PD-020 - Uplift modeling in selection of patients to either radiotherapy or radiochemotherapy in resectable rectal cancer: reassessment of data from the phas...

Date 04 July 2015
Event WorldGI 2015
Session Posters
Topics Anti-Cancer Agents & Biologic Therapy
Colon Cancer
Rectal Cancer
Surgery and/or Radiotherapy of Cancer
Presenter L. Wyrwicz
Citation Annals of Oncology (2015) 26 (suppl_4): 101-107. 10.1093/annonc/mdv234
Authors L. Wyrwicz1, S. Jaroszewicz2, P. Rzepakowski3, K. Bujko4
  • 1The Maria Sklodowska-Curie Memorial Cancer Center and Institute of Oncology, Warsaw/PL
  • 2Institute of Computer Science Polish Academy of Sciences, Warsaw/PL
  • 3National Institute of Telecommunications, Warsaw/PL
  • 4Maria Sklodowska-Curie Memorial Cancer Center and Institute of Oncology, Warsaw/PL

Abstract

Introduction

Uplift modeling is a machine learning technique which allows for estimation, at the level of individuals, the gain in success rate of a treatment with respect to an alternative treatment. The technique is becoming popular in marketing. For the current analysis we apply an uplift regression technique which directly models the difference between treatment and control success probabilities, but uses the correlation between the difference in two treatment outcomes and general patient prospects to improve the model. Here we apply uplift modelling to clinical data to improve selection of the preoperative modality in resectable T3/T4 or N+ rectal cancer patients.

Methods

The individual patients data from previously reported study were reanalyzed (resectable rectal cancer patients treated preoperatively either with 5x5 Gy radiotherapy with immediate surgery or the conventional radiochemotherapy with delayed surgery; Bujko K et al. Br J Surg. 2006 Oct;93(10):1215–23).

Results

The identified factors contributing to selection of better outcome included sex, age and distance from anal verge. Also post-hoc “need for stoma by decision of surgeon” was significantly related with outcome (OS), but due to obvious reasons it was not included in the prospective model. To assess model effectiveness, we tested the difference between survival probabilities, which would occur taking treatment decisions according to the model and opposite to model's advice. We test difference in survival curves at data cut off point of 42 months (median survival time in the sample) using a chi-squared test. The obtained p-value of 0.006 is highly significant.

Conclusion

Uplift modeling can be applied to clinical data and might be helpful in the selection between alternative methods of treatment with similar outcome. The utility of such technique is similar to multivariate analysis of clinical data, but in general it seems to be more sensitive in small or heterogenic datasets.