ESTRO 36 Abstract Book

278 / 1082

S265 ESTRO 36 2017
_______________________________________________________________________________________________
Conclusion
Our findings suggest that PD-L1 may serve as a promising
biomarker for poor prognosis as well as risk stratification
and even therapeutic targets in HNSCC. Further well-
designed studies and long-term follow up are warranted to
verify these results.
PV-0509 Failure type specific prognostic model for
selection of HNSCC patients for experimental
treatments
K. Håkansson1, J.H. Rasmussen2, G.B. Rasmussen1, J.
Friborg1, T.A. Gerds3, S.M. Bentzen4,5, L. Specht1, I.R.
Vogelius1
1 Rigshospitalet- University of Copenhagen, Department
of Oncology- Section of Radiotherapy, Copenhagen,
Denmark
2 Rigshospitalet- University of Copenhagen, Department
of Otorhinolaryngology- Head & Neck Surgery and
Audiology, Copenhagen, Denmark
3 University of Copenhagen, Department of Biostatistics,
Copenhagen, Denmark
4 University of Maryland Greenebaum Cancer Center,
Division of Biostatistics and Bioinformatics, Baltimore,
USA
5 University of Maryland School of Medicine, Department
of Epidemiology and Public Health, Baltimore, USA
Purpose or Objective
Most clinical trials involve simple inclusion/exclusion
criteria without support by prognostic models. Here, we
present a multivariate model on multiple endpoints to
generate an individual risk profile. We then examine the
risk profile of patients actually referred to a dose
escalation trial and patients that would be candidates for
the RTOG 1016 de-intensification trial.
Material and Methods
Data from 600 HNSCC patients receiving intensity-
modulated radiotherapy at our institution from 2005-2012
were retrospectively analyzed. Outcome was time from
start of radiotherapy to the first occurrence of loco-
regional failure (LRF), distant metastasis (DM) or death
with no evidence of disease (death NED), and was
censored in case of event-free at last follow-up. Three
cause-specific Cox models were built using clinical,
functional and morphological imaging input as candidate
predictors, and using a cross validation technique to
reduce the model to the prediction variables included in
Table 1.
Individualized estimates of 3-yr LRF, DM and death NED
were obtained combining the three Cox regression
models1, thus taking competing risks into consideration.
The performance of the risk predictions was quantified by
cause-specific concordance (C)-indices2 (ideal C-index=1,
coin flip C-index=0.5). The risk profiles of patients
referred to an in-house dose escalation study were
examined, as were the risk profiles of those of the 600
patients from model building which fulfilled the published
inclusion criteria for the RTOG 1016 de-intensification
trial.
Results
In the final analysis, 547 patients with complete data were
included. The observed 3-year incidences were: LRF 25%,
DM 10% and death NED 14%. Figure 1a presents a
visualization of the individual risk of all patients (note that
all probabilities add to 100%). The C-indices for the risk
predictions were: LRF: 0.72, DM: 0.67, Death NED: 0.65.
Of the 547 patients, 131 would have met the inclusion
criteria of the RTOG 1016 de-escalation trial. The risk
profiles of these patients (Figure 1b) show that 27 (21%)
of them had an estimated risk of failure (LRF and DM)
exceeding 20%. Of the 15 patients included in our local
dose escalation study, 11 had a risk of loco-regional failure
of less than 20% (Figure 1c).
Conclusion
The prediction model performed well for LRF, but death
NED and DM risk were only moderately well predicted.
Using the model to examine the profile of patients that
are candidates for a de-intensification schedule, we
document that several patients at a relatively high risk of
failure could be included.Conversely, our own dose
escalation study included several low risk patients,
despite focusing on p16 negative patients or heavy