Clustering on longitudinal quality‐of‐life measurements using growth mixture models for clinical prognosis: Implementation on <scp>CCTG</scp>/<scp>AGITG CO</scp>.20 trial

Jiahui Zhang, Weili Kong, Pingzhao Hu, Derek J. Jonker, Malcolm J. Moore, Jolie Ringash, Jeremy Shapiro, John Zalcberg, John Simes, Dongsheng Tu, Chris J. O’Callaghan, Geoffrey Liu, Wei XuCancer 2022 · DOI

TL;DR

This study uses Growth Mixture Models (GMM) to identify distinct patient subgroups based on their longitudinal quality-of-life (QoL) trajectories in a colorectal cancer clinical trial. These identified "classes" of patients show significant differences in their clinical characteristics and, crucially, in their overall survival, suggesting GMM can be a powerful tool for clinical prognostication.

What Problem Does This Paper Address?

Understanding how a patient's quality of life (QoL) changes over time during cancer treatment and how these changes relate to survival is crucial for personalized medicine. While many studies have looked at QoL at single time points, few have explored the dynamic, longitudinal nature of QoL and its prognostic implications. The challenge lies in identifying meaningful patterns within this complex, time-dependent data and linking them to patient outcomes. Specifically, the authors aim to uncover previously unobserved patient subpopulations with distinct QoL trajectories and assess if these subgroups have different clinical prognoses.

Methods

The core methodology employed is Growth Mixture Modeling (GMM). GMM is a statistical technique that combines finite mixture analysis with growth modeling. It's an unsupervised learning approach that can identify latent subgroups (classes) of individuals who exhibit similar patterns of change over time.

Here's a breakdown of the key steps:

  • Longitudinal QoL Data: The study utilizes QoL data collected at multiple time points (baseline and every 2 weeks for 16 weeks) from the EORTC QLQ-C30 questionnaire.
  • Clustering with GMM: GMM is applied separately to the experimental and control arms of the CO.20 trial. The model aims to identify the optimal number of distinct QoL trajectory classes for each arm.
  • Model Selection: Bayesian Information Criterion (BIC) and entropy values are used to determine the optimal number of classes and assess the quality of the clustering.
  • Clinical Variable Association: Once classes are identified, the authors compare the clinical characteristics (e.g., age, sex, ECOG performance status) of patients within each class using chi-square and t-tests.
  • Survival Analysis: Kaplan-Meier curves and Cox regression are used to analyze the association between the identified QoL classes and overall survival (OS), both from baseline and using a landmark analysis (starting from 16 weeks). They also explore the relationship between individual QoL parameters (baseline, slope, quadratic) and OS.

What's novel here is the application of GMM to longitudinal QoL data specifically for identifying prognostic subgroups in a clinical trial setting, moving beyond single-time-point analyses.

Datasets

The study uses data from the CO.20 clinical trial. This was a Phase III randomized trial involving 750 chemotherapy-refractory metastatic colorectal cancer (CRC) patients. Patients were randomized to receive either Brivanib + Cetuximab (experimental arm, n=376) or Cetuximab + Placebo (standard arm, n=374). QoL data was collected using the EORTC QLQ-C30 questionnaire at multiple time points.

Key Insights and Ideas

  • Hidden Patient Heterogeneity: GMM successfully identified distinct QoL trajectory "classes" within both treatment arms, revealing that patients are not a monolithic group regarding their QoL experience during treatment.
  • QoL Trajectories Predict Survival: The identified QoL classes were significantly associated with overall survival. This is a crucial finding, as it demonstrates that the pattern of QoL change over time, not just a single QoL score, holds prognostic value.
  • Sex and ECOG as Key Correlates: Sex and ECOG performance status emerged as significant predictors of QoL, influencing the trajectories patients followed. This highlights the importance of considering these factors when interpreting QoL data.
  • Prognostic Power Beyond Single Time Points: The study underscores the limitations of single-time-point QoL assessments and emphasizes the value of longitudinal data and advanced modeling techniques like GMM for uncovering deeper insights into patient prognosis.

Key Results

  • Three QoL Classes Identified: In both the experimental and control arms, GMM identified three distinct patient clusters based on their longitudinal QoL trajectories.
  • Significant Survival Differences: Kaplan-Meier analysis showed statistically significant differences in overall survival between these QoL classes within each treatment arm (log-rank p=0.00058 for the experimental arm, p<0.0001 for the control arm).
  • Prognostic Parameters: In a multivariable analysis of all patients, the baseline QoL, the slope of QoL change, and the quadratic term of QoL all showed significant associations with overall survival (p < 0.001).
  • Clinical Variable Differences: Significant differences in sex and ECOG performance status were observed between the identified QoL classes. For instance, in some classes, females had poorer QoL than males.

Limitations and Future Work

  • Sample Limitations: The study's sample was limited to CRC patients in Australia and Canada, which might affect the generalizability of the findings to other populations or cancer types.
  • Missing Data Handling: While the study used Full Information Maximum Likelihood (FIML) to handle missing QoL data, it acknowledges that in some studies, QoL measures might be missing not at random. Future work could explore methodologies to address other missing data patterns.
  • Time Period for Prediction: The study used 16 weeks of QoL data to predict survival beyond that period. While sensitivity analyses with shorter time periods (8 weeks) yielded consistent results, the authors suggest further exploration of different time windows for QoL data to predict survival.
  • Validation in Other Cohorts: Future research is needed to validate these findings in other independent cohorts to confirm the robustness and generalizability of the GMM-based clustering and prognostic implications.

Summary updated

Abstract

GMM can be used to analyze longitudinal QoL data in cancer studies, by identifying unobserved subpopulations (patient clusters). As demonstrated by CO.20 data, these classes can have important implications, including clinical prognostication.
BibTeX
@inproceedings{zhang2022clustering,
  author = {Jiahui Zhang and Weili Kong and Pingzhao Hu and Derek J. Jonker and Malcolm J. Moore and Jolie Ringash and Jeremy Shapiro and John Zalcberg and John Simes and Dongsheng Tu and Chris J. O’Callaghan and Geoffrey Liu and Wei Xu},
  title = {{Clustering on longitudinal quality‐of‐life measurements using growth mixture models for clinical prognosis: Implementation on <scp>CCTG</scp>/<scp>AGITG CO</scp>.20 trial}},
  booktitle = {Cancer},
  year = {2022},
  doi = {10.1002/cam4.5341},
}
Figure