A prediction model for treatment decisions in high-grade extremity soft-tissue sarcomas: Personalised sarcoma care (PERSARC)

Abstract

FREE ACCESS TO THE ARTICLE (LINK IS VALID TILL SEPTEMBER 27, 2017): https://authors.elsevier.com/a/1VWQJ3QE--9cO9

BACKGROUND.

To support shared decision-making, we developed the first prediction model for patients with primary soft-tissue sarcomas of the extremities (ESTS) which takes into account treatment modalities, including applied radiotherapy (RT) and achieved surgical margins. The PERsonalised SARcoma Care (PERSARC) model, predicts overall survival (OS) and the probability of local recurrence (LR) at 3, 5 and 10 years.

AIM

Development and validation, by internal validation, of the PERSARC prediction model.

METHODS

The cohort used to develop the model consists of 766 ESTS patients who underwent surgery, between 2000 and 2014, at five specialised international sarcoma centres. To assess the effect of prognostic factors on OS and on the cumulative incidence of LR (CILR), a multivariate Cox proportional hazard regression and the Fine and Gray model were estimated. Predictive performance was investigated by using internal cross validation (CV) and calibration. The discriminative ability of the model was determined with the C-index.

RESULTS

Multivariate Cox regression revealed that age and tumour size had a significant effect on OS. More importantly, patients who received RT showed better outcomes, in terms of OS and CILR, than those treated with surgery alone. Internal validation of the model showed good calibration and discrimination, with a C-index of 0.677 and 0.696 for OS and CILR, respectively.

CONCLUSIONS

The PERSARC model is the first to incorporate known clinical risk factors with the use of different treatments and surgical outcome measures. The developed model is internally validated to provide a reliable prediction of post-operative OS and CILR for patients with primary high-grade ESTS.

Full text

A Prediction Model for Treatment decisions in High-Grade Extremity Soft Tissue sarcomas
Personalized Sarcoma Care (PERSARC)
Veroniek M van Praag1±, PhD-candidate orthopaedic surgery and Anja J Rueten-Budde2±, PhD-candidate
mathematics, Lee M Jeys3, Professor orthopaedic oncology, Minna K Laitinen3, Consultant orthopaedic
oncology Rob Pollock4, Consultant Surgical oncology, Will Aston4, Consultant Surgical oncology Jos A van
der Hage5, Consultant orthopaedic oncology PD Sander Dijkstra1, Consultant orthopaedic oncology, Peter C
Ferguson6, Consultant orthopaedic oncology Anthony M Griffin5, research manager, Julie J Willeumier1,
PhD-candidate orthopaedic surgery, Jay S Wunder6, Consultant orthopaedic oncology, Michiel AJ van de
Sande1*, Consultant orthopaedic oncology and Marta Fiocco2,7*, Associate professor in biostatistics
±Shared first authorship, *Shared last authorship
1Department of Orthopaedic Surgery, Leiden University Medical Centre, Albinusdreef 2, 2333 ZA Leiden,
the Netherlands
2Mathematical Institute, Leiden University, Niels Bohrweg 1, 2333 CA Leiden, the Netherlands
3Department of Orthopaedic Surgery, Royal Orthopaedic Hospital, Bristol Road South, Northfield,
Birmingham, B31 2AP, United Kingdom
4Department of Orthopaedic Surgery, Royal National Orthopaedic Hospital, Brockley Hill, Stanmore, HA7
4LP, United Kingdom
5Sarcoma Unit, Netherlands Cancer Institute, Department of Surgery, Postbus 90203 1006
BE Amsterdam, the Netherlands
6University Musculoskeletal Oncology Unit, Mount Sinai Hospital, and Division of Orthopaedics,
Department of Surgery, University of Toronto, 600 University Avenue, Toronto, ON M5G 1X5, Canada
7Department of Medical Statistics and Bioinformatics, Leiden University Medical Centre, Albinusdreef 2,
2333 ZA, Leiden, the Netherlands
Corresponding author:
Michiel AJ van de Sande, MD PhD
Department of Orthopaedic Surgery, LUMC

2
Albinusdreef 2, 2300 RC Leiden, The Netherlands
majvandesande@lumc.nl; +31715263606
Conflict of interest: nothing to declare.
Level of significance: level III

3
ABSTRACT
BACKGROUND: To support shared decision making we developed the first prediction model for patients
with primary soft tissue sarcomas of the extremities (ESTS) which takes into account treatment
modalities, including applied radiotherapy and achieved surgical margins. The PERsonalized SARcoma
Care (PERSARC) model, predicts overall survival and the probability of local recurrence at 3, 5 and 10
years.
AIM: Development and validation, by internal validation, of the PERSARC prediction model.
METHODS: The cohort used to develop the model consists of 766 ESTS patients who underwent surgery,
between 2000-2014, at five specialised international sarcoma centres. To assess the effect of prognostic
factors on overall survival (OS) and on the cumulative incidence of local recurrence (CILR) a multivariate
Cox proportional hazard regression and the Fine and Gray model were estimated. Predictive performance
was investigated by using internal cross validation (CV) and calibration. The discriminative ability of the
model was determined with the C-index.
RESULTS: Multivariate Cox regression revealed that age and tumour size had a significant effect on OS.
More importantly patients who received radiotherapy showed better outcomes, in terms of OS and CILR,
than those treated with surgery alone. Internal validation of the model showed good calibration and
discrimination, with a C-index of 0.677 and 0.696 for OS and CILR respectively.
CONCLUSIONS: The PERSARC model is the first to incorporate known clinical risk factors with the use of
different treatments and surgical outcomes measures. The developed model is internally validated to
provide a reliable prediction of postoperative OS and CILR for patients with primary high-grade ESTS.
Highlights
• PERSARC model gives reliable patient-specific prediction for different treatments.
• Radiotherapy associated with survival and diminished risk of local recurrences.
• Higher age and larger tumour size decreased survival.
• Wider margins and smaller tumour size decreased risk of developing local recurrences.
• The 10-year overall survival rate in grade III patients was 38.5%.

4
INTRODUCTION
Multimodality treatment of high-grade soft tissue sarcomas of the extremities (ESTS) has improved over
the years; however, local recurrence (LR), distant metastasis (DM) and poor survival remain of great
concern.1 Although the effect of several patient related prognostic factors on overall survival (OS) and LR
is well described, the lack of a validated prediction model that includes treatment modalities, complicates
decision-making aimed at balancing oncologic cure and minimizing the risk of disability after treatment.
Factors, such as vascular invasion2, peripheral tumour growth2, tumour size2-5, infiltrative growth2,
necrosis2, site3, adjuvant chemo- and/or radiotherapy6, histological grade3-5 (for fibro- and
liposarcomas7) and histological subtype3,4 have been shown to have a significant impact on survival.
While some studies indicate that the prognostic value of tumour depth2, state at presentation6, tumour
site8 and age8 remains unclear, others found some of these factors to be good predictors of outcome.3-5
The effect of limb sparing surgery and neo-adjuvant chemo- and/or radiotherapy remains debatable.6
Surgical margins have an impact on LR1,4 but no clear association with OS has been established.1,4
In 2003 a prognostic model based on 175 patients with ESTS became available9 and expanded twice.10,11
The first update included patients who were diagnosed at a time (1967) when MRI was not part of
standard care. Prognostic factors included in those studies were tumour size, vascular invasion, necrosis,
grade, peripheral growth, depth and location. Age, gender, recurrence and metastasis, margins and
histology were not included in the model. Callegaro et al. (2016) developed two nomograms for soft
tissue sarcomas of the extremities and trunk using age, tumour size, histological- grade and subtype,
using exclusively patients with macroscopically complete surgical resections.12 In addition, several
models only provide prognosis for OS and DM, whereas others underline the relevance of LR. Willeumier
et al. (2017) underlined the importance of individual prognostication of LR and OS based on different
combinations of surgical margins and possible (neo-) and/or adjuvant therapy, while also taking different
patient- and tumour characteristics into account.13
In order to support shared decision making between patients and physicians, this study aims to develop a
prognostic personalized sarcoma care model (PERSARC) to predict the cumulative incidence of local
recurrence (CILR) and OS for a patient with high-grade ESTS with specific clinical characteristics and
possible treatment modalities at baseline. The prediction model is internally validated by calibration and
discrimination.

5
METHODS
This multicentre study was approved by each of our hospitals’ human subjects review boards.
Study population. The study population included a consecutive series of 838 patients with primary
high-grade ESTS who underwent surgical treatment at one of the five international collaborating
hospitals between January 2001 and December 2014. Due to missing values for 72 patients, 766
individuals were included in development of the PERSARC model. Eligible diagnoses included high-grade
(FNCLCC grade 3) angiosarcoma, malignant peripheral nerve sheath tumour (MPNST), synovial sarcoma,
spindle cell sarcoma, myxofibrosarcoma and (pleomorphic) soft tissue sarcomas not-otherwise-specified
(NOS). Excluded patients include those that were treated without curative intent, had LR or DM within
two months after primary treatment (ruled out by pre-treatment and follow-up staging with lung CT-
scan), had a tumour in their abdomen, thorax, head or neck or received (neo-)adjuvant treatment other
than radio- or chemotherapy.
All collaborating sarcoma centres implemented the ESMO guidelines for soft tissue sarcoma follow-up.14
Patients visited the outpatient clinic for their scheduled clinical and radiographic follow-up: every 3-4
months in the first 2-3 years, then every 6 months and after 5 years yearly. It was common that follow-up
was ended after ten years evidence of no disease.
Study Design. This was a retrospective observational study in which clinical information was gathered
retrospectively (medical records) and by using existing prospective sarcoma databases; (including
documentation of clinic visits, operation reports, histology and radiographic reports). This information
included demographics (centre, patient gender and age at presentation, event and follow-up), tumour
characteristics (presentation, localisation, depth, diameter, histology, grade), treatment characteristics
(goal, time of operation (weeks), resection margin and categorical, type and dose of (neo-)adjuvant
therapy), development of LR and/or DM and last known status. All patients had a minimal follow-up (FU)
of 2 years or experienced an event (LR, DM or death) before. The primary outcome measure was survival,
if the patient was alive at their last documented visit information on the tumour status was gathered.
Secondary outcome measure was local recurrence. Long-term follow-up was obtained through reviewing
documentation of all clinical and radiographic follow-up.
A sarcoma was considered primary if it was previously untreated, a biopsy or whoops excision had been
performed prior to presentation at one of the five contributing specialised sarcoma centres, with no
evidence of metastatic disease. Local recurrence was defined as the presence of viable tumour at the site
of the original tumour bed confirmed by clinical findings, pathological tissue diagnosis or radiological

6
reports more than two months after primary surgery. Distant recurrence was defined by clinical or
radiological evidence of systemic spread of tumour outside the primary tumour bed.
Tumour size was defined as maximum diameter at pathologic analysis. In patients that received
neoadjuvant radiotherapy and/or chemotherapy, tumour size was defined as maximum diameter
measured by computed tomography or magnetic resonance imaging prior to treatment. Surgical margin
was defined as follows: intraleasional for tumour cells present at the margin of the resection specimen
(<0.1 mm), marginal for tumour cells found within 0.1-2 millimetres of the margin and free for tumour
cells found at least 2 mm away from the margin. 1,13,15 Tumour grade was classified as high-grade based
on established criteria Fédération Nationale des Centres de Lutte Contre le Cancer (FNCLCC).
Statistical analysis:
Multivariate Cox regression model. To assess the effect of prognostic factors on OS a multivariate Cox
proportional hazards regression model was used. Predictor variables incorporated in the model were
age, tumour size, depth, histology subtype, surgical margin, and radiotherapy. Initially tumour site, and
tumour presentation were considered; however, previous studies 12 and an initial multivariate analysis
(Wald test p value: tumour site p= 0.818, tumour presentation p= 0.696) showed no strong predictive
value.
Fine and Gray model. To estimate the effect of risk factors on the cumulative incidence of LR (CILR), a
competing risks model, which accounts for the competing event death was used (Appendix 1) 16. After
surgery, a patient may be alive with no evidence of disease (ANED). He may then develop LR or die. The
cumulative incidence function is defined as the probability of the event occurring before a certain time
point. Fine and Gray’s method models the effect of covariates on the cumulative incidence in the presence
of competing events. Subdistribution hazard ratios (sHR) estimate the effect of risk factors on the
probability of event occurrence directly. The same covariates used in the Cox model were considered.
Prediction and validation. Predictions for OS and LR at 3, 5 and 10 years after surgery together with
95% confidence intervals (95%CI), which indicate the uncertainty surrounding the estimates are
provided. To justify their use in clinical practice predictive performance of the prediction models was
assessed internally by using leave-one-out cross validation (CV). CV is a technique to simulate model
performance on new data.
Following van Houwelingen (2000) a prognostic model is defined as a rule to compute probabilities given
relevant covariates and its validity can be argued on the basis of model calibration.

7
Calibration refers to how well predicted probabilities agree with observed probabilities. A common
practice is to group patients from “good” to “bad” prognosis. A model is well calibrated if true and
predicted group probabilities do not differ.
The prediction model can be used to categorize patients based on their prognosis. A patient’s risk factor
information can be summarised into a prognostic index (PI), which is a weighted mean of prognostic
variables, where weights are derived from the prognostic model. Patients with a higher value of PI have a
higher predicted risk. Hence the PI can be used to divide data into 4 equal sized groups: “good prognosis”,
“fairly good prognosis”, “fairly poor prognosis” and “poor prognosis”.
Calibration plots visualize model calibration on a given set of data.17,18 Data are divided into prognostic
groups. At specific time points the groups’ observed outcome (OS or CILR) is plotted against their
predicted outcome. If the points are scattered around the diagonal (x=y) the model is valid without
adjustment. To investigate calibration for data subgroups, one-sampled T-tests are used, where predicted
outcomes were considered the ‘fixed” value and observed outcomes as the evaluated variable.19
Discrimination refers to the ability of the model to assign higher predicted risk to patients who
experience the event earlier compared to those experiencing the event later or not at all. To visualize this
aspect, nonparametric curves are plotted showing the observed outcome (OS or CILR) for different
prognostic groups 20. The spread of the curves indicates how well a model can discriminate. The C-index
quantifies discrimination as the proportion of patient pairs that experience events in the order of risk
predicted.21 It can be adjusted for competing risks18 and can be interpreted as follows: a C-index of 1
means that the model has perfect discrimination and a C-index of 0.5 means that the model predicts just
as well as flipping a coin.22
All statistical analysis was conducted using R software.23 A p-value of 0.05 was defined as statistically
significant.
RESULTS
Table 1 summarises patients’ characteristics at baseline for the included 766 patients from the five
international sarcoma centres. The median follow-up was 71.8 months (95%CI: 67.6-75.9), assessed with
the reverse Kaplan-Meier method. In total 369 patients died and 116 developed a LR. The majority of
patients with a LR died (n=83; 72%). OS was estimated to be equal to 63%, 53%, and 39% at 3, 5 and 10
years respectively. Cumulative incidence of LR was estimated to be equal to 13.3% (95%CI: 10.9-15.8),

8
15.1% (95%CI: 12.4-17.7) and 17.2% (95%CI: 13.9-20.5) at 3, 5 and 10 years respectively. The centre
effect on disease progression was investigated but no significant effect was found.
Table 1. Patient characteristics .
Characteristic
N(%)
Total
766
Age (mean (SD))
58.28 (19.39)
Age (%)
30-60 years
281 (36)
<30 years
82 (11)
>60 years
403 (53)
Sex (%)
male
435 (57)
female
331 (43)
Depth (%)*
Deep
579 (76)
Superficial
134 (17 )
Deep & superficial
53 (7)
Size in cm (mean (SD))
10.06 (6.21)
Extremity (%)
Upper
182 (24)
Lower
584 (76)
Presentation (%)
Primary
622 (81)
Whoops
144 (18)
Histology (%)
Myxofibrosarcoma
238 (31 )
MPNST
91 (12)
Synovial sarcoma
142 (18 )
Spindle cell sarcoma
167 (22)
MFH/UPS
77 (10 )
Other
51 (7)
Margin (%)
0 mm
140 (18 )
0.1-2 mm
343 (45)
>2 mm
283 (37)
Limb sparing (%)

9
No
81 (11)
Yes
685 (89 )
RT (%)
Neoadjuvant
184 (24)
Adjuvant
400 (52)
No RT
182 (24)
MFH/UPS: Malignant Fibrous Histiocytoma/Undifferentiated Pleomorphic Sarcoma; MPNST: malignant peripheral nerve
sheath tumour; RT: radiotherapy.
*Depth: relative to the investing fascia.
Age, tumour size and additional radiotherapy show an independent significant effect on OS (Table 2).
Patients with larger tumours have a significantly increased risk of dying with HR equal to 1.068 (95%CI:
1.052-1.085) for a unit increase of 1 cm. Older age is associated with a higher risk of death with HR equal
to 1.195 (95%CI: 1.116-1.268) for a 10-year increase in age. Note that age and size are included as linear
terms in the model, implying that a ‘k*10’ year change in age and a ‘k’ cm change in size multiply the
hazard by HRk. Surgical margin has a marginally significant effect on OS, with HR equal to 0.786 (95%CI:
0.599-1.033) and 0.711 (95%CI: 0.524-0.964) for margin equal to 0.1-2 mm and >2 mm respectively
(reference category 0 mm). Radiotherapy treatment is associated with a decreased risk of dying
compared to surgery alone with HRs equal to 0.548 (95%CI: 0.399-0.753) and 0.638 (95%CI: 0.486-
0.837) for neoadjuvant and adjuvant radiotherapy respectively.
Table 2. Multivariate Cox model for OS: hazard ratio (HR) along with 95% confidence interval (n = 766).
HR
95%CI
P value
Age
1.195
1.116-1.268
<0.001
Size
1.068
1.052-1.085
<0.001
Depth*
0.377
Deep
1.000
Superficial
0.813
0.591-1.117
Deep and superficial
1.110
0.736-1.674
Histology
0.492
Myxofibrosarcoma
1.000
MPNST
1.422
0.989-2.044
Synovial sarcoma
1.261
0.869-1.831
Spindle cell sarcoma
1.211
0.884-1.661

10
MFH/UPS
1.293
0.890-1.876
Margin
0.080
0 mm
1.000
0.1-2 mm
0.786
0.599-1.033
>2 mm
0.711
0.524-0.964
RT
<0.001
No RT
1.000
Neoadjuvant
0.548
0.399-0.753
Adjuvant
0.638
0.486-0.837
The HR of age corresponds to a unit increase of 10 years and the HR of size corresponds to a unit increase of 1 cm. MFH/UPS:
Malignant Fibrous Histiocytoma/Undifferentiated Pleomorphic Sarcoma; MPNST: malignant peripheral nerve sheath tumour;
RT: radiotherapy. *Depth: relative to the investing fascia.
Figure 1 shows calibration plots for OS at 3, 5 and 10 years. The 3-, 5- and 10-year calibration plots show
points (representing risk groups) scattered close to the diagonal, which is contained in the 95%CIs of the
observed group survival.
Fig. 1: Calibration plots for overall survival
Observed survival obtained using Kaplan-Meier estimator is plotted against predicted survival for patients in 8 equal sized risk
groups identified by their predicted survival at (A) 3 years, (B) 5 years and (C) 10 years, as assessed by cross validation.
A detailed comparison of observed and predicted survival at 3, 5 and 10 years for data subgroups is given
in Table 3. Observed and predicted outcome do not differ significantly; however, for smaller and medium
sized tumours (<5cm, 5cm-10cm) survival is underestimated at 3 and 5, and 10 years respectively.

11
Table 3. Comparing observed and predicted OS, assessed by cross validation, for subgroups of data at 3,
5 and 10 years.
Figure 2 shows good discrimination of the model visualised by the spread of the Kaplan-Meier estimates
(solid lines). Model-based estimates (dotted lines) show the mean predicted survival per group close to
the observed survival, indicating good calibration.
The C-index for overall survival was estimated to be 0.677 (95%CI 0.643-0.701).
n (%)
Predicted
3 years
Observed (s.e.)
Difference (95% CI)
Predicted
5 years
Observed (s.e.)
Difference (95% CI)
Predicted
10 years
Observed (s.e.)
Difference (95% CI)
Age
30-60
281 (36.7)
68.9
70.7 (2.8)
-1.8 (-7.3 to 3.7)
60.2
60.5 (3.1)
-0.3 (-6.4 to 5.8)
46.4
45.4 (4.7)
1.0 ( -8.2 to 10.2)
<30
82 (10.7)
77.8
74.6 (4.9)
3.2 (-6.4 to 12.8)
70.7
68.9 (5.3)
1.8 (-8.6 to 12.2)
58.7
58.3 (6.9)
0.4 (-13.1 to 13.9)
>60
403 (52.6)
54.0
54.9 (2.6)
-0.9 (-6.0 to 4.2)
43.6
44.4 (2.7)
-0.8 (-6.1 to 4.5)
29.1
28.9 (4.1)
0.2 ( -7.8 to 8.2)
Size
<5cm
123 (16.1)
77.2
87.0 (3.1)
-9.8 (-15.9 to -3.7)
69.8
78.2 (4.1)
-8.4 (-16.4 to -0.4)
57.2
57.8 (8.1)
-0.6 (-16.5 to 15.3)
5cm-10cmS
295 (38.5)
69.4
68.7 (2.8)
0.7 ( -4.8 to 6.2)
60.3
59.7 (3.0)
0.6 ( -5.3 to 6.5)
45.8
53.1 (3.6)
-7.3 (-14.4 to -0.2)
>=10cm
348 (45.4)
50.4
49.3 (2.8)
1.1 ( -4.4 to 6.6)
40.0
38.6 (2.8)
1.4 ( -4.1 to 6.9)
26.0
20.7 (3.8)
5.3 ( -2.1 to 12.7)
Depth*
Deep
579 (75.6)
60.9
62.8 (2.1)
-1.9 (-6.0 to 2.2)
51.4
52.7 (2.2)
-1.3 (-5.6 to 3.0)
37.3
37.5 ( 3.1)
-0.2 ( -6.3 to 5.9)
Superficial
134 (17.5)
69.3
67.6 (4.2)
1.7 (-6.5 to 9.9)
60.7
60.9 (4.5)
-0.2 (-9.0 to 8.6)
46.9
56.1 ( 4.9)
-9.2 (-18.8 to 0.4)
Deep and superficial
53 (6.9)
55.0
51.2 (7.0)
3.8 (-9.9 to 17.5)
45.7
37.9 (7.4)
7.8 (-6.7 to 22.3)
32.7
19.0 (10.2)
13.7 ( -6.3 to 33.7)
Histology
Myxofibrosarcoma
238 (31.1)
62.7
62.4 (3.2)
0.3 ( -6.0 to 6.6)
53.3
53.4 (3.4)
-0.1 ( -6.8 to 6.6)
39.1
36.3 ( 5.0)
2.8 ( -7.0 to 12.6)
MPNST
91 (11.9)
60.0
57.7 (5.2)
2.3 ( -7.9 to 12.5)
50.2
50.1 (5.4)
0.1 (-10.5 to 10.7)
35.9
33.4 ( 7.7)
2.5 (-12.6 to 17.6)
Synovial
sarcoma
142 (18.5)
73.8
75.1 (3.8)
-1.3 ( -8.7 to 6.1)
65.8
66.6 (4.2)
-0.8 ( -9.0 to 7.4)
52.7
50.9 ( 5.9)
1.8 ( -9.8 to 13.4)
Spindle cell sarcoma
167 (21.8)
55.6
59.9 (3.9)
-4.3 (-11.9 to 3.3)
45.4
47.4 (4.3)
-2.0 (-10.4 to 6.4)
30.9
41.5 ( 5.0)
-10.6 (-20.4 to -0.8)
MFH/UPS
77 (10.1)
53.7
54.8 (5.9)
-1.1 (-12.7 to 10.5)
43.6
44.8 (6.1)
-1.2 (-13.2 to 10.8)
29.6
29.2 ( 9.0)
0.4 (-17.2 to 18.0)
Margin
0 mm
140 (18.3)
52.3
51.5 (4.3)
0.8 (-7.6 to 9.2)
42.4
46.8 (4.4)
-4.4 (-13.0 to 4.2)
28.8
37.1 (4.8)
-8.3 (-17.7 to 1.1)
0.1-2 mm
343 (44.8)
63.1
64.6 (2.6)
-1.5 (-6.6 to 3.6)
53.5
52.4 (2.8)
1.1 ( -4.4 to 6.6)
39.1
37.9 (3.7)
1.2 ( -6.1 to 8.5)
> 2 mm
283 (36.9)
65.5
66.4 (2.9)
-0.9 (-6.6 to 4.8)
56.5
57.5 (3.2)
-1.0 ( -7.3 to 5.3)
42.9
39.5 (6.3)
3.4 ( -8.9 to 15.7)
RT
No RT
182 (23.8)
50.9
53.6 (3.8)
-2.7 (-10.1 to 4.7)
40.8
44.5 (3.9)
-3.7 (-11.3 to 3.9)
27.2
29.5 (6.1)
-2.3 (-14.3 to 9.7)
Neoadjuvant
184 (24)
69.3
69.2 (3.5)
0.1 ( -6.8 to 7.0)
60.7
60.3 (3.8)
0.4 ( -7.0 to 7.8)
47.0
40.3 (6.0)
6.7 ( -5.1 to 18.5)
Adjuvant
400 (52.2)
63.7
64.1 (2.5)
-0.4 ( -5.3 to 4.5)
54.3
53.7 (2.7)
0.6 ( -4.7 to 5.9)
40.0
43.3 (3.1)
-3.3 ( -9.4 to 2.8)

12
Fig. 2. Survival curves for 4 PI groups
Kaplan-Meier survival curves (solid lines) plotted with the model based survival curves (dotted lines) for 4 different prognostic
index groups. Number of patients at risk were 423, 265, and 33 at 3, 5 and 10 years respectively. Black: patients with good; red:
fairly good; green: fairly poor and blue: poor prognosis.
In the Fine and Gray model, tumour size, surgical margin and radiotherapy show a significant effect on
CILR (Table 4). Bigger tumours are associated with a higher probability of LR with sHR equal to 1.031
(95%CI: 1.001-1.063) for a unit increase of 1 cm. Patients with larger margins have a significantly lower
CILR with sHR equal to 0.635 (95%CI: 0.406-0.992) and 0.282 (95%CI: 0.159-0.500) for 0.1-2mm and
>2mm respectively. Radiotherapy treatment is associated with a lower CILR compared to surgery alone
with sHRs equal to 0.312 (95%CI: 0.146-0.668) and 0.700 (95%CI: 0.417-1.175) for neoadjuvant and
adjuvant radiotherapy respectively.

13
Table 4. Fine and Gray model for LR. Subdistribution hazard ratio (sHR) along with 95% confidence
interval (n = 766).
sHR
95%CI
P value
Age
1.051
0.942-1.184
0.337
Size
1.031
1.001-1.063
0.042
Depth*
0.559
Deep
1.000
Superficial
0.907
0.536-1.535
Deep and superficial
0.563
0.198-1.604
Histology
0.864
Myxofibrosarcoma
1.000
MPNST
1.079
0.580-2.009
Synovial sarcoma
0.779
0.379-1.602
Spindle cell sarcoma
0.979
0.570-1.681
MFH/UPS
1.096
0.557-2.156
Margin
<0.001
0 mm
1.000
0.1-2 mm
0.635
0.406-0.992
>2 mm
0.282
0.159-0.500
RT
0.010
No RT
1.000
Neoadjuvant
0.312
0.146-0.668
Adjuvant
0.700
0.417-1.175
The sHR of age corresponds to a unit increase of 10 years and the sHR of size corresponds to a unit increase of 1 cm.
MFH/UPS: Malignant Fibrous Histiocytoma/Undifferentiated Pleomorphic Sarcoma; MPNST: malignant peripheral nerve
sheath tumour; RT: radiotherapy. *Depth: relative to the investing fascia.
Calibration plots for LR are shown in Figure 3. Points are scattered around the lower diagonal that lies
within the 95%CIs of the observed cumulative incidence, indicating a good calibration. However, the
small distance between lower risk groups and the fact that groups observed outcome not always
monotonically increases indicate the relative difficulty to discriminate among patients with lower risk
profiles.

14
Fig. 3: Calibration plots for local recurrence.
Observed LR is plotted against predicted LR for patients in 8 equal sized risk groups identified by their predicted probability for
LR, as assessed by cross validation.
Figure 4 shows crude cumulative incidence curves (solid lines) and model-based estimates (dotted lines)
computed as the mean predicted cumulative incidence for LR. The high-risk groups can clearly be
distinguished from the rest. However, the curves of the lower risk groups are located very close to each
other, which indicates some difficulty to discriminate between patients with low risk resulting from the
small number of LRs observed in those groups.

15
Fig. 4: Cumulative incidence of local recurrence for 4 PI groups
Crude cumulative incidence curves (solid lines) plotted with the model based Cumulative incidence curves (dotted lines) for 4
different prognostic index groups. Number of patients at risk were 423, 265, and 33 at 3, 5 and 10 years
respectively. Black: patients with good; red: fairly good; green: fairly poor and blue: poor prognosis.
Figure 5 shows the effect of RT on OS and CILR for two patients with the same risk factors (70 years old, 9
cm tumour size, deep depth, MFH/UPS, resection margin of 0.1-2mm) with and without neoadjuvant RT.
The patient without RT (red lines) has worse OS and higher cumulative incidence of LR.

16
Fig. 5: Survival and CIFLR for patient of 70 years, tumour size 9cm, deep depth, MFH/UPS, resection
margin 0.1-2mm
In red: curves for patient treated with neoadjuvant RT. In black: patient without RT. Solid lines: survival curves. Dotted lines:
cumulative incidence for LR.
Detailed comparisons of observed and predicted probabilities for LR for data subgroups are shown in
Table 5. No significant differences between observed and predicted outcomes were evident. The C-index
for LR was 0.696 (95%CI 0.629-0.743).

17
Table 5. Comparing observed cumulative incidence and predicted probabilities of LR, assessed by cross
validation, for subgroups of data at 3, 5 and 10 years.
DISCUSSION
In cancer care, patient characteristics are generally set at presentation, whereas the combination and
timing of treatment(s) is a clinical decision based on each patient’s specific circumstances. Previously we
developed a multistate model to investigate how these variables affect patient outcomes.13 In this study
we developed the PERSARC model which uniquely presents clinicians with the possibility to accurately
predict outcome of OS and CILR, and compare different treatment modalities, for patients with high-
grade ESTS that undergo surgical resection with curative intent. It clearly shows the possible added value
of (neo)adjuvant radiotherapy at an individual patient level (figure 5). Surgical margins, adjuvant
therapies, and individual baseline characteristics are all incorporated in this model. To assess the
predictive value of this model, internal validation was performed.
n (%)
Predicted
3 years
Observed (s.e.)
Difference (95%
CI)
Predicted
5 years
Observed (s.e.)
Difference (95%
CI)
Predicted
10 years
Observed (s.e.)
Difference (95%
CI)
Age
30-60
281 (36.7)
11.4
10.2 (1.8)
1.2 ( -2.3 to 4.7)
13.0
11.5 (2.0)
1.5 ( -2.4 to 5.4)
14.8
12.9 (2.1)
1.9 ( -2.2 to 6.0)
<30
82 (10.7)
9.6
12.6 (3.7)
-3.0 (-10.3 to 4.3)
10.8
15.7 (4.2)
-4.9 (-13.1 to 3.3)
12.4
15.7 (4.2)
-3.3 (-11.5 to 4.9)
>60
403 (52.6)
15.5
15.6 (1.8)
-0.1 ( -3.6 to 3.4)
17.5
17.4 (2.0)
0.1 ( -3.8 to 4.0)
19.9
20.8 (2.8)
-0.9 ( -6.4 to 4.6)
Size
<5cm
123 (16.1)
9.9
8.4 (2.6)
1.5 (-3.6 to 6.6)
11.3
11.9 (3.2)
-0.6 (-6.9 to 5.7)
12.9
18.2 (5.6)
-5.3 (-16.3 to 5.7)
5cm-10cmS
295 (38.5)
11.5
10.1 (1.8)
1.4 (-2.1 to 4.9)
13.1
11.9 (2.0)
1.2 (-2.7 to 5.1)
15.0
14.1 (2.5)
0.9 ( -4.0 to 5.8)
>=10cm
348 (45.4)
16.2
17.8 (2.1)
-1.6 (-5.7 to 2.5)
18.3
18.8 (2.1)
-0.5 (-4.6 to 3.6)
20.7
19.4 (2.2)
1.3 ( -3.0 to 5.6)
Depth*
Deep
579 (75.6)
13.9
13.9 (1.5)
0.0 (-2.9 to 2.9)
15.7
15.6 (1.6)
0.1 (-3.0 to 3.2)
17.8
17.9 (1.9)
-0.1 (-3.8 to 3.6)
Superficial
134 (17.5)
13.4
12.3 (2.9)
1.1 (-4.6 to 6.8)
15.2
14.8 (3.3)
0.4 (-6.1 to 6.9)
17.3
16.7 (3.7)
0.6 (-6.7 to 7.9)
Deep and
superficial
53 (6.9)
8.1
9.6 (4.1)
-1.5 (-9.5 to 6.5)
9.2
9.6 (4.1)
-0.4 (-8.4 to 7.6)
10.5
9.6 (4.1)
0.9 (-7.1 to 8.9)
Histology
Myxofibrosarcoma
238 (31.1)
12.1
11.9 (2.1)
0.2 ( -3.9 to 4.3)
13.7
12.9 (2.2)
0.8 ( -3.5 to 5.1)
15.7
15.7 (3.0)
0.0 ( -5.9 to 5.9)
MPNST
91 (11.9)
15.6
17.7 (4.0)
-2.1 ( -9.9 to 5.7)
17.6
17.7 (4.0)
-0.1 ( -7.9 to 7.7)
20.0
19.6 (4.3)
0.4 ( -8.0 to 8.8)
Synovial sarcoma
142 (18.5)
7.2
4.5 (1.8)
2.7 ( -0.8 to 6.2)
8.2
9.1 (2.6)
-0.9 ( -6.0 to 4.2)
9.4
10.4 (2.9)
-1.0 ( -6.7 to 4.7)
Spindle cell
sarcoma
167 (21.8)
16.8
16.7 (2.9)
0.1 ( -5.6 to 5.8)
19.0
19.0 (3.3)
0.0 ( -6.5 to 6.5)
21.6
26.3 (7.3)
-4.7 (-19.0 to 9.6)
MFH/UPS
77 (10.1)
18.1
19.3 (4.6)
-1.2 (-10.2 to 7.8)
20.3
20.9 (4.8)
-0.6 (-10.0 to 8.8)
23.0
20.9 (4.8)
2.1 ( -7.3 to 11.5)
Margin
0 mm
140 (18.3)
23.9
26.2 (3.8)
-2.3 (-9.7 to 5.1)
26.9
26.2 (3.8)
0.7 (-6.7 to 8.1)
30.3
26.2 (3.8)
4.1 (-3.3 to 11.5)
0.1-2 mm
343 (44.8)
14.5
13.4 (1.9)
1.1 (-2.6 to 4.8)
16.4
15.9 (2.0)
0.5 (-3.4 to 4.4)
18.7
19.3 (2.6)
-0.6 (-5.7 to 4.5)
> 2 mm
283 (36.9)
6.8
6.7 (1.5)
0.1 (-2.8 to 3.0)
7.8
8.3 (1.8)
-0.5 (-4.0 to 3.0)
9.0
9.1 (1.9)
-0.1 (-3.8 to 3.6)
RT
No RT
182 (23.8)
16.4
15.3 (2.7)
1.1 (-4.2 to 6.4)
18.5
18.6 (3.1)
-0.1 (-6.2 to 6.0)
20.9
19.9 (3.3)
1.0 (-5.5 to 7.5)
Neoadjuvant
184 (24)
6.0
7.3 (2.0)
-1.3 (-5.2 to 2.6)
6.9
7.3 (2.0)
-0.4 (-4.3 to 3.5)
7.9
7.3 (2.0)
0.6 (-3.3 to 4.5)
Adjuvant
400 (52.2)
15.4
15.1 (1.8)
0.3 (-3.2 to 3.8)
17.4
17.0 (1.9)
0.4 (-3.3 to 4.1)
19.9
21.0 (2.8)
-1.1 (-6.6 to 4.4)

18
This prognostic model illustrates that as the tumour size increases, the prognosis worsens for LR and OS
with sHR equal to 1.031 (95%CI: 1.001-1.063) and HR equal to 1.068 (95%CI: 1.052-1.085) respectively.
These findings are similar to results reported by other groups.12 As expected, age was an adverse
prognostic risk factor for OS,3 which can only be partially explained by comorbidities. Margins are clearly
associated with LR, and seem to have a marginally significant effect on OS (Table 2 and 4). The effect of
recurrence on OS might be attributed to biological aggressiveness of the tumour rather than margins
itself (Table 2 and 4).1,24
Patients who received radiotherapy seem to have better outcomes than those who did not (Table 2 and
4).25 These patients may have been selected out of the total group of ESTS patients based on clinical
characteristics, presenting scenarios, or capability to undergo neoadjuvant radiotherapy.26 All patients
included in this study were treated at one of the five high volume sarcoma centres following discussion of
their case at a multidisciplinary tumour board. Although selection bias may be present, it only reflects
every day care decisions. There are two prospective randomised trials on this topic; in both studies
adjuvant radiotherapy has shown a decrease in LR but had no significant impact on survival. However,
both studies also included patients with low-grade tumours. Furthermore, due to low number of events
(death) per arm they could only detect a minimal benefit of 20% (as mentioned in the trail that had the
most patients per arm).27,28 Previous studies along with the results from this investigation suggest that
neoadjuvant radiotherapy should be considered at multidisciplinary tumour board discussions for all
patients undergoing surgery for primary high-grade ESTS.25,29-32 Patients treated with neoadjuvant
radiation are at significantly increased risk of wound healing complications, whether they receive
conventional treatment or intensity-modulated radiotherapy.25 Therefore certain patients such as the
elderly, those with significant medical comorbidities, or those with prosthetic implants adjacent to the
location of the sarcoma, may be considered inappropriate candidates for neoadjuvant radiation.
The outcomes presented above must be interpreted with caution, since this model is based on clinical
routine data and is therefore susceptible to selection bias. In addition, margin categories are based on
millimetres and histology was not re-evaluated centrally. Therefore, margin assessment and evaluation of
specific ‘close’ margins to anatomic structures; e.g. periosteum, perineurium, or fascia may be subjective
to variability.26 For patients treated in centres where other margin criteria (e.g. Enneking) are in place,
this model may be less applicable. Further research should focus on evaluating the different classification
methods and agreeing on one standardised margin description for patients with ESTS.33-35

19
While some patients may accept the increased risk of a LR and potential need for subsequent treatment
by opting for less aggressive therapy including minimal margins, others may want to minimize the risk of
another surgery, for example because of age and comorbidities or because of the potential effect on
survival. These trade-offs are delicate and have to be based on clinical experience and substantial
evidence. The prediction model developed in this study provides some indication about the possible
evolution of the disease and helps in shared decision-making. The Personalised Sarcoma Care model is
freely available in the Appstore and Google apps.
Funding: The Dutch Cancer Society (DCS) - KWF Kankerbestrijding
Role of the funding source:
This funding source had no role in the design of this study as well as any role during its execution,
analyses, interpretation of the data, in the writing of the report, or decision to submit the article for
publication
REFERENCES
1. Willeumier J, Fiocco M, Nout R, et al. High-grade soft tissue sarcomas of the extremities: surgical margins influence only local recurrence
not overall survival. Int Orthop 2015; 39(5): 935-41.
2. Engellau J, Samuelsson V, Anderson H, et al. Identification of low-risk tumours in histological high-grade soft tissue sarcomas. European
journal of cancer (Oxford, England : 1990) 2007; 43(13): 1927-34.
3. Mendenhall WM, Indelicato DJ, Scarborough MT, et al. The management of adult soft tissue sarcomas. American journal of clinical
oncology 2009; 32(4): 436-42.
4. Gronchi A, Lo Vullo S, Colombo C, et al. Extremity soft tissue sarcoma in a series of patients treated at a single institution: local control
directly impacts survival. Annals of surgery 2010; 251(3): 506-11.
5. Biau DJ, Ferguson PC, Chung P, et al. Local recurrence of localized soft tissue sarcoma: a new look at old predictors. Cancer 2012; 118(23):
5867-77.
6. Davis LE, Ryan CW. Preoperative therapy for extremity soft tissue sarcomas. Current treatment options in oncology 2015; 16(6): 25.
7. Trojani M, Contesso G, Coindre JM, et al. Soft-tissue sarcomas of adults; study of pathological prognostic variables and definition of a
histopathological grading system. International journal of cancer 1984; 33(1): 37-42.
8. Maretty-Nielsen K, Aggerholm-Pedersen N, Safwat A, et al. Prognostic factors for local recurrence and mortality in adult soft tissue
sarcoma of the extremities and trunk wall: a cohort study of 922 consecutive patients. Acta orthopaedica 2014; 85(3): 323-32.
9. Gustafson P, Akerman M, Alvegard TA, et al. Prognostic information in soft tissue sarcoma using tumour size, vascular invasion and
microscopic tumour necrosis-the SIN-system. European journal of cancer (Oxford, England : 1990) 2003; 39(11): 1568-76.
10. Carneiro A, Bendahl PO, Engellau J, et al. A prognostic model for soft tissue sarcoma of the extremities and trunk wall based on size,
vascular invasion, necrosis, and growth pattern. Cancer 2011; 117(6): 1279-87.

20
11. Sampo M, Tarkkanen M, Tukiainen E, et al. A web-based prognostic tool for extremity and trunk wall soft tissue sarcomas and its external
validation. British journal of cancer 2012; 106(6): 1076-82.
12. Callegaro D, Miceli R, Bonvalot S, et al. Development and external validation of two nomograms to predict overall survival and occurrence
of distant metastases in adults after surgical resection of localised soft-tissue sarcomas of the extremities: a retrospective analysis. The Lancet
Oncology 2016; 17(5): 671-80.
13. Willeumier JJ, Rueten-Budde AJ, Jeys LM, et al. Individualised risk assessment for local recurrence and distant metastases in a
retrospective transatlantic cohort of 687 patients with high-grade soft tissue sarcomas of the extremities: a multistate model. BMJ open 2017; 7(2):
e012930.
14. Group. TEESNW. Soft tissue and visceral sarcomas: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up. Annals of
Oncology 25 (Supplement 3): iii102–iii112 2014.
15. Kainhofer V, Smolle MA, Szkandera J, et al. The width of resection margins influences local recurrence in soft tissue sarcoma patients.
European journal of surgical oncology : the journal of the European Society of Surgical Oncology and the British Association of Surgical Oncology
2016; 42(6): 899-906.
16. Putter H, Fiocco M, Geskus RB. Tutorial in biostatistics: competing risks and multi-state models. Statistics in medicine 2007; 26(11): 2389-
430.
17. van Houwelingen HC. Validation, calibration, revision and combination of prognostic survival models. Statistics in medicine 2000; 19(24):
3401-15.
18. Wolbers M, Koller MT, Witteman JC, Steyerberg EW. Prognostic models with competing risks: methods and application to coronary risk
prediction. Epidemiology (Cambridge, Mass) 2009; 20(4): 555-61.
19. de Glas NA, Bastiaannet E, Engels CC, et al. Validity of the online PREDICT tool in older patients with breast cancer: a population-based
study. British journal of cancer 2016; 114(4): 395-400.
20. Royston P, Altman DG. External validation of a Cox prognostic model: principles and methods. BMC medical research methodology 2013;
13: 33.
21. Van Houwelingen H, and Putter, H. . Dynamic Prediction in Clinical Survival Analysis. Boca Raton, Florida: CRC Press/Chapman & Hall;
2012.
22. Akobeng AK. Understanding diagnostic tests 3: Receiver operating characteristic curves. Acta paediatrica (Oslo, Norway : 1992) 2007;
96(5): 644-7.
23. Team RC. R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing.; 2016.
24. Guerrero WM, Deneve JL. Local Recurrence of Extremity Soft Tissue Sarcoma. The Surgical clinics of North America 2016; 96(5): 1157-74.
25. O'Sullivan B, Davis AM, Turcotte R, et al. Preoperative versus postoperative radiotherapy in soft-tissue sarcoma of the limbs: a
randomised trial. Lancet (London, England) 2002; 359(9325): 2235-41.
26. O'Donnell PW, Griffin AM, Eward WC, et al. The effect of the setting of a positive surgical margin in soft tissue sarcoma. Cancer 2014;
120(18): 2866-75.
27. Brennan MF, Casper ES, Harrison LB, Shiu MH, Gaynor J, Hajdu SI. The role of multimodality therapy in soft-tissue sarcoma. Annals of
surgery 1991; 214(3): 328-36; discussion 36-8.
28. Beane JD, Yang JC, White D, Steinberg SM, Rosenberg SA, Rudloff U. Efficacy of adjuvant radiation therapy in the treatment of soft tissue
sarcoma of the extremity: 20-year follow-up of a randomized prospective trial. Annals of surgical oncology 2014; 21(8): 2484-9.
29. Al-Absi E, Farrokhyar F, Sharma R, et al. A systematic review and meta-analysis of oncologic outcomes of pre- versus postoperative
radiation in localized resectable soft-tissue sarcoma. Annals of surgical oncology 2010; 17(5): 1367-74.
30. Muller DA, Beltrami G, Scoccianti G, Frenos F, Capanna R. Combining limb-sparing surgery with radiation therapy in high-grade soft tissue
sarcoma of extremities - Is it effective? European journal of surgical oncology : the journal of the European Society of Surgical Oncology and the
British Association of Surgical Oncology 2016; 42(7): 1057-63.

21
31. Sampath S, Schultheiss TE, Hitchcock YJ, Randall RL, Shrieve DC, Wong JY. Preoperative versus postoperative radiotherapy in soft-tissue
sarcoma: multi-institutional analysis of 821 patients. International journal of radiation oncology, biology, physics 2011; 81(2): 498-505.
32. Zagars GK, Ballo MT, Pisters PW, Pollock RE, Patel SR, Benjamin RS. Preoperative vs. postoperative radiation therapy for soft tissue
sarcoma: a retrospective comparative evaluation of disease outcome. International journal of radiation oncology, biology, physics 2003; 56(2): 482-
8.
33. Enneking WF, Spanier SS, Goodman MA. A system for the surgical staging of musculoskeletal sarcoma. Clinical orthopaedics and related
research 1980; (153): 106-20.
34. Kawaguchi N, Ahmed AR, Matsumoto S, Manabe J, Matsushita Y. The concept of curative margin in surgery for bone and soft tissue
sarcoma. Clinical orthopaedics and related research 2004; (419): 165-72.
35. Hoang K, Gao Y, Miller BJ. The Variability in Surgical Margin Reporting in Limb Salvage Surgery for Sarcoma. The Iowa orthopaedic journal
2015; 35: 181-6.
Appendix I. Competing risk model
A patient enters the state of being alive with no evidence of disease (ANED) after surgery and may move
to the state of local recurrence (LR) or Death.