A comparison of approaches for combining predictive markers for personalised treatment recommendations.
Pierce, Matthias, Emsley, Richard · Trials · 2021 · DOI
Quick Summary
This study compared three statistical methods for using blood test results (biomarkers) to predict which ME/CFS patients would benefit most from rehabilitation treatment. The researchers tested these methods using computer simulations and real data from a rehabilitation trial, to figure out the best way to combine multiple biomarkers to personalize treatment recommendations for individual patients.
Why It Matters
ME/CFS has significant clinical heterogeneity, and identifying which patients will respond to specific treatments is crucial for improving outcomes. This study directly addresses a key gap in precision medicine for ME/CFS by testing practical statistical methods to combine multiple biomarkers for treatment personalization, using real data from an ME/CFS rehabilitation trial.
Observed Findings
Regression approach outperformed Kraemer's approach across all simulated data-generating scenarios
Modified Kraemer approach improved treatment recommendations except when strong unobserved prognostic biomarkers were present
In the FINE trial example, the regression method indicated only weak improvement under its personalized treatment recommendation algorithm
All three methods were sensitive to misspecification of parametric models
Multiple correlated biomarkers showed evidence of treatment effect heterogeneity in the FINE trial data
Inferred Conclusions
Regression-based methods are more reliable than Kraemer's approach for combining multiple biomarkers to guide personalized treatment recommendations
Prognostic score matching can improve upon standard biomarker combination methods but has limitations when unobserved confounders exist
The moderate performance of all methods on real ME/CFS data suggests that additional biomarkers or alternative modeling approaches may be needed to achieve clinically meaningful personalization
Remaining Questions
Which specific biomarkers in ME/CFS most effectively predict differential treatment response to rehabilitation and other interventions?
What This Study Does Not Prove
This study does not establish which specific biomarkers are causally related to treatment response in ME/CFS, nor does it validate any particular biomarker combination clinically. The methods are theoretical frameworks tested on simulation and existing trial data; they do not prove that personalized recommendations based on these biomarkers would improve patient outcomes in practice.
About the PEM badge: “PEM required” means post-exertional malaise was an explicit required diagnostic criterion for participant inclusion in this study — not that PEM was studied, observed, or discussed. Studies using criteria that do not require PEM (e.g. Fukuda, Oxford) are tagged “PEM not required”. How the atlas works →
How can researchers account for unobserved prognostic factors that may limit the effectiveness of current statistical methods?
Would improved biomarker panels or alternative machine learning approaches outperform these parametric methods for treatment personalization in ME/CFS?
Can any of these methods generate personalized recommendations that would be validated and actionable in clinical practice?