal insights. Real-world validations in lung cancer and SLE demonstrate the method’s scalability and its ability to uncover known etiologies and plausible novel causal signals, while highlighting ongoing challenges in causal discovery evaluation and the focus on interpretability over in-sample predictive gain.

Abstract

We provide an accessible description of a peer-reviewed generalizable causal machine learning pipeline to (i) discover latent causal sources of large-scale electronic health records observations, and (ii) quantify the source causal effects on clinical outcomes. We illustrate how imperfect multimodal clinical data can be processed, decomposed into probabilistic independent latent sources, and used to train taskspecific causal models from which individual causal effects can be estimated. We summarize the findings of the two real-world applications of the approach to date as a demonstration of its versatility and utility for medical discovery at scale.

A tutorial on discovering and quantifying the effect of latent causal sources of multimodal EHR data

TL;DR

The paper tackles causal effect estimation in large, imperfect multimodal EHR data where latent confounding is pervasive. It presents a pipeline that first preprocesses and time-aligns diverse EHR modalities, then uses FastICA to identify probabilistically independent latent sources with mixing matrix , situating them as exogenous factors in a linear structural causal model. A causally predictive model is trained to predict outcomes from these sources, with sample-level ITEs derived via SHAP on the causal model, offering interpretable, patient-specific causal insights. Real-world validations in lung cancer and SLE demonstrate the method’s scalability and its ability to uncover known etiologies and plausible novel causal signals, while highlighting ongoing challenges in causal discovery evaluation and the focus on interpretability over in-sample predictive gain.

Abstract

We provide an accessible description of a peer-reviewed generalizable causal machine learning pipeline to (i) discover latent causal sources of large-scale electronic health records observations, and (ii) quantify the source causal effects on clinical outcomes. We illustrate how imperfect multimodal clinical data can be processed, decomposed into probabilistic independent latent sources, and used to train taskspecific causal models from which individual causal effects can be estimated. We summarize the findings of the two real-world applications of the approach to date as a demonstration of its versatility and utility for medical discovery at scale.
Paper Structure (5 sections, 1 figure)

This paper contains 5 sections, 1 figure.

Table of Contents

  1. Introduction
  2. Preparing EHR data for machine learning
  3. Discovering causal sources and their EHR signatures
  4. Causally predictive models
  5. Real-world examples

Figures (1)

  • Figure 1: Causal ML pipeline summary. Information flows from multimodal episodic EHR observations that are preprocessed for A) independent latent causal sources discovery, to B) task-specific supervised causal models $H_c(S)$. $H_c$ learns to predict $Y$ from the source expressions of a small labeled cohort. Sample-specific Shapley additive explanations $\phi$ estimate individual treatment effects at the $(j,t)$ patient-timepoint level for each of $H_c(S)$$k$ input sources. C) The source inputs to $H_c(S)$ are latent exogenous nodes of the underlying linear SCM causal graph. When active, predictive sources leave a signature in the observed variables and ultimately affect $Y$.