PALM: Patient-centered Treatment Ranking via Large-scale Multivariate Network Meta-analysis

Rui Duan, Jiayi Tong, Lifeng Lin, Lisa D. Levine, Mary D. Sammel, Joel Stoddard, Tianjing Li, Christopher H. Schmid, Haitao Chu, Yong ChenmedRxiv 2020 · DOI

Abstract

The growing number of available treatment options has led to urgent needs for reliable answers when choosing the best course of treatment for a patient. As it is often infeasible to compare a large number of treatments in a single randomized controlled trial, multivariate network meta-analyses (NMAs) are used to synthesize evidence from existing trials of a subset of the available treatments, where outcomes related to both efficacy and safety are considered simultaneously. However, these large-scale multiple-outcome NMAs have created challenges to existing methods due to the increasingly complexity of the unknown correlation structures between different outcomes and treatment comparisons. In this paper, we proposed a new framework for PAtient-centered treatment ranking via Large-scale Multivariate network meta-analysis, termed as PALM, which includes a parsimonious modeling approach, a fast algorithm for parameter estimation and inference, a novel visualization tool for comparing treatments with multivariate outcomes termed as the star plot, as well as personalized treatment ranking procedures taking into account the individual’s considerations on multiple outcomes. In application to an NMA that compares 14 treatment options for labor induction over five modalities, we provided a comprehensive illustration of the proposed framework and demonstrated its computational efficiency and practicality. Our analysis leads to new insights on comparing these 14 treatment options based on joint inference of multiple outcomes that cannot be obtained from univariate NMAs, and novel visualizations of evidence to support patient-centered clinical decision making.

Introduction

Choosing the right medical treatment is rarely a matter of looking at a single variable. Clinicians and patients must constantly weigh efficacy against safety, balancing the "best" result against the "least bad" side effects. While Network Meta-Analysis (NMA) is the gold standard for comparing multiple treatments across different trials, doing so for many treatments and many outcomes simultaneously—a "multivariate" approach—has historically been a computational nightmare.

In this paper, we introduce PALM (PAtient-centered treatment ranking via Large-scale Multivariate network meta-analysis), a framework designed to make these complex comparisons fast, robust, and visually intuitive. PALM addresses the "curse of dimensionality" in correlation modeling, providing a computationally efficient way to rank numerous treatments across multiple outcomes to support personalized clinical decision-making.

The Challenge of Multivariate NMA

When comparing a wide array of treatments, such as 20 different antidepressants, a standard NMA typically evaluates one outcome at a time (e.g., remission rate). However, a patient also considers weight gain, insomnia, and nausea. A multivariate NMA analyzes these together, capturing the inherent trade-offs.

The primary challenge lies in the "correlation structure." To perform this correctly, researchers must understand how different outcomes are correlated within each study. Unfortunately, individual trials almost never report these correlations, and the number of parameters to estimate grows exponentially as more treatments and outcomes are added. Existing Bayesian methods are often too slow or fail to converge when the network scale increases.

The PALM Framework

The PALM framework proposes three major innovations to solve the scaling problem:

1. Parsimonious Modeling via Pseudolikelihood

Instead of attempting to model the full, complex joint distribution of all outcomes—which requires knowing missing correlations—PALM utilizes a "composite likelihood" approach. By focusing on the marginal distributions of each outcome, the model remains robust to misspecifications of the correlation structure while still providing consistent estimates.

2. Fast Iterative Algorithm

We developed an estimation algorithm that separates the calculation of treatment effects from the calculation of heterogeneity (the variance between studies). This allows the model to converge in seconds, even for networks that would take hours or days using standard Bayesian Markov Chain Monte Carlo (MCMC) methods.

3. The Star Plot

Traditional radar charts are often used to visualize multiple outcomes, but their area changes depending on the order in which outcomes are listed. The PALM Star Plot adds auxiliary axes to ensure the area of the "star" is invariant to the order, providing a mathematically stable "overall" performance score for each treatment.

Application: Labor Induction NMA

We evaluated PALM using a massive real-world dataset comprising 280 randomized controlled trials (RCTs) involving 48,068 women. This dataset compares 14 treatment options for labor induction across 5 clinical outcomes, including C-section risk, neonatal morbidity, and maternal morbidity. Additionally, simulation studies were conducted to test the algorithm’s performance regarding bias, efficiency, and coverage under various conditions.

Key Insights and Results

Computational Efficiency

On a standard computer cluster, the PALM algorithm analyzed the 14-treatment/5-outcome labor induction network in less than 3 seconds. This efficiency is driven by "information orthogonality," where the estimation of population-level effect size is relatively insensitive to the estimation of between-study variance.

Clinical Findings

In the labor induction study, Vaginal PGE2 gel was identified as having the best overall effect across all five outcomes. However, PALM’s personalized ranking procedure allows for a "weighted utility function." If a patient's sole priority is minimizing C-section risk, low-dose misoprostol solution ranks first. This demonstrates PALM's ability to tailor recommendations to individual patient concerns.

Statistical Validity

Simulations confirmed that PALM maintains a 95% coverage probability. This indicates that its confidence intervals are reliable even when the underlying data correlations are complex. While full-likelihood models are theoretically more efficient, the efficiency gain is often marginal compared to the massive gain in computational stability and speed offered by PALM.

Conclusion

PALM provides a robust, fast, and patient-centered approach to multivariate network meta-analysis. While the pseudolikelihood approach may "borrow" less information across outcomes than a full Bayesian model in cases of extremely high, known correlations, its practical utility in large-scale clinical settings is significant. Future work will focus on applying PALM to even larger databases and addressing inherent meta-analysis vulnerabilities such as publication and outcome reporting bias.

Summary updated

Abstract

Abstract The growing number of available treatment options have led to urgent needs for reliable answers when choosing the best course of treatment for a patient. As it is often infeasible to compare a large number of treatments in a single randomized controlled trial, multivariate network meta-analyses (NMAs) are used to synthesize evidence from existing trials of a subset of the available treatments, where outcomes related to both efficacy and safety are considered simultaneously. However, these large-scale multiple-outcome NMAs have created challenges to existing methods due to the increasingly complexity of the unknown correlation structures between different outcomes and treatment comparisons. In this paper, we proposed a new framework for PAtient-centered treatment ranking via Large-scale Multivariate network meta-analysis, termed as PALM, which includes a parsimonious modeling approach, a fast algorithm for parameter estimation and inference, a novel visualization tool for comparing treatments with multivariate outcomes termed as the star plot, as well as personalized treatment ranking procedures taking into account the individual’s considerations on multiple outcomes. In application to an NMA that compares 14 treatment options for labor induction over five modalities, we provided a comprehensive illustration of the proposed framework and demonstrated its computational efficiency and practicality. Our analysis leads to new insights on comparing these 14 treatment options based on joint inference of multiple outcomes that cannot be obtained from univariate NMAs, and novel visualizations of evidence to support patient-centered clinical decision making.
BibTeX
@misc{duan2020palm,
  author = {Rui Duan and Jiayi Tong and Lifeng Lin and Lisa D. Levine and Mary D. Sammel and Joel Stoddard and Tianjing Li and Christopher H. Schmid and Haitao Chu and Yong Chen},
  title = {{PALM: Patient-centered Treatment Ranking via Large-scale Multivariate Network Meta-analysis}},
  year = {2020},
  doi = {10.1101/2020.11.18.20234252},
  howpublished = {medRxiv},
}
Figure