Improving survival models in healthcare: a novel matching approach

Sample size: 1799 publication 10 minutes Evidence: high

Author Information

Author(s): Bertsimas Dimitris, Ning Catherine, Lønning Per Eystein, Baba Hideo, Endo Itaru, Burkhart Richard, Aucejo Federico N, Balzer Felix, Kreis Martin E, Margonis Georgios Antonios

Primary Institution: Massachusetts Institute of Technology

Hypothesis

Our matching algorithm, applied to patient subgroups stratified by prognostic scores, allows to match on not only observed patient characteristics but also latent confounding factors.

Conclusion

The proposed two-stage matching approach significantly improves the prognostic accuracy of Cox regression models in colorectal cancer liver metastases.

Supporting Evidence

Models trained on matched cohorts showed an improvement of up to 20% in bootstrapped C-indices compared to models trained on the entire cohort.
Calibration metrics indicated a 6- to 10-fold improvement in models using the proposed matching approach.
The methodology effectively corrected statistically significant imbalances in nine of the thirteen prognostic variables between treated and untreated groups.

Takeaway

This study shows that by carefully selecting a balanced group of patients, we can make better predictions about their survival.

Methodology

A two-stage matching approach was used to create balanced cohorts of treated and untreated patients, followed by training Cox regression models on these matched groups.