EURURO Vol. 72 No. 3

E68 476

Letter to the Editor
Re: Fatemeh Seyednasrollah, Mehrad Mahmoudian,
Liisa Rautakorpi, et al. How Reliable are Trial-based
Prognostic Models in Real-world Patients with Meta-
static Castration-resistant Prostate Cancer? Eur Urol[1_TD$DIFF].
[5_TD$DIFF]2017;71:838–40
Clinical Utility of Trial-estimated Prognostic Models
Seyednasrollah et al posed a relevant question in their
recent Research Letter. Once shown to be reliable enough in
real-world clinical data, trial-based prognostic models
could provide practical guidance in tailoring treatment
strategies for patients with metastatic castration-resistant
prostate cancer (mCRPC). Unfortunately, the letter missed
several important points about these models and their
proper application to predict overall survival of mCRPC
patients, which may lead to confusion regarding their
practical utility in clinical decision-making, and should
therefore be addressed.
Seyednasrollah et al took the open-source imple-
mentations of the three best-performing models from the
recent prostate cancer DREAM challenge It should be
obvious that the use of others’ models without consulting
the model developers will lead to suboptimal results when
the models are applied to different types of patient data. For
instance, as mentioned in the abstract of the DREAM
challenge paper , aspartate aminotransferase (AST) was
shared by the top-performing models as one of the most
important prognostic biomarkers. However, AST was
missing in this relatively limited real-world data set, along
with many other key markers such as lactate dehydroge-
nase. To give a more balanced view of the models’
performance, it would be important to understand and
acknowledge the effect of the absence of these key
predictors that are rarely measured in real-world prostate
cancer patients.
To deal with the missing clinical variables, Seyednas-
rollah et al applied quite crude procedures that cannot
really compensate for this lack of clinical measurements.
Missing values for many laboratory tests were simply
imputed using median values, even though it is known that
such a procedure performs poorly in heterogeneous data
sets. For instance, in our top-performing prognostic model
(Team 1), we implemented and made available an improved
model-based imputation method that led to the most
accurate predictions in the challenge . It remains unclear
why Seyednasrollah et al chose to use simple median
imputation instead [6_TD$DIFF]in their[7_TD$DIFF] closed-data results . Further
confusion was introduced by simply assuming that all the
patients had bone lesions, even though this information was
actually missing in this real-world patient data set.
The real-world patient cohort used by Seyednasrollah
et al to evaluate the three prognostic models was collected
on the basis of the following inclusion criteria: clinical
diagnosis of prostate carcinoma (ICD10:C61), a prescription
for antiandrogen therapy (ACT code G03HA), and chemo-
therapy with docetaxel as first-line treatment; patients
with other malignancies were excluded, resulting in
289 patients. Compared to the standardized inclusion
criteria of the clinical trials used in the DREAM challenge
to estimate the best-performing prognostic models (histo-
logically or cytologically confirmed prostate adenocarcino-
ma; metastatic disease; progressive disease while receiving
hormonal therapy or after surgical castration; effective
castration), it can be argued that this real-world cohort
represents a rather heterogeneous patient subpopulation,
making it less attractive for the application of mCRPC-
specific prognostic models.
In conclusion, we commend Seyednasrollah et al for
raising this important question and showing that even such
suboptimal application of the trial-estimated prognostic
models led to surprisingly high predictive performance for
these real-world patient data, especially at the clinically
important early time points. However, owing to the above
issues, their letter falls short in fully addressing the question
posed; instead, we are concerned it raises confusion and
more questions regarding the correct application of prog-
nostic models to real-world patient data. Better-designed
studies in much larger and better-characterized mCRPC
cohorts will be needed to reveal robust prognostic factors
and models to help in the treatment of mCRPC patients.
Conflicts of interest: [8_TD$DIFF]Teams [9_TD$DIFF]led [10_TD$DIFF]by [11_TD$DIFF]Tero [12_TD$DIFF]Aittokallio [13_TD$DIFF]and Laura Elo
participated the same prostate cancer DREAM challenge.
available at
journal homepage:
DOI of original article: .
0302-2838/# 2017 European Association of Urology. Published by Elsevier B.V. All rights reserved.