gms | German Medical Science

64. Jahrestagung der Deutschen Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie e. V. (GMDS)

Deutsche Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie

08. - 11.09.2019, Dortmund

Analysis of Colorectal Cancer Data Using Semiparametric Distributional Regression

Meeting Abstract

  • Alexander Seipp - Carl von Ossietzky Universität Oldenburg, Oldenburg, Germany
  • Verena Uslar - Universitätsklinik für Viszeralchirurgie, Pius-Hospital Oldenburg, Oldenburg, Germany
  • Dirk Weyhe - Universitätsklinik für Viszeralchirurgie, Pius-Hospital Oldenburg, Oldenburg, Germany
  • Antje Timmer - Carl von Ossietzky Universität Oldenburg, Oldenburg, Germany
  • Fabian Otto-Sobotka - Carl von Ossietzky Universität Oldenburg, Oldenburg, Germany

Deutsche Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie. 64. Jahrestagung der Deutschen Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie e.V. (GMDS). Dortmund, 08.-11.09.2019. Düsseldorf: German Medical Science GMS Publishing House; 2019. DocAbstr. 235

doi: 10.3205/19gmds093, urn:nbn:de:0183-19gmds0934

Published: September 6, 2019

© 2019 Seipp et al.
This is an Open Access article distributed under the terms of the Creative Commons Attribution 4.0 License. See license information at http://creativecommons.org/licenses/by/4.0/.


Outline

Text

Introduction: The benefits of postoperative adjuvant chemotherapy for patients with stage III colon cancer have been shown in multiple clinical trials [1]. However, in most of these and similar trials with time-to-event data the focus of the statistical analysis lies on the average patient. Usually, either a ratio of average survival times (Accelerated Failure Time Model) or a ratio of average hazards (Proportional Hazards Model) is estimated. In contrast to this predominant trend, our research question is to ask how the whole conditional distribution of overall survival is changed through chemotherapy and the number of examined lymph nodes. We present results from an analysis of colon cancer patients, registered and treated at the Pius Hospital in Oldenburg. The observational data currently consists of 1028 patients with a median follow up of 745 days and 66.5% censoring.

Methods: Quantiles and expectiles can characterize a distribution. While the 0.5 quantile (median) and the 0.5 expectile (mean) describe the center of the distribution, lower (< 0.5) and upper (> 0.5) quantiles and expectiles characterize tails of the distribution. Distributional regression can, for example, inform us about the impact of covariates on a patient with the in comparison 10% shortest survival time. Additionally, we can analyze the shape of the distribution, inferring risks involved with treatment. While single quantiles have a direct interpretation, we do not have the same intuitions about expectiles. We can, however, calculate the expected shortfall, measuring expected survival time, given the patient survived to a certain threshold [2]. It thus has a similar interpretation as the hazard rate.

Analyzing the colon cancer data, we first use propensity scores to adjust for covariate imbalance, since the data are non-randomized. The logarithm of the survival time is then regressed on multiple covariates, including age, sex, kind of therapy, number of examined lymph nodes and lymph node ratio.

Results: The number of examined lymph nodes has only negligible effects on overall survival. Chemotherapy is positively linked with overall survival. Corresponding coefficients for mean log survival time and median log survival time are 0.68 and 0.73 respectively. The positive impact is higher for the lower tail of the distribution, i.e. shorter survival (0.2 expectile: 1.02, 0.2 quantile: 1.14) than for the upper tail, i.e. longer survival (0.8 expectile: 0.47, 0.8 quantile: 0.23). The expected shortfall equally shows the highest impact of chemotherapy on patients with shorter survival.

Discussion: The results add value to our understanding of colon cancer by providing other information than just average effects. This is done nonparametrically, using no additional assumptions like proportional-hazards or parametric residual assumptions. Treatment and inclusion in the registry is still ongoing and our results are considered preliminary.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.


References

1.
Ragnhammar P, Hafström L, Nygren P, Glimelius B. A Systematic Overview of Chemotherapy Effects in Colorectal Cancer. Acta Oncologica. 2001;40(2-3):282-308. DOI: 10.1080/02841860121543 External link
2.
Taylor JW. Estimating value at risk and expected shortfall using expectiles. Journal of Financial Econometrics. 2008 Feb 9;6(2):231-52. DOI: 10.1093/jjfinec/nbn001 External link