Professional Documents
Culture Documents
Bertil Damato MD PhD1, Antonio Eleuteri PhD2, Anthony C Fisher PhD CSci2 Sarah E
From the 1Ocular Oncology Service Royal Liverpool University Hospital, Prescot St,
Pathology, Duncan Building, Royal Liverpool University Hospital, Daulby St, Liverpool
L7 8XP, UK.
Correspondence to: Bertil Damato MD PhD FRCOphth, Ocular Oncology Service, Royal
Liverpool University Hospital, Prescot St, Liverpool L7 8XP, UK; Tel: +44 (0) 151
always involves the liver. It is not known whether treatment of the ocular tumour prevents
metastasis and if so in whom. Most patients with metastatic disease die within a year of the
onset of symptoms and treatment only rarely seems to prolong life significantly. Studies on
As with other cancers, accurate prognostication is required, so that patients with a good
survival probability can be reassured and prevented from unnecessary and prolonged
screening for metastasis. Such intensive care would be reserved for high-risk patients,
thereby conserving resources. More reliable prognostication would also enhance prospects
for randomized studies evaluating whether systemic adjuvant therapy prevents or delays
metastatic disease.
Many factors are known to be predictive of metastatic disease, the most important being:
largest basal tumour diameter; ciliary body involvement; extraocular spread; epithelioid
The 7th edition of the TNM (Tumour, Node Metastasis) staging system for cancer
categorizes uveal melanomas according to: involvement of choroid, ciliary body and iris;
basal tumour diameter; tumour height and extraocular extension. A list of pathological risk
factors recommended for collection is included, together with a system for defining
histological grade according to melanoma cell type. There are no instructions or guidelines
for combining histological data with clinical and genetic predictive factors. As a result,
individuals. Given the choice, the large majority of patients want to know their prognosis,
oncologists also depend on reliable prognostication because they tend to plan systemic
For several years, the authors have been developing methods of integrating pathological,
clinical and genetic data to enhance prognostic accuracy so that estimates of survival
probability are relevant to individual patients. These used neural networks because the
survival predictors correlated with each other and with metastasis in a non-linear fashion.
Two versions of the model were developed: one based only on clinical information; the
other also including histological and genetic factors. These generated two survival curves,
respectively showing the all-cause morality of our patients with uveal melanoma and of the
general population matched for age and sex so that metastatic mortality could be estimated
by comparing one curve with the other. This strategy avoided the need for diagnosis of
unreliable when only biopsy specimens were analysed, as in patients treated with
assessment of extravascular matrix patterns and because the neural networks were unable
to compensate for the missing data. Such missing information has become more common
in recent years because of the way in which prognostic biopsy of irradiated melanomas has
become routine.
The aim of this study was to create a prognostic model that combined pathological, clinical
and genetic data, using imputation techniques to compensate for missing information.
Patients
Patients were selected from the database of the Liverpool Ocular Oncology Centre for the
time period 1984 – 2009 if: (1) diagnosed with uveal melanoma, clinically or
histopathologically; (2) primarily treated by the first author (BD) or an associate at the
Liverpool University Hospital between January 1993 and July 2006; and (3) resident in
mainland Britain. Patients were excluded because of: (1) bilateral melanoma; (2) missing
data regarding basal tumor dimension or anterior tumor extension; (3) iris or ciliary body
tumor not involving choroid; or (4) residence overseas, including Northern Ireland.
Clinical data
et al., 2005] Briefly, the following features were defined by slit-lamp examination and
binocular indirect ophthalmoscopy: coronal tumor location; anterior and posterior tumor
margins; circumferential spread around disk margin, ciliary body, iris and angle; and
centripetal spread towards vitreous. Tumor basal dimensions and thickness were measured
Pathological data
Tumor cell type and mitotic rate were determined by light microscopy using hematoxylin
and eosin staining.[Font et al., 2006] Closed connective tissue loops were identified using
sections stained with periodic-acid Schiff stain without counterstaining and viewed under a
Genetic data
Between 1999 and 2007, chromosome 3 loss and chromosome 8 gains were identified by
fluorescent in-situ hybridization (FISH), which was performed as a clinical service to all
amplification (MLPA).
Survival data
Information on every patient with newly-diagnosed uveal melanoma was sent to the
National Health Service (NHS) Cancer Registry, which automatically informed us of date
and cause of death, usually within three months of this event. Survival probabilities of the
general population were estimated using the UK Government Actuary’s Department (GAD)
Data management
From 1984, clinical information was collected and computerized prospectively, using a
succession of customized databases. The tenets of the Helsinki Declaration were followed.
Institutional ethical committee approval for outcomes analysis was not required.
Mathematical methods
The data were described mathematically using an Accelerated Failure Time model. This
described the relationships between the variables and survival. The formula developed for
this model, linking the survival time T to the vector of prognostic factors X was the
following:
log(T ) = f ( X ) + se (1)
where e is a logistic random variable with scale s . The function f was a linear combination
of prognostic factors. Age and basal tumour diameter were continuous so that complex
introduced to achieve a better fit. With these continuous variables, the data were modeled
splines, which are a class of mathematical models. Mitotic count was treated differently,
being expanded nonlinearly by a combination of linear plus dummy factors (i.e., creating
non-uniform categories). This was necessary because of the sparsity of mitotic counts (e.g.,
few tumours with a count of, say, 5 per 40 high power fields as opposed to many with a
mitotic count of 1 per 40 high power fields). The model was further simplified by fixing a
parametric family for the distribution of events. In other words, survival times were
Two versions of the model were created: one with clinical factors only and the other
including pathological and genetic data. The same mathematical methods were used in both
models.
Missing data
Values for missing data were estimated using the Alternating Conditional Expectations
algorithm, which essentially estimated each of the missing variables as a function of the
other variables. For example, if mitotic count was missing, this was estimated by modelling
its relationships with all the other baseline variables. In other words, if large, epithelioid
melanomas tended to have a higher mitotic count than small, spindle cell melanomas, then
the former kind of tumour would mathematically be given a higher mitotic score than the
latter. Mathematically speaking, this process approximated the joint distribution of the
baseline variables.
Bootstrap re-sampling was then used to estimate the model parameters. In other words, the
entire dataset was repeatedly and randomly split into training and test datasets, and the
model was then fitted to the training data and its performance tested using the test data.
This was done 200 times so that the statistical variability of the performance of the model
could be assessed, using the C index and residuals. Bayesian regularization was applied, to
reduce the chance of model overfitting the data. In other words, the model was simplified
and tailored to the complexity of the data so that predictions could be extrapolated to all
patients and not only relevant to those used to train the model.
Validation
described the ability of the model to rank the outcomes as a function of the prognostic
factors (e.g., checking that tumours with more malignant histological features were indeed
associated with a worse prognosis). Discrimination was expressed in terms of the c-index,
which is a mathematical way of generalising to censored data the diagnostic power of a test
and which is estimated from the area under the receiver operating characteristic (ROC)
curve. As a reminder, the ROC curve describes how sensitivity and specificity change as
the value of a test result is varied. For example, an ROC curve may describe the sensitivity
and specificity with which an increase in basal tumour diameter predicts metastatic death.
Calibration described the precision of the predictions (e.g., checking that approximately
50% of patients with a particular combination of risk factors actually survived five years
when such a survival probability was predicted). Calibration was measured by calculating
the censored prediction residuals, which for a “good” model should follow the log-logistic
Online interface
A web-based decision support system was developed using the Matlab Webserver toolbox
the tool and provided links to our departmental website and Terms and Conditions. On
agreeing to proceed, the user was taken to the data entry screen, which displayed a form for
entry of the following data: age; gender; largest basal tumor diameter; anterior tumor
extension in relation to ora serrata; extraocular spread; years since treatment; epithelioid
cell type; presence of closed loops; mitotic rate; and monosomy 3. Survival curves showed
actuarial survival rates both for patients with choroidal melanoma and for the age-matched
general population (i.e. +/- five years) of the same gender (Figure 1). The patients’ survival
curve was shown with its 95% credibility interval (i.e. the probabilistic equivalent of the
‘frequentist’ 95% confidence interval). Survival was modeled for the first ten post-
operative years, except if cytogenetic data were used, in which case survival was predicted
for only seven years. The facility was provided to adjust the survival estimate if the patient
Cohort
The cohort comprised 3201 patients (1543 female; 1658 male) with a median age of 62.2
years (range, 12.4 – 96.9). The tumours were located in the left eye in 1603 patients
(50.1%) and the right eye in 1598 patients (49.9%). The anterior margin extended anterior
to the ora serrata in 843 eyes (26.3%) and involving the anterior chamber in 195 eyes
(6.1%). Posteriorly, the tumour extended post-equatorially in 2805 eyes 87.6%, involving
optic disc in 482 eyes (15.1%). Extraocular spread was detected in 55 cases (1.7%). The
median largest basal tumour diameter was 12.0 mm (range, 1.2-23.6) and the median
Histological data were available for 1551 tumours, of which 973 (30.4%) contained
epithelioid cells. Closed loops were present in 49% of 960 tumours in which this feature
was assessed and the mitotic count per 40 high power fields exceeded 4 in 29.6% of 1409
tumours in which this examination was possible. Chromosome 3 loss was detected in 295
A total of 1091 (34.1%) of patients had died, the certified cause of death being metastasis
from uveal melanoma in 622 patients. The follow-up time had a median of 5.3 years (range,
0.02 – 25.1), in surviving patients exceeding five years in 1236 patients, ten years in 698
Discrimination was expressed in terms of the c-index, which was 0.75 (95% confidence
interval [CI], 0.74 -0.76) for the clinical model and 0.79 (95% CI, 0.76 - 0.82) for the
cytogenetic model.
Calibration
Calibration followed the log-logistic distribution postulated in equation 1 (Fig. xx). With
both the clinical and cytogenetic models, agreement between ??? and ??? was excellent.
1.0
0.8
Survival Probability
0.6
0.4
0.2
0.0
-12.0 -8.8 -7.2 -5.6 -4.0 -2.4 -0.8 0.8 2.4 4.0
Residual
1.0
0.8
Survival Probability
0.6
0.4
0.2
0.0
Residual
Figure xx shows the predicted and observed survival according to the estimated risk of
metastatic death at five years (i.e., low, medium or high), both for the clinical model (Fig
xxa) and the cytogenetic model (Fig xxb). All pairs of survival curves correlated well with
each other.