When deep learning meets causal inference: a computational framework for drug repurposing from real-world data

Ruoqi Liu, Lai Wei, Ping ZhangarXiv 2020 · OpenReview

TL;DR

This paper introduces a novel computational framework that combines deep learning and causal inference to identify new uses for existing drugs (drug repurposing) using real-world healthcare data. By emulating randomized clinical trials on large-scale claims data, the framework successfully identified six previously unindicated drugs that significantly improve outcomes for coronary artery disease, offering a faster and more reliable path to new treatments.

What Problem Does This Paper Address?

Drug repurposing is a promising avenue for accelerating drug development, as it leverages existing drugs with known safety profiles. However, many current methods rely on pre-clinical data, which doesn't always translate well to human efficacy. While real-world data (RWD) like electronic health records and insurance claims offer a direct view of drug effects in large patient populations, analyzing this data presents significant challenges. RWD is often messy, longitudinal, and prone to confounding factors (variables that influence both treatment assignment and outcomes), making it difficult to establish causal relationships. Existing RWD-based repurposing efforts are often limited in scope and rely on pre-existing hypotheses rather than proactively generating new drug candidates. This paper aims to bridge this gap by developing a robust computational framework that can efficiently and reliably identify potential drug repurposing candidates from RWD by emulating clinical trials and rigorously controlling for confounding.

Methods

The core of the proposed solution is a computational framework that emulates randomized clinical trials (RCTs) using RWD. For each drug candidate, the framework:

  1. Cohort Definition: Identifies "user" cohorts (patients prescribed the drug) and "non-user" cohorts (patients prescribed alternative drugs) from a large claims database.
  2. Feature Extraction: Collects a comprehensive set of features (confounders) for each patient, including demographic information, diagnosis codes, and prescription history.
  3. Causal Inference with Deep Learning: This is the novel part. To address the confounding inherent in RWD, the authors employ a deep learning model to estimate propensity scores. This model uses:
  4. Embedding Module: Converts high-dimensional, sparse input features (diagnoses, prescriptions) into lower-dimensional representations.
  5. Recurrent Neural Network (LSTM) and Attention Mechanism: Processes the temporal sequence of patient visits (diagnoses, prescriptions) to capture evolving health status. An attention mechanism helps the model focus on the most important time points.
  6. Prediction Module: Uses the processed features to predict the probability of a patient receiving the treatment (propensity score).
  7. Weighting and Effect Estimation: The estimated propensity scores are used to compute balancing weights (Inverse Probability of Treatment Weighting - IPTW) to create a pseudo-population where treatment assignment is more balanced. The average treatment effect (ATE) on disease outcomes is then estimated using these weights.
  8. Candidate Selection: Drugs are considered repurposed candidates if they show a statistically significant beneficial effect (negative ATE) and meet a strict feature balancing criterion (unbalanced feature ratio below 2%).

The authors also compare their deep learning-based IPTW (LSTM-IPTW) with a traditional logistic regression-based IPTW (LR-IPTW) to demonstrate the superiority of their approach in balancing covariates and estimating treatment effects.

Datasets

The primary dataset used is the MarketScan claims data (from 2012 to 2017). This is a large-scale, de-identified healthcare claims database containing information from employers, health plans, hospitals, Medicare, and Medicaid programs. It includes individual-level data on healthcare utilization and services. For their case study, they identified approximately 1.18 million Coronary Artery Disease (CAD) patients.

Key Insights and Ideas

  • Bridging the Pre-clinical to Clinical Gap: The framework directly addresses the translational issues of pre-clinical drug repurposing by using real-world human data.
  • Emulating RCTs with RWD: The core innovation is the ability to computationally "emulate" RCTs using observational data, allowing for rigorous causal inference without conducting new trials.
  • Deep Learning for Propensity Score Estimation: Using LSTMs and attention mechanisms to model temporal sequences of patient data for propensity score estimation is a significant advancement over traditional methods like logistic regression, especially for longitudinal RWD. This allows for better control of complex confounding.
  • Proactive Candidate Generation: Unlike previous methods that often rely on existing hypotheses, this framework can proactively generate novel drug repurposing candidates by screening a large number of drugs.
  • Interpretability through Attention: The attention mechanism not only improves model performance but also provides insights into which patient features (covariates) are most influential in determining treatment assignment and outcomes, aiding in understanding the causal pathways.

Key Results

  • The framework was applied to identify repurposing candidates for Coronary Artery Disease (CAD), evaluating 55 drug candidates.
  • After rigorous analysis and balancing, 6 drug candidates were identified as significantly improving CAD outcomes.
  • Crucially, these 6 drugs were not previously indicated for treating CAD according to their drug labels.
  • The proposed LSTM-IPTW model demonstrated superior covariate balancing compared to traditional logistic regression-based IPTW (LR-IPTW), leading to more accurate ATE estimations. This was evidenced by lower unbalanced feature ratios and more stable propensity score distributions.

Limitations and Future Work

  • Hypothesized Confounders: The study relied on a set of hypothesized confounders identified by domain experts. The authors suggest that future work could explore additional potential confounders, such as time elapsed from diagnosis to index date and baseline period outcome values.
  • Drug Combinations: The current framework focuses on single active ingredients. The authors propose extending the work to consider and estimate the effects of drug combinations.
  • Data Limitations: While MarketScan is comprehensive, it has limitations (e.g., incomplete race/ethnicity information). The authors also note that the "control group" is defined by alternative drug prescriptions, which might not perfectly mimic a placebo or no-treatment group in a true RCT.

The authors conclude that their framework successfully identifies drug candidates with beneficial effects that were previously unindicated, and their LSTM-IPTW model offers improved performance over traditional methods for causal inference in RWD.

Summary updated

Abstract

Drug repurposing is an effective strategy to identify new uses for existing drugs, providing the quickest possible transition from bench to bedside. Existing methods for drug repurposing that mainly focus on pre-clinical information may exist translational issues when applied to human beings. Real world data (RWD), such as electronic health records and insurance claims, provide information on large cohorts of users for many drugs. Here we present an efficient and easily-customized framework for generating and testing multiple candidates for drug repurposing using a retrospective analysis of RWDs. Building upon well-established causal inference and deep learning methods, our framework emulates randomized clinical trials for drugs present in a large-scale medical claims database. We demonstrate our framework in a case study of coronary artery disease (CAD) by evaluating the effect of 55 repurposing drug candidates on various disease outcomes. We achieve 6 drug candidates that significantly improve the CAD outcomes but not have been indicated for treating CAD, paving the way for drug repurposing.
BibTeX
@misc{liu2020when,
  author = {Ruoqi Liu and Lai Wei and Ping Zhang},
  title = {{When deep learning meets causal inference: a computational framework for drug repurposing from real-world data}},
  year = {2020},
  eprint = {2007.10152},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2007.10152},
}
Figure